feat(proxy-engine): upgrade the internal audio bus to 48kHz f32 with per-leg denoising and improve SIP leg routing

This commit is contained in:
2026-04-10 15:58:41 +00:00
parent 10ad432a4c
commit 73b28f5f57
10 changed files with 194 additions and 65 deletions

View File

@@ -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 +296,86 @@ 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).
pub fn decode_to_f32(&mut self, data: &[u8], pt: u8) -> Result<(Vec<f32>, u32), String> {
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.
pub fn encode_from_f32(&mut self, pcm: &[f32], pt: u8) -> Result<Vec<u8>, String> {
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)]