v1.18.0

feat(readme): expand documentation for voicemail, IVR, audio engine, and API capabilities
v1.17.2
2026-04-10 17:25:34 +00:00 · 2026-04-10 17:25:34 +00:00 · 2026-04-10 17:14:14 +00:00 · 2026-04-10 17:14:14 +00:00 · 2026-04-10 16:57:07 +00:00 · 2026-04-10 16:57:07 +00:00
25 changed files with 880 additions and 255 deletions
--- a/changelog.md
+++ b/changelog.md
@@ -1,5 +1,36 @@
 # Changelog

+## 2026-04-10 - 1.18.0 - feat(readme)
+expand documentation for voicemail, IVR, audio engine, and API capabilities
+
+- Updates the feature overview to document voicemail, IVR menus, call recording, enhanced TTS, and the 48kHz float audio engine
+- Refreshes the architecture section to describe the TypeScript control plane, Rust proxy-engine data plane, and JSON-over-stdio IPC
+- Clarifies REST API and WebSocket coverage with voicemail endpoints, incoming call events, and refined endpoint descriptions
+
+## 2026-04-10 - 1.17.2 - fix(proxy-engine)
+use negotiated SDP payload types when wiring SIP legs and enable default nnnoiseless features for telephony denoising
+
+- Select the negotiated codec payload type from SDP answers instead of always using the first offered codec
+- Preserve the device leg's preferred payload type from its own INVITE SDP when attaching it to the mixer
+- Enable default nnnoiseless features in codec-lib and proxy-engine dependencies
+
+## 2026-04-10 - 1.17.1 - fix(proxy-engine,codec-lib,sip-proto,ts)
+preserve negotiated media details and improve RTP audio handling across call legs
+
+- Use native Opus float encode/decode to avoid unnecessary i16 quantization in the f32 audio path.
+- Parse full RTP headers including extensions and sequence numbers, then sort inbound packets before decoding to keep codec state stable for out-of-order audio.
+- Capture negotiated codec payload types from SDP offers and answers and include codec, RTP port, remote media, and metadata in leg_added events.
+- Emit leg_state_changed and leg_removed events more consistently so the dashboard reflects leg lifecycle updates accurately.
+
+## 2026-04-10 - 1.17.0 - feat(proxy-engine)
+upgrade the internal audio bus to 48kHz f32 with per-leg denoising and improve SIP leg routing
+
+- switch mixer, prompt playback, and tool leg audio handling from 16kHz i16 to 48kHz f32 for higher-quality internal processing
+- add f32 decode/encode and resampling support plus standalone RNNoise denoiser creation in codec-lib
+- apply per-leg inbound noise suppression in the mixer before mix-minus generation
+- fix passthrough call routing by matching the actual leg from the signaling source address when Call-IDs are shared
+- correct dialed number extraction from bare SIP request URIs by parsing the user part directly
+
 ## 2026-04-10 - 1.16.0 - feat(proxy-engine)
 integrate Kokoro TTS generation into proxy-engine and simplify TypeScript prompt handling to use cached WAV files

--- a/nogit/voicemail/default/msg-1775840000387.wav
+++ b/nogit/voicemail/default/msg-1775840000387.wav
--- a/nogit/voicemail/default/msg-1775840014276.wav
+++ b/nogit/voicemail/default/msg-1775840014276.wav
--- a/nogit/voicemail/default/msg-1775840439400.wav
+++ b/nogit/voicemail/default/msg-1775840439400.wav
--- a/nogit/voicemail/default/msg-1775840447441.wav
+++ b/nogit/voicemail/default/msg-1775840447441.wav
--- a/nogit/voicemail/default/msg-1775840454835.wav
+++ b/nogit/voicemail/default/msg-1775840454835.wav
--- a/package.json
+++ b/package.json
@@ -1,6 +1,6 @@
 {
  "name": "siprouter",
-  "version": "1.16.0",
+  "version": "1.18.0",
  "private": true,
  "type": "module",
  "scripts": {
--- a/readme.md
+++ b/readme.md
@@ -1,6 +1,6 @@
 # @serve.zone/siprouter

-A production-grade **SIP B2BUA + WebRTC bridge** built with TypeScript and Rust. Routes calls between SIP providers, SIP hardware devices, and browser softphones — with real-time codec transcoding, ML noise suppression, neural TTS announcements, and a slick web dashboard.
+A production-grade **SIP B2BUA + WebRTC bridge** built with TypeScript and Rust. Routes calls between SIP providers, SIP hardware devices, and browser softphones — with real-time codec transcoding, ML noise suppression, neural TTS, voicemail, IVR menus, and a slick web dashboard.

 ## Issue Reporting and Security

@@ -12,14 +12,16 @@ For reporting bugs, issues, or security vulnerabilities, please visit [community

 siprouter sits between your SIP trunk providers and your endpoints — hardware phones, ATAs, browser softphones — and handles **everything** in between:

- 📞 **SIP B2BUA** — Terminates and re-originates calls with full RFC 3261 dialog state management
- 🌐 **WebRTC Bridge** — Browser-based softphone with bidirectional audio to the SIP network
- 🎛️ **Multi-Provider Trunking** — Register with multiple SIP providers simultaneously (sipgate, easybell, o2, etc.)
- 🔊 **Rust Codec Engine** — Real-time Opus ↔ G.722 ↔ PCMU ↔ PCMA transcoding in native Rust
- 🤖 **ML Noise Suppression** — RNNoise denoiser with per-direction state (to SIP / to browser)
- 🗣️ **Neural TTS** — Kokoro-powered "connecting your call" announcements, pre-encoded for instant playback
- 🔀 **Hub Model Calls** — N-leg calls with dynamic add/remove, transfer, and RTP fan-out
- 🖥️ **Web Dashboard** — Real-time SPA with live call monitoring, browser phone, contact management, provider config
+- 📞 **SIP B2BUA** — Terminates and re-originates calls with full RFC 3261 dialog state management, digest auth, and SDP negotiation
+- 🌐 **WebRTC Bridge** — Browser-based softphone with bidirectional Opus audio to the SIP network
+- 🎛️ **Multi-Provider Trunking** — Register with multiple SIP providers simultaneously (sipgate, easybell, etc.) with automatic failover
+- 🎧 **48kHz f32 Audio Engine** — High-fidelity internal audio bus at 48kHz/32-bit float with native Opus float encode/decode, FFT-based resampling, and per-leg ML noise suppression
+- 🔀 **N-Leg Mix-Minus Mixer** — Conference-grade mixing with dynamic leg add/remove, transfer, and per-source audio separation
+- 📧 **Voicemail** — Configurable voicemail boxes with TTS greetings, recording, and web playback
+- 🔢 **IVR Menus** — DTMF-navigable interactive voice response with nested menus, routing actions, and custom prompts
+- 🗣️ **Neural TTS** — Kokoro-powered announcements and greetings with 25+ voice presets, backed by espeak-ng fallback
+- 🎙️ **Call Recording** — Per-source separated WAV recording at 48kHz via tool legs
+- 🖥️ **Web Dashboard** — Real-time SPA with 9 views: live calls, browser phone, routing, voicemail, IVR, contacts, providers, and streaming logs

 ---

@@ -35,32 +37,38 @@ siprouter sits between your SIP trunk providers and your endpoints — hardware
 ┌──────────────────────────────────────┐
 │            siprouter                  │
 │                                      │
-│  ┌──────────┐  ┌──────────────────┐  │
-│  │ Call Hub  │  │  Rust Transcoder │  │
-│  │  N legs   │──│  Opus/G.722/PCM │  │
-│  │  fan-out  │  │  + RNNoise      │  │
-│  └────┬─────┘  └──────────────────┘  │
-│       │                              │
-│  ┌────┴─────┐  ┌──────────────────┐  │
-│  │ SIP Stack│  │  Kokoro TTS      │  │
-│  │ Dialog SM│  │  (ONNX Runtime)  │  │
-│  └────┬─────┘  └──────────────────┘  │
-│       │                              │
-│  ┌────┴──────────────────────────┐   │
-│  │   Local Registrar + Provider  │   │
-│  │   Registration Engine         │   │
-│  └───────────────────────────────┘   │
-└──────────┬──────────────┬────────────┘
-           │              │
-    ┌──────┴──────┐ ┌─────┴──────┐
-    │ SIP Devices │ │ SIP Trunk  │
-    │ (HT801, etc)│ │ Providers  │
-    └─────────────┘ └────────────┘
+│  TypeScript Control Plane            │
+│  ┌────────────────────────────────┐  │
+│  │ Config · WebRTC Signaling      │  │
+│  │ REST API · Web Dashboard       │  │
+│  │ Voicebox Manager · TTS Cache   │  │
+│  └────────────┬───────────────────┘  │
+│          JSON-over-stdio IPC         │
+│  ┌────────────┴───────────────────┐  │
+│  │ Rust proxy-engine (data plane) │  │
+│  │                                │  │
+│  │ SIP Stack · Dialog SM · Auth   │  │
+│  │ Call Manager · N-Leg Mixer     │  │
+│  │ 48kHz f32 Bus · RNNoise       │  │
+│  │ Codec Engine · RTP Port Pool   │  │
+│  │ WebRTC Engine · Kokoro TTS     │  │
+│  │ Voicemail · IVR · Recording    │  │
+│  └────┬──────────────────┬────────┘  │
+└───────┤──────────────────┤───────────┘
+        │                  │
+ ┌──────┴──────┐    ┌──────┴──────┐
+ │ SIP Devices │    │ SIP Trunk   │
+ │ (HT801 etc) │    │ Providers   │
+ └─────────────┘    └─────────────┘
 ```

-### The Hub Model
+### 🧠 Key Design Decisions

-Every call is a **hub** with N legs. Each leg is either a `SipLeg` (hardware device or provider) or a `WebRtcLeg` (browser). RTP flows through the hub — each leg's received audio is forwarded to all other legs, with codec transcoding handled transparently by the Rust engine.
+- **Hub Model** — Every call is a hub with N legs. Each leg is a `SipLeg` (device/provider) or `WebRtcLeg` (browser). Legs can be dynamically added, removed, or transferred without tearing down the call.
+- **Rust Data Plane** — All SIP protocol handling, codec transcoding, mixing, and RTP I/O runs in native Rust for real-time performance. TypeScript handles config, signaling, REST API, and dashboard.
+- **48kHz f32 Internal Bus** — Audio is processed at maximum quality internally. Encoding/decoding to wire format (G.722, PCMU, Opus) happens solely at the leg boundary.
+- **Per-Session Codec Isolation** — Each call leg gets its own encoder/decoder/resampler/denoiser state — no cross-call corruption.
+- **SDP Codec Negotiation** — Outbound encoding uses the codec actually negotiated in SDP answers, not just the first offered codec.

 ---

@@ -70,15 +78,16 @@ Every call is a **hub** with N legs. Each leg is either a `SipLeg` (hardware dev

 - **Node.js** ≥ 20 with `tsx` globally available
 - **pnpm** for package management
- **Rust** toolchain (for building the codec engine and TTS)
+- **Rust** toolchain (for building the proxy engine)
+- **espeak-ng** (optional, for TTS fallback)

 ### Install & Build

 ```bash
-# Clone and install
+# Clone and install dependencies
 pnpm install

-# Build the Rust binaries (opus-codec + tts-engine)
+# Build the Rust proxy-engine binary
 pnpm run buildRust

 # Bundle the web frontend
@@ -87,57 +96,92 @@ pnpm run bundle

 ### Configuration

-Create `.nogit/config.json` with your setup:
+Create `.nogit/config.json`:

 ```jsonc
 {
  "proxy": {
-    "lanIp": "192.168.1.100",     // Your server's LAN IP
-    "lanPort": 5070,               // SIP signaling port
-    "rtpPortRange": [20000, 20200],// RTP relay port pool (even ports)
-    "webUiPort": 3060              // Dashboard port
+    "lanIp": "192.168.1.100",          // Your server's LAN IP
+    "lanPort": 5070,                    // SIP signaling port
+    "publicIpSeed": "stun.example.com", // STUN server for public IP discovery
+    "rtpPortRange": { "min": 20000, "max": 20200 }, // RTP port pool (even ports)
+    "webUiPort": 3060                   // Dashboard + REST API port
  },
  "providers": [
    {
      "id": "my-trunk",
-      "name": "My SIP Provider",
-      "host": "sip.provider.com",
-      "port": 5060,
+      "displayName": "My SIP Provider",
+      "domain": "sip.provider.com",
+      "outboundProxy": { "address": "sip.provider.com", "port": 5060 },
      "username": "user",
      "password": "pass",
-      "codecs": ["G.722", "PCMA", "PCMU"],
-      "registerExpiry": 3600
+      "codecs": [9, 0, 8, 101],        // G.722, PCMU, PCMA, telephone-event
+      "registerIntervalSec": 300
    }
  ],
  "devices": [
    {
      "id": "desk-phone",
-      "name": "Desk Phone",
-      "type": "sip"
+      "displayName": "Desk Phone",
+      "expectedAddress": "192.168.1.50",
+      "extension": "100"
    }
  ],
  "routing": {
-    "inbound": {
-      "default": { "target": "all-devices", "ringBrowser": true }
+    "routes": [
+      {
+        "id": "inbound-default",
+        "name": "Ring all devices",
+        "priority": 100,
+        "direction": "inbound",
+        "match": {},
+        "action": {
+          "targets": ["desk-phone"],
+          "ringBrowsers": true,
+          "voicemailBox": "main",
+          "noAnswerTimeout": 25
+        }
+      },
+      {
+        "id": "outbound-default",
+        "name": "Route via trunk",
+        "priority": 100,
+        "direction": "outbound",
+        "match": {},
+        "action": { "provider": "my-trunk" }
+      }
+    ]
+  },
+  "voiceboxes": [
+    {
+      "id": "main",
+      "enabled": true,
+      "greetingText": "Please leave a message after the beep.",
+      "greetingVoice": "af_bella",
+      "noAnswerTimeoutSec": 25,
+      "maxRecordingSec": 120,
+      "maxMessages": 50
    }
-  }
+  ],
+  "contacts": [
+    { "id": "1", "name": "Alice", "number": "+491234567890", "starred": true }
+  ]
 }
 ```

 ### TTS Setup (Optional)

-For neural "connecting your call" announcements, download the Kokoro TTS model:
+For neural announcements and voicemail greetings, download the Kokoro TTS model:

 ```bash
 mkdir -p .nogit/tts
-# Download the full-quality model (310MB) + voices (27MB)
 curl -L -o .nogit/tts/kokoro-v1.0.onnx \
  https://github.com/mzdk100/kokoro/releases/download/V1.0/kokoro-v1.0.onnx
 curl -L -o .nogit/tts/voices.bin \
  https://github.com/mzdk100/kokoro/releases/download/V1.0/voices.bin
 ```

-If the model files aren't present, the announcement feature is simply disabled — everything else works fine.
+Without the model files, TTS falls back to `espeak-ng`. Without either, announcements are skipped — everything else works fine.

 ### Run

@@ -145,7 +189,7 @@ If the model files aren't present, the announcement feature is simply disabled
 pnpm start
 ```

-The SIP proxy starts on the configured port and the web dashboard is available at `http://<your-ip>:3060`.
+The SIP proxy starts on the configured port and the web dashboard is available at `https://<your-ip>:3060`.

 ### HTTPS (Optional)

@@ -157,68 +201,91 @@ Place `cert.pem` and `key.pem` in `.nogit/` for TLS on the dashboard.

 ```
 siprouter/
-├── ts/                        # TypeScript source
-│   ├── sipproxy.ts            # Main entry — bootstraps everything
-│   ├── config.ts              # Config loader & validation
-│   ├── registrar.ts           # Local SIP registrar for devices
-│   ├── providerstate.ts       # Per-provider upstream registration engine
-│   ├── frontend.ts            # Web dashboard HTTP/WS server + REST API
-│   ├── webrtcbridge.ts        # WebRTC signaling layer
-│   ├── opusbridge.ts          # Rust IPC bridge (smartrust)
-│   ├── codec.ts               # High-level RTP transcoding interface
-│   ├── announcement.ts        # Neural TTS announcement generator
-│   ├── sip/                   # Zero-dependency SIP protocol library
-│   │   ├── message.ts         #   SIP message parser/builder/mutator
-│   │   ├── dialog.ts          #   RFC 3261 dialog state machine
-│   │   ├── helpers.ts         #   SDP builder, digest auth, codec registry
-│   │   └── rewrite.ts         #   SIP URI + SDP body rewriting
-│   └── call/                  # Hub-model call management
-│       ├── call-manager.ts    #   Central registry, factory, routing
-│       ├── call.ts            #   Call hub — owns N legs, media fan-out
-│       ├── sip-leg.ts         #   SIP device/provider connection
-│       ├── webrtc-leg.ts      #   Browser WebRTC connection
-│       └── rtp-port-pool.ts   #   UDP port allocation
-├── ts_web/                    # Web frontend (Lit-based SPA)
-│   ├── elements/              #   Web components (dashboard, phone, etc.)
-│   └── state/                 #   App state, WebRTC client, notifications
-├── rust/                      # Rust workspace
+├── ts/                            # TypeScript control plane
+│   ├── sipproxy.ts                # Main entry — bootstraps everything
+│   ├── config.ts                  # Config loader & validation
+│   ├── proxybridge.ts             # Rust proxy-engine IPC bridge (smartrust)
+│   ├── frontend.ts                # Web dashboard HTTP/WS server + REST API
+│   ├── webrtcbridge.ts            # WebRTC signaling layer
+│   ├── registrar.ts               # Browser softphone registration
+│   ├── announcement.ts            # TTS announcement generator (espeak-ng / Kokoro)
+│   ├── voicebox.ts                # Voicemail box management
+│   └── call/
+│       └── prompt-cache.ts        # Named audio prompt WAV management
+│
+├── ts_web/                        # Web frontend (Lit-based SPA)
+│   ├── elements/                  # Web components (9 dashboard views)
+│   └── state/                     # App state, WebRTC client, notifications
+│
+├── rust/                          # Rust workspace (the data plane)
 │   └── crates/
-│       ├── opus-codec/        #   Real-time audio transcoder (Opus/G.722/PCM)
-│       └── tts-engine/        #   Kokoro neural TTS CLI
-├── html/                      # Static HTML shell
-├── .nogit/                    # Secrets, config, models (gitignored)
-└── dist_rust/                 # Compiled Rust binaries (gitignored)
+│       ├── codec-lib/             # Audio codec library (Opus/G.722/PCMU/PCMA)
+│       ├── sip-proto/             # Zero-dependency SIP protocol library
+│       └── proxy-engine/          # Main binary — SIP engine + mixer + RTP
+│
+├── html/                          # Static HTML shell
+├── .nogit/                        # Secrets, config, TTS models (gitignored)
+└── dist_rust/                     # Compiled Rust binary (gitignored)
 ```

 ---

-## 🎧 Codec Engine (Rust)
+## 🎧 Audio Engine (Rust)

-The `opus-codec` binary handles all real-time audio processing via a JSON-over-stdio IPC protocol:
+The `proxy-engine` binary handles all real-time audio processing with a **48kHz f32 internal bus** — encoding and decoding happens only at leg boundaries.

-| Codec | Payload Type | Sample Rate | Use Case |
-|-------|-------------|-------------|----------|
-| **Opus** | 111 | 48 kHz | WebRTC browsers |
-| **G.722** | 9 | 16 kHz | HD SIP devices |
+### Supported Codecs
+
+| Codec | PT | Native Rate | Use Case |
+|-------|:--:|:-----------:|----------|
+| **Opus** | 111 | 48 kHz | WebRTC browsers (native float encode/decode — zero i16 quantization) |
+| **G.722** | 9 | 16 kHz | HD SIP devices & providers |
 | **PCMU** (G.711 µ-law) | 0 | 8 kHz | Legacy SIP |
 | **PCMA** (G.711 A-law) | 8 | 8 kHz | Legacy SIP |

-**Features:**
- Per-call isolated codec sessions (no cross-call state corruption)
- FFT-based sample rate conversion via `rubato`
- **RNNoise ML noise suppression** with per-direction state — denoises audio flowing to SIP separately from audio flowing to the browser
- Raw PCM encoding for TTS frame processing
+### Audio Pipeline
+
+```
+Inbound:   Wire RTP → Decode → Resample to 48kHz → Denoise (RNNoise) → Mix Bus
+Outbound:  Mix Bus → Mix-Minus → Resample to codec rate → Encode → Wire RTP
+```
+
+- **FFT-based resampling** via `rubato` — high-quality sinc interpolation with cached resampler state for seamless inter-frame continuity
+- **ML noise suppression** via `nnnoiseless` (RNNoise) — per-leg inbound denoising with SIMD acceleration (AVX/SSE). Skipped for WebRTC legs (browsers already denoise via getUserMedia)
+- **Mix-minus mixing** — each participant hears everyone except themselves, accumulated in f64 precision
+- **In-tick packet reorder** — inbound RTP packets are sorted by sequence number before decoding, protecting G.722 ADPCM state from out-of-order delivery
+- **RFC 3550 compliant header parsing** — properly handles CSRC lists and header extensions

 ---

-## 🗣️ Neural TTS (Rust)
+## 🗣️ Neural TTS

-The `tts-engine` binary uses [Kokoro TTS](https://github.com/mzdk100/kokoro) (82M parameter neural model) to synthesize announcements at startup:
+Announcements and voicemail greetings are synthesized using [Kokoro TTS](https://github.com/mzdk100/kokoro) — an 82M parameter neural model running via ONNX Runtime directly in the Rust process:

 - **24 kHz, 16-bit mono** output
 - **25+ voice presets** — American/British, male/female (e.g., `af_bella`, `am_adam`, `bf_emma`, `bm_george`)
- **~800ms** synthesis time for a 3-second announcement
- Pre-encoded to G.722 + Opus for zero-latency RTP playback during call setup
+- **~800ms** synthesis time for a 3-second phrase
+- Lazy-loaded on first use — no startup cost if TTS is unused
+- Falls back to `espeak-ng` if the ONNX model is not available
+
+---
+
+## 📧 Voicemail
+
+- Configurable voicemail boxes with custom TTS greetings
+- Automatic routing on no-answer timeout
+- Recording with configurable max duration and message count
+- Web dashboard playback and management
+- WAV storage in `.nogit/voicemail/`
+
+---
+
+## 🔢 IVR (Interactive Voice Response)
+
+- DTMF-navigable menus with configurable entries
+- Actions: route to extension, route to voicemail, transfer, submenu, hangup, repeat prompt
+- Custom TTS prompts per menu
+- Nested menu support

 ---

@@ -228,33 +295,42 @@ The `tts-engine` binary uses [Kokoro TTS](https://github.com/mzdk100/kokoro) (82

 | View | Description |
 |------|-------------|
-| **Overview** | Stats tiles — uptime, providers, devices, active calls |
-| **Calls** | Active calls with leg details, codec info, packet counters. Add/remove legs, transfer, hangup |
-| **Phone** | Browser softphone — mic/speaker selection, audio meters, dial pad, incoming call popup |
-| **Contacts** | Contact management with click-to-call |
-| **Providers** | SIP trunk config with registration status |
-| **Log** | Live streaming log viewer |
+| 📊 **Overview** | Stats tiles — uptime, providers, devices, active calls |
+| 📞 **Calls** | Active calls with leg details, codec info, add/remove legs, transfer, hangup |
+| ☎️ **Phone** | Browser softphone — mic/speaker selection, audio meters, dial pad, incoming call popup |
+| 🔀 **Routes** | Routing rule management — match/action model with priority |
+| 📧 **Voicemail** | Voicemail box management + message playback |
+| 🔢 **IVR** | IVR menu builder — DTMF entries, TTS prompts, nested menus |
+| 👤 **Contacts** | Contact management with click-to-call |
+| 🔌 **Providers** | SIP trunk configuration and registration status |
+| 📋 **Log** | Live streaming log viewer |

 ### REST API

 | Endpoint | Method | Description |
 |----------|--------|-------------|
-| `/api/status` | GET | Full system status (providers, devices, calls) |
+| `/api/status` | GET | Full system status (providers, devices, calls, history) |
 | `/api/call` | POST | Originate a call |
 | `/api/hangup` | POST | Hang up a call |
-| `/api/call/:id/addleg` | POST | Add a leg to an active call |
-| `/api/call/:id/addexternal` | POST | Add an external participant |
+| `/api/call/:id/addleg` | POST | Add a device leg to an active call |
+| `/api/call/:id/addexternal` | POST | Add an external participant via provider |
 | `/api/call/:id/removeleg` | POST | Remove a leg from a call |
 | `/api/transfer` | POST | Transfer a call |
-| `/api/config` | GET/POST | Read or update configuration (hot-reload) |
+| `/api/config` | GET | Read current configuration |
+| `/api/config` | POST | Update configuration (hot-reload) |
+| `/api/voicemail/:box` | GET | List voicemail messages |
+| `/api/voicemail/:box/:id` | DELETE | Delete a voicemail message |
+| `/api/voicemail/:box/:id/audio` | GET | Stream voicemail audio |

 ### WebSocket Events

 Connect to `/ws` for real-time push:

 ```jsonc
-{ "type": "status", "data": { ... } }       // Full status snapshot (1s interval)
+{ "type": "status", "data": { ... } }           // Full status snapshot (1s interval)
 { "type": "log", "data": { "message": "..." } } // Log lines in real-time
+{ "type": "incoming_call", "data": { ... } }     // Incoming call notification
+{ "type": "call_ended", "data": { ... } }        // Call ended notification
 ```

 ---
@@ -264,7 +340,7 @@ Connect to `/ws` for real-time push:
 | Port | Protocol | Purpose |
 |------|----------|---------|
 | 5070 (configurable) | UDP | SIP signaling |
-| 20000–20200 (configurable) | UDP | RTP relay (even ports, per-call allocation) |
+| 20000–20200 (configurable) | UDP | RTP media (even ports, per-call allocation) |
 | 3060 (configurable) | TCP | Web dashboard + WebSocket + REST API |

 ---
@@ -275,23 +351,16 @@ Connect to `/ws` for real-time push:
 # Start in dev mode
 pnpm start

-# Build Rust crates
+# Build Rust proxy-engine
 pnpm run buildRust

 # Bundle web frontend
 pnpm run bundle

-# Restart background server (build + bundle + restart)
+# Build + bundle + restart background server
 pnpm run restartBackground
 ```

-### Key Design Decisions
-
- **Hub Model** — Calls are N-leg hubs, not point-to-point. This enables multi-party, dynamic leg manipulation, and transfer without tearing down the call.
- **Zero-dependency SIP library** — `ts/sip/` is a pure data-level SIP stack (parse/build/mutate/serialize). No transport or timer logic — those live in the application layer.
- **Rust for the hot path** — Codec transcoding and noise suppression run in native Rust for real-time performance. TypeScript handles signaling and orchestration.
- **Per-session codec isolation** — Each call gets its own Opus/G.722 encoder/decoder state in the Rust process, preventing stateful codec prediction from leaking between concurrent calls.
-
 ---

 ## License and Legal Information
--- a/rust/Cargo.lock
+++ b/rust/Cargo.lock
@@ -237,6 +237,17 @@ version = "1.1.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "1505bd5d3d116872e7271a6d4e16d81d0c8570876c8de68093a09ac269d8aac0"

+[[package]]
+name = "atty"
+version = "0.2.14"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d9b39be18770d11421cdb1b9947a45dd3f37e93092cbf377614828a319d5fee8"
+dependencies = [
+ "hermit-abi",
+ "libc",
+ "winapi",
+]
+
 [[package]]
 name = "audiopus"
 version = "0.3.0-rc.0"
@@ -487,6 +498,31 @@ dependencies = [
 "inout",
 ]

+[[package]]
+name = "clap"
+version = "3.2.25"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "4ea181bf566f71cb9a5d17a59e1871af638180a18fb0035c92ae62b705207123"
+dependencies = [
+ "atty",
+ "bitflags 1.3.2",
+ "clap_lex",
+ "indexmap 1.9.3",
+ "once_cell",
+ "strsim",
+ "termcolor",
+ "textwrap",
+]
+
+[[package]]
+name = "clap_lex"
+version = "0.2.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2850f2f5a82cbf437dd5af4d49848fbdfc27c157c3d010345776f952765261c5"
+dependencies = [
+ "os_str_bytes",
+]
+
 [[package]]
 name = "cmake"
 version = "0.1.58"
@@ -700,6 +736,125 @@ version = "0.3.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "06d2e3287df1c007e74221c49ca10a95d557349e54b3a75dc2fb14712c751f04"

+[[package]]
+name = "dasp"
+version = "0.11.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7381b67da416b639690ac77c73b86a7b5e64a29e31d1f75fb3b1102301ef355a"
+dependencies = [
+ "dasp_envelope",
+ "dasp_frame",
+ "dasp_interpolate",
+ "dasp_peak",
+ "dasp_ring_buffer",
+ "dasp_rms",
+ "dasp_sample",
+ "dasp_signal",
+ "dasp_slice",
+ "dasp_window",
+]
+
+[[package]]
+name = "dasp_envelope"
+version = "0.11.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8ec617ce7016f101a87fe85ed44180839744265fae73bb4aa43e7ece1b7668b6"
+dependencies = [
+ "dasp_frame",
+ "dasp_peak",
+ "dasp_ring_buffer",
+ "dasp_rms",
+ "dasp_sample",
+]
+
+[[package]]
+name = "dasp_frame"
+version = "0.11.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b2a3937f5fe2135702897535c8d4a5553f8b116f76c1529088797f2eee7c5cd6"
+dependencies = [
+ "dasp_sample",
+]
+
+[[package]]
+name = "dasp_interpolate"
+version = "0.11.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7fc975a6563bb7ca7ec0a6c784ead49983a21c24835b0bc96eea11ee407c7486"
+dependencies = [
+ "dasp_frame",
+ "dasp_ring_buffer",
+ "dasp_sample",
+]
+
+[[package]]
+name = "dasp_peak"
+version = "0.11.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5cf88559d79c21f3d8523d91250c397f9a15b5fc72fbb3f87fdb0a37b79915bf"
+dependencies = [
+ "dasp_frame",
+ "dasp_sample",
+]
+
+[[package]]
+name = "dasp_ring_buffer"
+version = "0.11.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "07d79e19b89618a543c4adec9c5a347fe378a19041699b3278e616e387511ea1"
+
+[[package]]
+name = "dasp_rms"
+version = "0.11.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a6c5dcb30b7e5014486e2822537ea2beae50b19722ffe2ed7549ab03774575aa"
+dependencies = [
+ "dasp_frame",
+ "dasp_ring_buffer",
+ "dasp_sample",
+]
+
+[[package]]
+name = "dasp_sample"
+version = "0.11.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0c87e182de0887fd5361989c677c4e8f5000cd9491d6d563161a8f3a5519fc7f"
+
+[[package]]
+name = "dasp_signal"
+version = "0.11.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "aa1ab7d01689c6ed4eae3d38fe1cea08cba761573fbd2d592528d55b421077e7"
+dependencies = [
+ "dasp_envelope",
+ "dasp_frame",
+ "dasp_interpolate",
+ "dasp_peak",
+ "dasp_ring_buffer",
+ "dasp_rms",
+ "dasp_sample",
+ "dasp_window",
+]
+
+[[package]]
+name = "dasp_slice"
+version = "0.11.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "4e1c7335d58e7baedafa516cb361360ff38d6f4d3f9d9d5ee2a2fc8e27178fa1"
+dependencies = [
+ "dasp_frame",
+ "dasp_sample",
+]
+
+[[package]]
+name = "dasp_window"
+version = "0.11.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "99ded7b88821d2ce4e8b842c9f1c86ac911891ab89443cc1de750cae764c5076"
+dependencies = [
+ "dasp_sample",
+]
+
 [[package]]
 name = "data-encoding"
 version = "2.10.0"
@@ -1214,6 +1369,12 @@ dependencies = [
 "subtle",
 ]

+[[package]]
+name = "hashbrown"
+version = "0.12.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8a9ee70c43aaf417c914396645a0fa852624801b24ebb7ae78fe8272889ac888"
+
 [[package]]
 name = "hashbrown"
 version = "0.15.5"
@@ -1246,6 +1407,15 @@ version = "0.5.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "2304e00983f87ffb38b55b444b5e3b60a884b5d30c0fca7d82fe33449bbe55ea"

+[[package]]
+name = "hermit-abi"
+version = "0.1.19"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "62b467343b94ba476dcb2500d242dadbb39557df889310ac77c5d99100aaac33"
+dependencies = [
+ "libc",
+]
+
 [[package]]
 name = "hex"
 version = "0.4.3"
@@ -1446,6 +1616,16 @@ dependencies = [
 "zstd",
 ]

+[[package]]
+name = "indexmap"
+version = "1.9.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "bd070e393353796e801d209ad339e89596eb4c8d430d18ede6a1cced8fafbd99"
+dependencies = [
+ "autocfg",
+ "hashbrown 0.12.3",
+]
+
 [[package]]
 name = "indexmap"
 version = "2.14.0"
@@ -1739,7 +1919,13 @@ version = "0.5.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "805d5964d1e7a0006a7fdced7dae75084d66d18b35f1dfe81bd76929b1f8da0c"
 dependencies = [
+ "anyhow",
+ "clap",
+ "dasp",
+ "dasp_interpolate",
+ "dasp_ring_buffer",
 "easyfft",
+ "hound",
 "once_cell",
 ]

@@ -1905,6 +2091,12 @@ dependencies = [
 "ureq",
 ]

+[[package]]
+name = "os_str_bytes"
+version = "6.6.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e2355d85b9a3786f481747ced0e0ff2ba35213a1f9bd406ed906554d7af805a1"
+
 [[package]]
 name = "p256"
 version = "0.11.1"
@@ -2179,6 +2371,7 @@ dependencies = [
 "codec-lib",
 "hound",
 "kokoro-tts",
+ "nnnoiseless",
 "ort",
 "rand 0.8.5",
 "regex-lite",
@@ -2882,6 +3075,21 @@ dependencies = [
 "windows-sys",
 ]

+[[package]]
+name = "termcolor"
+version = "1.4.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "06794f8f6c5c898b3275aebefa6b8a1cb24cd2c6c79397ab15774837a0bc5755"
+dependencies = [
+ "winapi-util",
+]
+
+[[package]]
+name = "textwrap"
+version = "0.16.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c13547615a44dc9c452a8a534638acdf07120d4b6847c8178705da06306a3057"
+
 [[package]]
 name = "thiserror"
 version = "1.0.69"
@@ -3243,7 +3451,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "bb0e353e6a2fbdc176932bbaab493762eb1255a7900fe0fea1a2f96c296cc909"
 dependencies = [
 "anyhow",
- "indexmap",
+ "indexmap 2.14.0",
 "wasm-encoder",
 "wasmparser",
 ]
@@ -3256,7 +3464,7 @@ checksum = "47b807c72e1bac69382b3a6fb3dbe8ea4c0ed87ff5629b8685ae6b9a611028fe"
 dependencies = [
 "bitflags 2.11.0",
 "hashbrown 0.15.5",
- "indexmap",
+ "indexmap 2.14.0",
 "semver",
 ]

@@ -3514,6 +3722,15 @@ version = "0.4.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ac3b87c63620426dd9b991e5ce0329eff545bccbbb34f3be09ff6fb6ab51b7b6"

+[[package]]
+name = "winapi-util"
+version = "0.1.11"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c2a7b1c03c876122aa43f3020e6c3c3ee5c05081c9a00739faf7503aeba10d22"
+dependencies = [
+ "windows-sys",
+]
+
 [[package]]
 name = "winapi-x86_64-pc-windows-gnu"
 version = "0.4.0"
@@ -3563,7 +3780,7 @@ checksum = "b7c566e0f4b284dd6561c786d9cb0142da491f46a9fbed79ea69cdad5db17f21"
 dependencies = [
 "anyhow",
 "heck",
- "indexmap",
+ "indexmap 2.14.0",
 "prettyplease",
 "syn 2.0.117",
 "wasm-metadata",
@@ -3594,7 +3811,7 @@ checksum = "9d66ea20e9553b30172b5e831994e35fbde2d165325bec84fc43dbf6f4eb9cb2"
 dependencies = [
 "anyhow",
 "bitflags 2.11.0",
- "indexmap",
+ "indexmap 2.14.0",
 "log",
 "serde",
 "serde_derive",
@@ -3613,7 +3830,7 @@ checksum = "ecc8ac4bc1dc3381b7f59c34f00b67e18f910c2c0f50015669dde7def656a736"
 dependencies = [
 "anyhow",
 "id-arena",
- "indexmap",
+ "indexmap 2.14.0",
 "log",
 "semver",
 "serde",
--- a/rust/crates/codec-lib/Cargo.toml
+++ b/rust/crates/codec-lib/Cargo.toml
@@ -7,4 +7,4 @@ edition = "2021"
 audiopus = "0.3.0-rc.0"
 ezk-g722 = "0.1"
 rubato = "0.14"
-nnnoiseless = { version = "0.5", default-features = false }
+nnnoiseless = "0.5"
--- a/rust/crates/codec-lib/src/lib.rs
+++ b/rust/crates/codec-lib/src/lib.rs
@@ -104,6 +104,8 @@ pub struct TranscodeState {
    g722_dec: libg722::decoder::Decoder,
    /// Cached FFT resamplers keyed by (from_rate, to_rate, chunk_size).
    resamplers: HashMap<(u32, u32, usize), FftFixedIn<f64>>,
+    /// Cached f32 FFT resamplers keyed by (from_rate, to_rate, chunk_size).
+    resamplers_f32: HashMap<(u32, u32, usize), FftFixedIn<f32>>,
    /// ML noise suppression for the SIP-bound direction.
    denoiser_to_sip: Box<DenoiseState<'static>>,
    /// ML noise suppression for the browser-bound direction.
@@ -133,6 +135,7 @@ impl TranscodeState {
            g722_enc,
            g722_dec,
            resamplers: HashMap::new(),
+            resamplers_f32: HashMap::new(),
            denoiser_to_sip: DenoiseState::new(),
            denoiser_to_browser: DenoiseState::new(),
        })
@@ -293,6 +296,126 @@ impl TranscodeState {
            _ => Err(format!("unsupported target PT {pt}")),
        }
    }
+
+    // ---- f32 API for high-quality internal bus ----------------------------
+
+    /// Decode an encoded audio payload to f32 PCM samples in [-1.0, 1.0].
+    /// Returns (samples, sample_rate).
+    ///
+    /// For Opus, uses native float decode (no i16 quantization).
+    /// For G.722/G.711, decodes to i16 then converts (codec is natively i16).
+    pub fn decode_to_f32(&mut self, data: &[u8], pt: u8) -> Result<(Vec<f32>, u32), String> {
+        match pt {
+            PT_OPUS => {
+                let mut pcm = vec![0.0f32; 5760]; // up to 120ms at 48kHz
+                let packet =
+                    OpusPacket::try_from(data).map_err(|e| format!("opus packet: {e}"))?;
+                let out =
+                    MutSignals::try_from(&mut pcm[..]).map_err(|e| format!("opus signals: {e}"))?;
+                let n: usize = self
+                    .opus_dec
+                    .decode_float(Some(packet), out, false)
+                    .map_err(|e| format!("opus decode_float: {e}"))?
+                    .into();
+                pcm.truncate(n);
+                Ok((pcm, 48000))
+            }
+            _ => {
+                // G.722, PCMU, PCMA: natively i16 codecs — decode then convert.
+                let (pcm_i16, rate) = self.decode_to_pcm(data, pt)?;
+                let pcm_f32 = pcm_i16.iter().map(|&s| s as f32 / 32768.0).collect();
+                Ok((pcm_f32, rate))
+            }
+        }
+    }
+
+    /// Encode f32 PCM samples ([-1.0, 1.0]) to an audio codec.
+    ///
+    /// For Opus, uses native float encode (no i16 quantization).
+    /// For G.722/G.711, converts to i16 then encodes (codec is natively i16).
+    pub fn encode_from_f32(&mut self, pcm: &[f32], pt: u8) -> Result<Vec<u8>, String> {
+        match pt {
+            PT_OPUS => {
+                let mut buf = vec![0u8; 4000];
+                let n: usize = self
+                    .opus_enc
+                    .encode_float(pcm, &mut buf)
+                    .map_err(|e| format!("opus encode_float: {e}"))?
+                    .into();
+                buf.truncate(n);
+                Ok(buf)
+            }
+            _ => {
+                // G.722, PCMU, PCMA: natively i16 codecs.
+                let pcm_i16: Vec<i16> = pcm
+                    .iter()
+                    .map(|&s| (s * 32767.0).round().clamp(-32768.0, 32767.0) as i16)
+                    .collect();
+                self.encode_from_pcm(&pcm_i16, pt)
+            }
+        }
+    }
+
+    /// High-quality sample rate conversion for f32 PCM using rubato FFT resampler.
+    /// Uses a separate cache from the i16 resampler.
+    pub fn resample_f32(
+        &mut self,
+        pcm: &[f32],
+        from_rate: u32,
+        to_rate: u32,
+    ) -> Result<Vec<f32>, String> {
+        if from_rate == to_rate || pcm.is_empty() {
+            return Ok(pcm.to_vec());
+        }
+
+        let chunk = pcm.len();
+        let key = (from_rate, to_rate, chunk);
+
+        if !self.resamplers_f32.contains_key(&key) {
+            let r =
+                FftFixedIn::<f32>::new(from_rate as usize, to_rate as usize, chunk, 1, 1)
+                    .map_err(|e| format!("resampler f32 {from_rate}->{to_rate}: {e}"))?;
+            self.resamplers_f32.insert(key, r);
+        }
+        let resampler = self.resamplers_f32.get_mut(&key).unwrap();
+
+        let input = vec![pcm.to_vec()];
+        let result = resampler
+            .process(&input, None)
+            .map_err(|e| format!("resample f32 {from_rate}->{to_rate}: {e}"))?;
+
+        Ok(result[0].clone())
+    }
+
+    /// Apply RNNoise ML noise suppression to 48kHz f32 PCM audio.
+    /// Processes in 480-sample (10ms) frames. State persists across calls.
+    /// Operates natively in f32 — no i16 conversion overhead.
+    pub fn denoise_f32(denoiser: &mut DenoiseState, pcm: &[f32]) -> Vec<f32> {
+        let frame_size = DenoiseState::FRAME_SIZE; // 480
+        let total = pcm.len();
+        let whole = (total / frame_size) * frame_size;
+        let mut output = Vec::with_capacity(total);
+        let mut out_buf = [0.0f32; 480];
+
+        // nnnoiseless expects f32 samples scaled as i16 range (-32768..32767).
+        for offset in (0..whole).step_by(frame_size) {
+            let input: Vec<f32> = pcm[offset..offset + frame_size]
+                .iter()
+                .map(|&s| s * 32768.0)
+                .collect();
+            denoiser.process_frame(&mut out_buf, &input);
+            output.extend(out_buf.iter().map(|&s| s / 32768.0));
+        }
+        if whole < total {
+            output.extend_from_slice(&pcm[whole..]);
+        }
+        output
+    }
+}
+
+/// Create a new standalone denoiser for per-leg inbound processing.
+pub fn new_denoiser() -> Box<DenoiseState<'static>> {
+    DenoiseState::new()
 }

 #[cfg(test)]
--- a/rust/crates/proxy-engine/Cargo.toml
+++ b/rust/crates/proxy-engine/Cargo.toml
@@ -10,6 +10,7 @@ path = "src/main.rs"
 [dependencies]
 codec-lib = { path = "../codec-lib" }
 sip-proto = { path = "../sip-proto" }
+nnnoiseless = "0.5"
 tokio = { version = "1", features = ["full"] }
 serde = { version = "1", features = ["derive"] }
 serde_json = "1"
--- a/rust/crates/proxy-engine/src/audio_player.rs
+++ b/rust/crates/proxy-engine/src/audio_player.rs
@@ -10,9 +10,9 @@ use tokio::net::UdpSocket;
 use tokio::time::{self, Duration};

 /// Mixing sample rate used by the mixer (must stay in sync with mixer::MIX_RATE).
-const MIX_RATE: u32 = 16000;
+const MIX_RATE: u32 = 48000;
 /// Samples per 20ms frame at the mixing rate.
-const MIX_FRAME_SIZE: usize = 320;
+const MIX_FRAME_SIZE: usize = 960;

 /// Play a WAV file as RTP to a destination.
 /// Returns when playback is complete.
@@ -178,9 +178,9 @@ pub async fn play_beep(
    Ok((seq, ts))
 }

-/// Load a WAV file and split it into 20ms PCM frames at 16kHz.
+/// Load a WAV file and split it into 20ms f32 PCM frames at 48kHz.
 /// Used by the leg interaction system to prepare prompt audio for the mixer.
-pub fn load_prompt_pcm_frames(wav_path: &str) -> Result<Vec<Vec<i16>>, String> {
+pub fn load_prompt_pcm_frames(wav_path: &str) -> Result<Vec<Vec<f32>>, String> {
    let path = Path::new(wav_path);
    if !path.exists() {
        return Err(format!("WAV file not found: {wav_path}"));
@@ -191,17 +191,17 @@ pub fn load_prompt_pcm_frames(wav_path: &str) -> Result<Vec<Vec<i16>>, String> {
    let spec = reader.spec();
    let wav_rate = spec.sample_rate;

-    // Read all samples as i16.
-    let samples: Vec<i16> = if spec.bits_per_sample == 16 {
+    // Read all samples as f32 in [-1.0, 1.0].
+    let samples: Vec<f32> = if spec.bits_per_sample == 16 {
        reader
            .samples::<i16>()
            .filter_map(|s| s.ok())
+            .map(|s| s as f32 / 32768.0)
            .collect()
    } else if spec.bits_per_sample == 32 && spec.sample_format == hound::SampleFormat::Float {
        reader
            .samples::<f32>()
            .filter_map(|s| s.ok())
-            .map(|s| (s * 32767.0).round().clamp(-32768.0, 32767.0) as i16)
            .collect()
    } else {
        return Err(format!(
@@ -214,24 +214,24 @@ pub fn load_prompt_pcm_frames(wav_path: &str) -> Result<Vec<Vec<i16>>, String> {
        return Ok(vec![]);
    }

-    // Resample to MIX_RATE (16kHz) if needed.
+    // Resample to MIX_RATE (48kHz) if needed.
    let resampled = if wav_rate != MIX_RATE {
        let mut transcoder = TranscodeState::new().map_err(|e| format!("codec init: {e}"))?;
        transcoder
-            .resample(&samples, wav_rate, MIX_RATE)
+            .resample_f32(&samples, wav_rate, MIX_RATE)
            .map_err(|e| format!("resample: {e}"))?
    } else {
        samples
    };

-    // Split into MIX_FRAME_SIZE (320) sample frames.
+    // Split into MIX_FRAME_SIZE (960) sample frames.
    let mut frames = Vec::new();
    let mut offset = 0;
    while offset < resampled.len() {
        let end = (offset + MIX_FRAME_SIZE).min(resampled.len());
        let mut frame = resampled[offset..end].to_vec();
        // Pad short final frame with silence.
-        frame.resize(MIX_FRAME_SIZE, 0);
+        frame.resize(MIX_FRAME_SIZE, 0.0);
        frames.push(frame);
        offset += MIX_FRAME_SIZE;
    }
--- a/rust/crates/proxy-engine/src/call_manager.rs
+++ b/rust/crates/proxy-engine/src/call_manager.rs
@@ -20,6 +20,35 @@ use std::net::SocketAddr;
 use std::sync::Arc;
 use tokio::net::UdpSocket;

+/// Emit a `leg_added` event with full leg information.
+/// Free function (not a method) to avoid `&self` borrow conflicts when `self.calls` is borrowed.
+fn emit_leg_added_event(tx: &OutTx, call_id: &str, leg: &LegInfo) {
+    let metadata: serde_json::Value = if leg.metadata.is_empty() {
+        serde_json::json!({})
+    } else {
+        serde_json::Value::Object(
+            leg.metadata
+                .iter()
+                .map(|(k, v)| (k.clone(), v.clone()))
+                .collect(),
+        )
+    };
+    emit_event(
+        tx,
+        "leg_added",
+        serde_json::json!({
+            "call_id": call_id,
+            "leg_id": leg.id,
+            "kind": leg.kind.as_str(),
+            "state": leg.state.as_str(),
+            "codec": sip_proto::helpers::codec_name(leg.codec_pt),
+            "rtpPort": leg.rtp_port,
+            "remoteMedia": leg.remote_media.map(|a| format!("{}:{}", a.ip(), a.port())),
+            "metadata": metadata,
+        }),
+    );
+}
+
 pub struct CallManager {
    /// All active calls, keyed by internal call ID.
    pub calls: HashMap<String, Call>,
@@ -120,7 +149,19 @@ impl CallManager {
        }

        // Passthrough-style routing for inbound/outbound device↔provider calls.
-        self.route_passthrough_message(&call_id, &leg_id, msg, from_addr, socket, config)
+        // The sip_index only stores one leg for shared Call-IDs, so we need to
+        // determine which leg the message actually belongs to by comparing from_addr.
+        let actual_leg_id = self
+            .calls
+            .get(&call_id)
+            .and_then(|call| {
+                call.legs
+                    .values()
+                    .find(|l| l.signaling_addr == Some(from_addr))
+                    .map(|l| l.id.clone())
+            })
+            .unwrap_or(leg_id);
+        self.route_passthrough_message(&call_id, &actual_leg_id, msg, from_addr, socket, config)
            .await
    }

@@ -155,7 +196,17 @@ impl CallManager {
        };
        // Mutable borrow on call/leg is now released.

-        let sip_pt = codecs.first().copied().unwrap_or(9);
+        let mut sip_pt = codecs.first().copied().unwrap_or(9);
+
+        // If the message has SDP (e.g., 200 OK answer), use the negotiated codec
+        // instead of the offered one.
+        if msg.has_sdp_body() {
+            if let Some(ep) = parse_sdp_endpoint(&msg.body) {
+                if let Some(pt) = ep.codec_pt {
+                    sip_pt = pt;
+                }
+            }
+        }

        match action {
            SipLegAction::None => {}
@@ -253,14 +304,27 @@ impl CallManager {
                                    dev_leg.state = LegState::Connected;
                                }
                            }
+                            emit_event(
+                                &self.out_tx,
+                                "leg_state_changed",
+                                serde_json::json!({ "call_id": call_id, "leg_id": dev_leg_id, "state": "connected" }),
+                            );

                            // Wire device leg to mixer.
+                            // Use the device's preferred codec from its INVITE SDP,
+                            // not the provider's negotiated codec.
+                            let dev_pt = device_invite
+                                .has_sdp_body()
+                                .then(|| parse_sdp_endpoint(&device_invite.body))
+                                .flatten()
+                                .and_then(|ep| ep.codec_pt)
+                                .unwrap_or(sip_pt);
                            if let Some(dev_remote_addr) = dev_remote {
                                let dev_channels = create_leg_channels();
                                spawn_sip_inbound(dev_rtp_socket.clone(), dev_channels.inbound_tx);
                                spawn_sip_outbound(dev_rtp_socket, dev_remote_addr, dev_channels.outbound_rx);
                                if let Some(call) = self.calls.get(call_id) {
-                                    call.add_leg_to_mixer(&dev_leg_id, sip_pt, dev_channels.inbound_rx, dev_channels.outbound_tx)
+                                    call.add_leg_to_mixer(&dev_leg_id, dev_pt, dev_channels.inbound_rx, dev_channels.outbound_tx)
                                        .await;
                                }
                            }
@@ -312,6 +376,8 @@ impl CallManager {
                        leg.state = LegState::Terminated;
                    }
                }
+                emit_event(&self.out_tx, "leg_state_changed",
+                    serde_json::json!({ "call_id": call_id, "leg_id": leg_id, "state": "terminated" }));
                emit_event(&self.out_tx, "call_ended",
                    serde_json::json!({ "call_id": call_id, "reason": reason, "duration": duration }));
                self.terminate_call(call_id).await;
@@ -517,21 +583,30 @@ impl CallManager {
                    if let Some(leg) = call.legs.get_mut(this_leg_id) {
                        leg.state = LegState::Ringing;
                    }
+                    emit_event(&self.out_tx, "leg_state_changed",
+                        serde_json::json!({ "call_id": call_id, "leg_id": this_leg_id, "state": "ringing" }));
                } else if code >= 200 && code < 300 {
                    let mut needs_wiring = false;
                    if let Some(leg) = call.legs.get_mut(this_leg_id) {
                        leg.state = LegState::Connected;
-                        // Learn remote media from SDP.
+                        // Learn remote media and negotiated codec from SDP answer.
                        if msg.has_sdp_body() {
                            if let Some(ep) = parse_sdp_endpoint(&msg.body) {
                                if let Ok(addr) = format!("{}:{}", ep.address, ep.port).parse() {
                                    leg.remote_media = Some(addr);
                                }
+                                // Use the codec from the SDP answer (what the remote actually selected).
+                                if let Some(pt) = ep.codec_pt {
+                                    leg.codec_pt = pt;
+                                }
                            }
                        }
                        needs_wiring = true;
                    }

+                    emit_event(&self.out_tx, "leg_state_changed",
+                        serde_json::json!({ "call_id": call_id, "leg_id": this_leg_id, "state": "connected" }));
+
                    if call.state != CallState::Connected {
                        call.state = CallState::Connected;
                        emit_event(&self.out_tx, "call_answered", serde_json::json!({ "call_id": call_id }));
@@ -677,15 +752,19 @@ impl CallManager {
        call.callee_number = Some(called_number);
        call.state = CallState::Ringing;

-        let codec_pt = provider_config.codecs.first().copied().unwrap_or(9);
+        let mut codec_pt = provider_config.codecs.first().copied().unwrap_or(9);

-        // Provider leg — extract media from SDP.
+        // Provider leg — extract media and negotiated codec from SDP.
        let mut provider_media: Option<SocketAddr> = None;
        if invite.has_sdp_body() {
            if let Some(ep) = parse_sdp_endpoint(&invite.body) {
                if let Ok(addr) = format!("{}:{}", ep.address, ep.port).parse() {
                    provider_media = Some(addr);
                }
+                // Use the codec from the provider's SDP offer (what they actually want to use).
+                if let Some(pt) = ep.codec_pt {
+                    codec_pt = pt;
+                }
            }
        }

@@ -755,6 +834,16 @@ impl CallManager {
        // Store the call.
        self.calls.insert(call_id.clone(), call);

+        // Emit leg_added for both initial legs.
+        if let Some(call) = self.calls.get(&call_id) {
+            if let Some(leg) = call.legs.get(&provider_leg_id) {
+                emit_leg_added_event(&self.out_tx, &call_id, leg);
+            }
+            if let Some(leg) = call.legs.get(&device_leg_id) {
+                emit_leg_added_event(&self.out_tx, &call_id, leg);
+            }
+        }
+
        Some(call_id)
    }

@@ -842,6 +931,14 @@ impl CallManager {
            .insert(sip_call_id, (call_id.clone(), leg_id));

        self.calls.insert(call_id.clone(), call);
+
+        // Emit leg_added for the provider leg.
+        if let Some(call) = self.calls.get(&call_id) {
+            for leg in call.legs.values() {
+                emit_leg_added_event(&self.out_tx, &call_id, leg);
+            }
+        }
+
        Some(call_id)
    }

@@ -866,11 +963,18 @@ impl CallManager {
        let lan_port = config.proxy.lan_port;
        let device_sip_call_id = invite.call_id().to_string();

+        // Extract just the user part from the request URI (e.g., "sip:16196000@10.0.0.1" → "16196000").
+        // extract_uri is for header values with angle brackets, not bare request URIs.
        let dialed_number = invite
            .request_uri()
-            .and_then(|uri| SipMessage::extract_uri(uri))
-            .unwrap_or(invite.request_uri().unwrap_or(""))
-            .to_string();
+            .map(|uri| {
+                let stripped = uri
+                    .strip_prefix("sip:")
+                    .or_else(|| uri.strip_prefix("sips:"))
+                    .unwrap_or(uri);
+                stripped.split('@').next().unwrap_or(stripped).to_string()
+            })
+            .unwrap_or_default();

        let provider_dest: SocketAddr = match provider_config.outbound_proxy.to_socket_addr() {
            Some(a) => a,
@@ -983,6 +1087,14 @@ impl CallManager {
            .insert(provider_sip_call_id, (call_id.clone(), provider_leg_id));

        self.calls.insert(call_id.clone(), call);
+
+        // Emit leg_added for both initial legs (device + provider).
+        if let Some(call) = self.calls.get(&call_id) {
+            for leg in call.legs.values() {
+                emit_leg_added_event(&self.out_tx, &call_id, leg);
+            }
+        }
+
        Some(call_id)
    }

@@ -1050,17 +1162,11 @@ impl CallManager {
        let call = self.calls.get_mut(call_id).unwrap();
        call.legs.insert(leg_id.clone(), leg_info);

-        emit_event(
-            &self.out_tx,
-            "leg_added",
-            serde_json::json!({
-                "call_id": call_id,
-                "leg_id": leg_id,
-                "kind": "sip-provider",
-                "state": "inviting",
-                "number": number,
-            }),
-        );
+        if let Some(call) = self.calls.get(call_id) {
+            if let Some(leg) = call.legs.get(&leg_id) {
+                emit_leg_added_event(&self.out_tx, call_id, leg);
+            }
+        }

        Some(leg_id)
    }
@@ -1126,17 +1232,11 @@ impl CallManager {
        let call = self.calls.get_mut(call_id).unwrap();
        call.legs.insert(leg_id.clone(), leg_info);

-        emit_event(
-            &self.out_tx,
-            "leg_added",
-            serde_json::json!({
-                "call_id": call_id,
-                "leg_id": leg_id,
-                "kind": "sip-device",
-                "state": "inviting",
-                "device_id": device_id,
-            }),
-        );
+        if let Some(call) = self.calls.get(call_id) {
+            if let Some(leg) = call.legs.get(&leg_id) {
+                emit_leg_added_event(&self.out_tx, call_id, leg);
+            }
+        }

        Some(leg_id)
    }
@@ -1223,6 +1323,13 @@ impl CallManager {
            None => return false,
        };

+        // Emit leg_removed for source call.
+        emit_event(
+            &self.out_tx,
+            "leg_removed",
+            serde_json::json!({ "call_id": source_call_id, "leg_id": leg_id }),
+        );
+
        // Update SIP index to point to the target call.
        if let Some(sip_cid) = &leg_info.sip_call_id {
            self.sip_index.insert(
@@ -1255,15 +1362,12 @@ impl CallManager {
        let target_call = self.calls.get_mut(target_call_id).unwrap();
        target_call.legs.insert(leg_id.to_string(), leg_info);

-        emit_event(
-            &self.out_tx,
-            "leg_transferred",
-            serde_json::json!({
-                "leg_id": leg_id,
-                "source_call_id": source_call_id,
-                "target_call_id": target_call_id,
-            }),
-        );
+        // Emit leg_added for target call.
+        if let Some(target) = self.calls.get(target_call_id) {
+            if let Some(leg) = target.legs.get(leg_id) {
+                emit_leg_added_event(&self.out_tx, target_call_id, leg);
+            }
+        }

        // Check if source call has too few legs remaining.
        let source_call = self.calls.get(source_call_id).unwrap();
@@ -1366,6 +1470,11 @@ impl CallManager {
                }
            }
            leg.state = LegState::Terminated;
+            emit_event(
+                &self.out_tx,
+                "leg_state_changed",
+                serde_json::json!({ "call_id": call_id, "leg_id": leg.id, "state": "terminated" }),
+            );
        }

        emit_event(
@@ -1484,6 +1593,13 @@ impl CallManager {
        );
        self.calls.insert(call_id.to_string(), call);

+        // Emit leg_added for the provider leg.
+        if let Some(call) = self.calls.get(call_id) {
+            for leg in call.legs.values() {
+                emit_leg_added_event(&self.out_tx, call_id, leg);
+            }
+        }
+
        // Build recording path.
        let timestamp = std::time::SystemTime::now()
            .duration_since(std::time::UNIX_EPOCH)
--- a/rust/crates/proxy-engine/src/leg_io.rs
+++ b/rust/crates/proxy-engine/src/leg_io.rs
@@ -35,7 +35,8 @@ pub fn create_leg_channels() -> LegChannels {
 }

 /// Spawn the inbound I/O task for a SIP leg.
-/// Reads RTP from the socket, strips the 12-byte header, sends payload to the mixer.
+/// Reads RTP from the socket, parses the variable-length header (RFC 3550),
+/// and sends the payload to the mixer.
 /// Returns the JoinHandle (exits when the inbound_tx channel is dropped).
 pub fn spawn_sip_inbound(
    rtp_socket: Arc<UdpSocket>,
@@ -51,12 +52,29 @@ pub fn spawn_sip_inbound(
                    }
                    let pt = buf[1] & 0x7F;
                    let marker = (buf[1] & 0x80) != 0;
+                    let seq = u16::from_be_bytes([buf[2], buf[3]]);
                    let timestamp = u32::from_be_bytes([buf[4], buf[5], buf[6], buf[7]]);
-                    let payload = buf[12..n].to_vec();
+
+                    // RFC 3550: header length = 12 + (CC * 4) + optional extension.
+                    let cc = (buf[0] & 0x0F) as usize;
+                    let has_extension = (buf[0] & 0x10) != 0;
+                    let mut offset = 12 + cc * 4;
+                    if has_extension {
+                        if offset + 4 > n {
+                            continue; // Malformed: extension header truncated.
+                        }
+                        let ext_len = u16::from_be_bytes([buf[offset + 2], buf[offset + 3]]) as usize;
+                        offset += 4 + ext_len * 4;
+                    }
+                    if offset >= n {
+                        continue; // No payload after header.
+                    }
+
+                    let payload = buf[offset..n].to_vec();
                    if payload.is_empty() {
                        continue;
                    }
-                    if inbound_tx.send(RtpPacket { payload, payload_type: pt, marker, timestamp }).await.is_err() {
+                    if inbound_tx.send(RtpPacket { payload, payload_type: pt, marker, seq, timestamp }).await.is_err() {
                        break; // Channel closed — leg removed.
                    }
                }
--- a/rust/crates/proxy-engine/src/main.rs
+++ b/rust/crates/proxy-engine/src/main.rs
@@ -677,6 +677,10 @@ async fn handle_webrtc_link(
            "leg_id": session_id,
            "kind": "webrtc",
            "state": "connected",
+            "codec": "Opus",
+            "rtpPort": 0,
+            "remoteMedia": null,
+            "metadata": {},
        }));

        respond_ok(out_tx, &cmd.id, serde_json::json!({
@@ -1125,8 +1129,11 @@ async fn handle_add_tool_leg(
            "call_id": call_id,
            "leg_id": tool_leg_id,
            "kind": "tool",
-            "tool_type": tool_type_str,
            "state": "connected",
+            "codec": null,
+            "rtpPort": 0,
+            "remoteMedia": null,
+            "metadata": { "tool_type": tool_type_str },
        }),
    );

--- a/rust/crates/proxy-engine/src/mixer.rs
+++ b/rust/crates/proxy-engine/src/mixer.rs
@@ -3,9 +3,12 @@
 //! Each Call spawns one mixer task. Legs communicate with the mixer via
 //! tokio mpsc channels — no shared mutable state, no lock contention.
 //!
+//! Internal bus format: 48kHz f32 PCM (960 samples per 20ms frame).
+//! All encoding/decoding happens at leg boundaries. Per-leg inbound denoising at 48kHz.
+//!
 //! The mixer runs a 20ms tick loop:
-//! 1. Drain inbound channels, decode to PCM, resample to 16kHz
-//! 2. Compute total mix (sum of all **participant** legs' PCM as i32)
+//! 1. Drain inbound channels, decode to f32, resample to 48kHz, denoise per-leg
+//! 2. Compute total mix (sum of all **participant** legs' f32 PCM as f64)
 //! 3. For each participant leg: mix-minus = total - own, resample to leg codec rate, encode, send
 //! 4. For each isolated leg: play prompt frame or silence, check DTMF
 //! 5. For each tool leg: send per-source unmerged audio batch
@@ -13,16 +16,18 @@

 use crate::ipc::{emit_event, OutTx};
 use crate::rtp::{build_rtp_header, rtp_clock_increment};
-use codec_lib::{codec_sample_rate, TranscodeState};
+use codec_lib::{codec_sample_rate, new_denoiser, TranscodeState};
+use nnnoiseless::DenoiseState;
 use std::collections::{HashMap, VecDeque};
 use tokio::sync::{mpsc, oneshot};
 use tokio::task::JoinHandle;
 use tokio::time::{self, Duration, MissedTickBehavior};

-/// Mixing sample rate — 16kHz. G.722 is native, G.711 needs 2× upsample, Opus needs 3× downsample.
-const MIX_RATE: u32 = 16000;
+/// Mixing sample rate — 48kHz. Opus is native, G.722 needs 3× upsample, G.711 needs 6× upsample.
+/// All processing (denoising, mixing) happens at this rate in f32 for maximum quality.
+const MIX_RATE: u32 = 48000;
 /// Samples per 20ms frame at the mixing rate.
-const MIX_FRAME_SIZE: usize = 320; // 16000 * 0.020
+const MIX_FRAME_SIZE: usize = 960; // 48000 * 0.020

 /// A raw RTP payload received from a leg (no RTP header).
 pub struct RtpPacket {
@@ -30,6 +35,8 @@ pub struct RtpPacket {
    pub payload_type: u8,
    /// RTP marker bit (first packet of a DTMF event, etc.).
    pub marker: bool,
+    /// RTP sequence number for reordering.
+    pub seq: u16,
    /// RTP timestamp from the original packet header.
    pub timestamp: u32,
 }
@@ -47,8 +54,8 @@ enum LegRole {
 }

 struct IsolationState {
-    /// PCM frames at MIX_RATE (320 samples each) queued for playback.
-    prompt_frames: VecDeque<Vec<i16>>,
+    /// PCM frames at MIX_RATE (960 samples each, 48kHz f32) queued for playback.
+    prompt_frames: VecDeque<Vec<f32>>,
    /// Digits that complete the interaction (e.g., ['1', '2']).
    expected_digits: Vec<char>,
    /// Ticks remaining before timeout (decremented each tick after prompt ends).
@@ -88,8 +95,8 @@ pub struct ToolAudioBatch {
 /// One participant's 20ms audio frame.
 pub struct ToolAudioSource {
    pub leg_id: String,
-    /// PCM at 16kHz, MIX_FRAME_SIZE (320) samples.
-    pub pcm_16k: Vec<i16>,
+    /// PCM at 48kHz f32, MIX_FRAME_SIZE (960) samples.
+    pub pcm_48k: Vec<f32>,
 }

 /// Internal storage for a tool leg inside the mixer.
@@ -122,8 +129,8 @@ pub enum MixerCommand {
    /// DTMF from the leg is checked against expected_digits.
    StartInteraction {
        leg_id: String,
-        /// PCM frames at MIX_RATE (16kHz), each 320 samples.
-        prompt_pcm_frames: Vec<Vec<i16>>,
+        /// PCM frames at MIX_RATE (48kHz f32), each 960 samples.
+        prompt_pcm_frames: Vec<Vec<f32>>,
        expected_digits: Vec<char>,
        timeout_ms: u32,
        result_tx: oneshot::Sender<InteractionResult>,
@@ -149,10 +156,12 @@ pub enum MixerCommand {
 struct MixerLegSlot {
    codec_pt: u8,
    transcoder: TranscodeState,
+    /// Per-leg inbound denoiser (48kHz, 480-sample frames).
+    denoiser: Box<DenoiseState<'static>>,
    inbound_rx: mpsc::Receiver<RtpPacket>,
    outbound_tx: mpsc::Sender<Vec<u8>>,
-    /// Last decoded PCM frame at MIX_RATE (320 samples). Used for mix-minus.
-    last_pcm_frame: Vec<i16>,
+    /// Last decoded+denoised PCM frame at MIX_RATE (960 samples, 48kHz f32).
+    last_pcm_frame: Vec<f32>,
    /// Number of consecutive ticks with no inbound packet.
    silent_ticks: u32,
    // RTP output state.
@@ -220,9 +229,10 @@ async fn mixer_loop(
                        MixerLegSlot {
                            codec_pt,
                            transcoder,
+                            denoiser: new_denoiser(),
                            inbound_rx,
                            outbound_tx,
-                            last_pcm_frame: vec![0i16; MIX_FRAME_SIZE],
+                            last_pcm_frame: vec![0.0f32; MIX_FRAME_SIZE],
                            silent_ticks: 0,
                            rtp_seq: 0,
                            rtp_ts: 0,
@@ -311,16 +321,18 @@ async fn mixer_loop(
            continue;
        }

-        // ── 2. Drain inbound packets, decode to 16kHz PCM. ─────────
+        // ── 2. Drain inbound packets, decode to 48kHz f32 PCM. ────
        //    DTMF (PT 101) packets are collected separately.
+        //    Audio packets are sorted by sequence number and decoded
+        //    in order to maintain codec state (critical for G.722 ADPCM).
        let leg_ids: Vec<String> = legs.keys().cloned().collect();
        let mut dtmf_forward: Vec<(String, RtpPacket)> = Vec::new();

        for lid in &leg_ids {
            let slot = legs.get_mut(lid).unwrap();

-            // Drain channel — collect DTMF packets separately, keep latest audio.
-            let mut latest_audio: Option<RtpPacket> = None;
+            // Drain channel — collect DTMF separately, collect ALL audio packets.
+            let mut audio_packets: Vec<RtpPacket> = Vec::new();
            loop {
                match slot.inbound_rx.try_recv() {
                    Ok(pkt) => {
@@ -328,33 +340,47 @@ async fn mixer_loop(
                            // DTMF telephone-event: collect for processing.
                            dtmf_forward.push((lid.clone(), pkt));
                        } else {
-                            latest_audio = Some(pkt);
+                            audio_packets.push(pkt);
                        }
                    }
                    Err(_) => break,
                }
            }

-            if let Some(pkt) = latest_audio {
+            if !audio_packets.is_empty() {
                slot.silent_ticks = 0;
-                match slot.transcoder.decode_to_pcm(&pkt.payload, pkt.payload_type) {
-                    Ok((pcm, rate)) => {
-                        // Resample to mixing rate if needed.
-                        let pcm_mix = if rate == MIX_RATE {
-                            pcm
-                        } else {
-                            slot.transcoder
-                                .resample(&pcm, rate, MIX_RATE)
-                                .unwrap_or_else(|_| vec![0i16; MIX_FRAME_SIZE])
-                        };
-                        // Pad or truncate to exactly MIX_FRAME_SIZE.
-                        let mut frame = pcm_mix;
-                        frame.resize(MIX_FRAME_SIZE, 0);
-                        slot.last_pcm_frame = frame;
-                    }
-                    Err(_) => {
-                        // Decode failed — use silence.
-                        slot.last_pcm_frame = vec![0i16; MIX_FRAME_SIZE];
+
+                // Sort by sequence number for correct codec state progression.
+                // This prevents G.722 ADPCM state corruption from out-of-order packets.
+                audio_packets.sort_by_key(|p| p.seq);
+
+                // Decode ALL packets in order (maintains codec state),
+                // but only keep the last decoded frame for mixing.
+                for pkt in &audio_packets {
+                    match slot.transcoder.decode_to_f32(&pkt.payload, pkt.payload_type) {
+                        Ok((pcm, rate)) => {
+                            // Resample to 48kHz mixing rate if needed.
+                            let pcm_48k = if rate == MIX_RATE {
+                                pcm
+                            } else {
+                                slot.transcoder
+                                    .resample_f32(&pcm, rate, MIX_RATE)
+                                    .unwrap_or_else(|_| vec![0.0f32; MIX_FRAME_SIZE])
+                            };
+                            // Per-leg inbound denoising at 48kHz.
+                            // Only for SIP telephony legs — WebRTC browsers
+                            // already apply noise suppression via getUserMedia.
+                            let processed = if slot.codec_pt != codec_lib::PT_OPUS {
+                                TranscodeState::denoise_f32(&mut slot.denoiser, &pcm_48k)
+                            } else {
+                                pcm_48k
+                            };
+                            // Pad or truncate to exactly MIX_FRAME_SIZE.
+                            let mut frame = processed;
+                            frame.resize(MIX_FRAME_SIZE, 0.0);
+                            slot.last_pcm_frame = frame;
+                        }
+                        Err(_) => {}
                    }
                }
            } else if dtmf_forward.iter().any(|(src, _)| src == lid) {
@@ -364,17 +390,18 @@ async fn mixer_loop(
                slot.silent_ticks += 1;
                // After 150 ticks (3 seconds) of silence, zero out to avoid stale audio.
                if slot.silent_ticks > 150 {
-                    slot.last_pcm_frame = vec![0i16; MIX_FRAME_SIZE];
+                    slot.last_pcm_frame = vec![0.0f32; MIX_FRAME_SIZE];
                }
            }
        }

        // ── 3. Compute total mix from PARTICIPANT legs only. ────────
-        let mut total_mix = vec![0i32; MIX_FRAME_SIZE];
+        //    Accumulate as f64 to prevent precision loss when summing f32.
+        let mut total_mix = vec![0.0f64; MIX_FRAME_SIZE];
        for slot in legs.values() {
            if matches!(slot.role, LegRole::Participant) {
                for (i, &s) in slot.last_pcm_frame.iter().enumerate().take(MIX_FRAME_SIZE) {
-                    total_mix[i] += s as i32;
+                    total_mix[i] += s as f64;
                }
            }
        }
@@ -387,27 +414,27 @@ async fn mixer_loop(
        for (lid, slot) in legs.iter_mut() {
            match &mut slot.role {
                LegRole::Participant => {
-                    // Mix-minus: total minus this leg's own contribution.
+                    // Mix-minus: total minus this leg's own contribution, clamped to [-1.0, 1.0].
                    let mut mix_minus = Vec::with_capacity(MIX_FRAME_SIZE);
                    for i in 0..MIX_FRAME_SIZE {
-                        let sample = (total_mix[i] - slot.last_pcm_frame[i] as i32)
-                            .clamp(-32768, 32767) as i16;
-                        mix_minus.push(sample);
+                        let sample =
+                            (total_mix[i] - slot.last_pcm_frame[i] as f64) as f32;
+                        mix_minus.push(sample.clamp(-1.0, 1.0));
                    }

-                    // Resample from 16kHz to the leg's codec native rate.
+                    // Resample from 48kHz to the leg's codec native rate.
                    let target_rate = codec_sample_rate(slot.codec_pt);
                    let resampled = if target_rate == MIX_RATE {
                        mix_minus
                    } else {
                        slot.transcoder
-                            .resample(&mix_minus, MIX_RATE, target_rate)
+                            .resample_f32(&mix_minus, MIX_RATE, target_rate)
                            .unwrap_or_default()
                    };

-                    // Encode to the leg's codec.
+                    // Encode to the leg's codec (f32 → i16 → codec inside encode_from_f32).
                    let encoded =
-                        match slot.transcoder.encode_from_pcm(&resampled, slot.codec_pt) {
+                        match slot.transcoder.encode_from_f32(&resampled, slot.codec_pt) {
                            Ok(e) if !e.is_empty() => e,
                            _ => continue,
                        };
@@ -456,21 +483,21 @@ async fn mixer_loop(
                            frame
                        } else {
                            state.prompt_done = true;
-                            vec![0i16; MIX_FRAME_SIZE]
+                            vec![0.0f32; MIX_FRAME_SIZE]
                        };

-                        // Encode prompt frame to the leg's codec (reuses existing encode path).
+                        // Encode prompt frame to the leg's codec.
                        let target_rate = codec_sample_rate(slot.codec_pt);
                        let resampled = if target_rate == MIX_RATE {
                            pcm_frame
                        } else {
                            slot.transcoder
-                                .resample(&pcm_frame, MIX_RATE, target_rate)
+                                .resample_f32(&pcm_frame, MIX_RATE, target_rate)
                                .unwrap_or_default()
                        };

                        if let Ok(encoded) =
-                            slot.transcoder.encode_from_pcm(&resampled, slot.codec_pt)
+                            slot.transcoder.encode_from_f32(&resampled, slot.codec_pt)
                        {
                            if !encoded.is_empty() {
                                let header = build_rtp_header(
@@ -523,7 +550,7 @@ async fn mixer_loop(
                .filter(|(_, s)| matches!(s.role, LegRole::Participant))
                .map(|(lid, s)| ToolAudioSource {
                    leg_id: lid.clone(),
-                    pcm_16k: s.last_pcm_frame.clone(),
+                    pcm_48k: s.last_pcm_frame.clone(),
                })
                .collect();

@@ -533,7 +560,7 @@ async fn mixer_loop(
                        .iter()
                        .map(|s| ToolAudioSource {
                            leg_id: s.leg_id.clone(),
-                            pcm_16k: s.pcm_16k.clone(),
+                            pcm_48k: s.pcm_48k.clone(),
                        })
                        .collect(),
                };
--- a/rust/crates/proxy-engine/src/tool_leg.rs
+++ b/rust/crates/proxy-engine/src/tool_leg.rs
@@ -2,7 +2,7 @@
 //!
 //! Tool legs are observer legs that receive individual audio streams from each
 //! participant in a call. The mixer pipes `ToolAudioBatch` every 20ms containing
-//! each participant's decoded PCM@16kHz tagged with source leg ID.
+//! each participant's decoded PCM@48kHz f32 tagged with source leg ID.
 //!
 //! Consumers:
 //! - **Recording**: writes per-source WAV files for speaker-separated recording.
@@ -37,20 +37,25 @@ pub fn spawn_recording_tool(

        while let Some(batch) = rx.recv().await {
            for source in &batch.sources {
-                // Skip silence-only frames (all zeros = no audio activity).
-                let has_audio = source.pcm_16k.iter().any(|&s| s != 0);
+                // Skip silence-only frames (near-zero = no audio activity).
+                let has_audio = source.pcm_48k.iter().any(|&s| s.abs() > 1e-6);
                if !has_audio && !recorders.contains_key(&source.leg_id) {
                    continue; // Don't create a file for silence-only sources.
                }

                let recorder = recorders.entry(source.leg_id.clone()).or_insert_with(|| {
                    let path = format!("{}/{}-{}.wav", base_dir, call_id, source.leg_id);
-                    Recorder::new_pcm(&path, 16000, None).unwrap_or_else(|e| {
+                    Recorder::new_pcm(&path, 48000, None).unwrap_or_else(|e| {
                        panic!("failed to create recorder for {}: {e}", source.leg_id);
                    })
                });

-                if !recorder.write_pcm(&source.pcm_16k) {
+                // Convert f32 [-1.0, 1.0] to i16 for WAV writing.
+                let pcm_i16: Vec<i16> = source.pcm_48k
+                    .iter()
+                    .map(|&s| (s * 32767.0).round().clamp(-32768.0, 32767.0) as i16)
+                    .collect();
+                if !recorder.write_pcm(&pcm_i16) {
                    // Max duration reached — stop recording this source.
                    break;
                }
@@ -88,7 +93,7 @@ pub fn spawn_recording_tool(

 /// Spawn a transcription tool leg.
 ///
-/// The plumbing is fully real: it receives per-source unmerged PCM@16kHz from
+/// The plumbing is fully real: it receives per-source unmerged PCM@48kHz f32 from
 /// the mixer every 20ms. The consumer is a stub that accumulates audio and
 /// reports metadata on close. Future: will stream to a Whisper HTTP endpoint.
 pub fn spawn_transcription_tool(
@@ -105,7 +110,7 @@ pub fn spawn_transcription_tool(
        while let Some(batch) = rx.recv().await {
            for source in &batch.sources {
                *source_samples.entry(source.leg_id.clone()).or_insert(0) +=
-                    source.pcm_16k.len() as u64;
+                    source.pcm_48k.len() as u64;

                // TODO: Future — accumulate chunks and stream to Whisper endpoint.
                // For now, the audio is received and counted but not processed.
@@ -118,7 +123,7 @@ pub fn spawn_transcription_tool(
            .map(|(leg_id, samples)| {
                serde_json::json!({
                    "source_leg_id": leg_id,
-                    "duration_ms": (samples * 1000) / 16000,
+                    "duration_ms": (samples * 1000) / 48000,
                })
            })
            .collect();
--- a/rust/crates/proxy-engine/src/webrtc_engine.rs
+++ b/rust/crates/proxy-engine/src/webrtc_engine.rs
@@ -290,8 +290,9 @@ async fn browser_to_mixer_loop(
                        .send(RtpPacket {
                            payload: payload.to_vec(),
                            payload_type: PT_OPUS,
-                            marker: false,
-                            timestamp: 0,
+                            marker: rtp_packet.header.marker,
+                            seq: rtp_packet.header.sequence_number,
+                            timestamp: rtp_packet.header.timestamp,
                        })
                        .await;
                }
--- a/rust/crates/sip-proto/src/helpers.rs
+++ b/rust/crates/sip-proto/src/helpers.rs
@@ -197,10 +197,11 @@ pub fn compute_digest_auth(

 use crate::Endpoint;

-/// Parse the audio media port and connection address from an SDP body.
+/// Parse the audio media port, connection address, and preferred codec from an SDP body.
 pub fn parse_sdp_endpoint(sdp: &str) -> Option<Endpoint> {
    let mut addr: Option<&str> = None;
    let mut port: Option<u16> = None;
+    let mut codec_pt: Option<u8> = None;

    let normalized = sdp.replace("\r\n", "\n");
    for raw in normalized.split('\n') {
@@ -208,10 +209,16 @@ pub fn parse_sdp_endpoint(sdp: &str) -> Option<Endpoint> {
        if let Some(rest) = line.strip_prefix("c=IN IP4 ") {
            addr = Some(rest.trim());
        } else if let Some(rest) = line.strip_prefix("m=audio ") {
+            // m=audio <port> RTP/AVP <pt1> [<pt2> ...]
            let parts: Vec<&str> = rest.split_whitespace().collect();
            if !parts.is_empty() {
                port = parts[0].parse().ok();
            }
+            // parts[1] is "RTP/AVP" or similar, parts[2..] are payload types.
+            // The first PT is the preferred codec.
+            if parts.len() > 2 {
+                codec_pt = parts[2].parse::<u8>().ok();
+            }
        }
    }

@@ -219,6 +226,7 @@ pub fn parse_sdp_endpoint(sdp: &str) -> Option<Endpoint> {
        (Some(a), Some(p)) => Some(Endpoint {
            address: a.to_string(),
            port: p,
+            codec_pt,
        }),
        _ => None,
    }
--- a/rust/crates/sip-proto/src/lib.rs
+++ b/rust/crates/sip-proto/src/lib.rs
@@ -9,9 +9,11 @@ pub mod dialog;
 pub mod helpers;
 pub mod rewrite;

-/// Network endpoint (address + port).
+/// Network endpoint (address + port + optional negotiated codec).
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct Endpoint {
    pub address: String,
    pub port: u16,
+    /// First payload type from the SDP `m=audio` line (the preferred codec).
+    pub codec_pt: Option<u8>,
 }
--- a/rust/crates/sip-proto/src/rewrite.rs
+++ b/rust/crates/sip-proto/src/rewrite.rs
@@ -92,7 +92,7 @@ pub fn rewrite_sdp(body: &str, ip: &str, port: u16) -> (String, Option<Endpoint>
        .collect();

    let original = match (orig_addr, orig_port) {
-        (Some(a), Some(p)) => Some(Endpoint { address: a, port: p }),
+        (Some(a), Some(p)) => Some(Endpoint { address: a, port: p, codec_pt: None }),
        _ => None,
    };

--- a/ts/00_commitinfo_data.ts
+++ b/ts/00_commitinfo_data.ts
@@ -3,6 +3,6 @@
 */
 export const commitinfo = {
  name: 'siprouter',
-  version: '1.16.0',
+  version: '1.18.0',
  description: 'undefined'
 }
--- a/ts/sipproxy.ts
+++ b/ts/sipproxy.ts
@@ -425,9 +425,9 @@ async function startProxyEngine(): Promise<void> {
        id: data.leg_id,
        type: data.kind,
        state: data.state,
-        codec: null,
-        rtpPort: null,
-        remoteMedia: null,
+        codec: data.codec ?? null,
+        rtpPort: data.rtpPort ?? null,
+        remoteMedia: data.remoteMedia ?? null,
        metadata: data.metadata || {},
      });
    }
--- a/ts_web/00_commitinfo_data.ts
+++ b/ts_web/00_commitinfo_data.ts
@@ -3,6 +3,6 @@
 */
 export const commitinfo = {
  name: 'siprouter',
-  version: '1.16.0',
+  version: '1.18.0',
  description: 'undefined'
 }
Author	SHA1	Message	Date
Juergen Kunz	b6950e11d2	v1.18.0	2026-04-10 17:25:34 +00:00
Juergen Kunz	e4935fbf21	feat(readme): expand documentation for voicemail, IVR, audio engine, and API capabilities	2026-04-10 17:25:34 +00:00
Juergen Kunz	f543ff1568	v1.17.2	2026-04-10 17:14:14 +00:00
Juergen Kunz	c63a759689	fix(proxy-engine): use negotiated SDP payload types when wiring SIP legs and enable default nnnoiseless features for telephony denoising	2026-04-10 17:14:14 +00:00
Juergen Kunz	a02146633b	v1.17.1	2026-04-10 16:57:07 +00:00
Juergen Kunz	f78639dd19	fix(proxy-engine,codec-lib,sip-proto,ts): preserve negotiated media details and improve RTP audio handling across call legs	2026-04-10 16:57:07 +00:00
Juergen Kunz	2aca5f1510	v1.17.0	2026-04-10 15:58:41 +00:00
Juergen Kunz	73b28f5f57	feat(proxy-engine): upgrade the internal audio bus to 48kHz f32 with per-leg denoising and improve SIP leg routing	2026-04-10 15:58:41 +00:00