feat: add StreamingResampler

[wavekat-eason]

Summary

Adds StreamingResampler — a stateful resampler for real-time pipelines that feed short frames in sequence.

AudioFrame::resample is fine for one-shot conversions, but it constructs a fresh rubato per call. A consumer that calls it on every 20 ms G.711 packet pays the sinc reconstruction's edge cost on every chunk — the resampler sees an abrupt step from zero to the chunk's first sample and produces audible boundary artifacts at the frame rate (50 Hz for 20 ms packets). On a real playback path the user hears continuous noise/buzz over the voice. This bit wavekat-voice's M1 media work after the short-input fix from #23 made resample itself succeed.

StreamingResampler::new(source_rate, target_rate, chunk_size) builds rubato once. process accumulates input until it has chunk_size, runs one rubato step, and emits output. Internal filter state is preserved across calls.

Also factors the sinc parameters + builder out into a private build_sinc_resampler helper shared with AudioFrame::resample, so the two paths can't drift on quality settings.

API

use wavekat_core::StreamingResampler;

let mut r = StreamingResampler::new(8000, 44100, 160)?; // chunk = 1 G.711 frame
let mut out = Vec::new();
for packet in incoming_rtp_packets {
    let samples: Vec<f32> = /* G.711 decode */;
    r.process(&samples, &mut out)?;
    // `out` now has device-rate samples ready for the audio sink
}

Pass-through fast path when source_rate == target_rate: no resampler is built, process is a copy.

Test plan

• cargo test --features resample — 53 unit tests + 5 doctests pass.
• 6 new StreamingResampler tests:
  • same_rate_is_passthrough — proves the no-op fast path doesn't go through rubato.
  • short_input_chunked_calls — 10× 160-sample inputs at 8 kHz → 44.1 kHz produce roughly 10 × 882 output samples (the exact RTP receive shape).
  • buffers_across_partial_calls — splitting an input across two process calls produces sample-for-sample-identical output to one big call.
  • avoids_per_frame_edge_artifacts — the regression guard: a smooth 600 Hz sine, resampled chunk-by-chunk with StreamingResampler vs AudioFrame::resample. The streaming output's max consecutive-sample delta stays inside the expected smooth bound (under 10 spikes in steady state); the stateless per-chunk output has 5×+ more spikes because of the chunk-boundary transients.
  • rejects_zero_rate / rejects_zero_chunk_size — input validation.
• cargo clippy --workspace --all-features -- -D warnings clean.
• cargo fmt --all --check clean.

Downstream

wavekat-voice's LocalDeviceSink will switch from per-frame AudioFrame::resample to StreamingResampler once this lands and bumps to 0.0.11.

🤖 Generated with Claude Code

[sentry]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!