test(parity): rust runner for silero VAD parity harness

[al8n]

Adds tests/parity/Cargo.toml and src/main.rs for the silero-parity-runner binary that loads a 16 kHz mono WAV via ffmpeg-next, runs silero::detect_speech with the bundled ONNX model, and emits a JSON segment list. Pairs with the Python runner (next commit) for side-by-side comparison against upstream silero-vad.

The runner uses ffmpeg-next (not hound) for audio loading so the f32 buffer the model sees is byte-identical to what upstream Python silero-vad consumes via torchaudio/ffmpeg, letting the parity score verify both runners decoded the audio the same way before flagging any output divergence as a model issue. Same pattern whispery's parity harness uses.

Cargo.lock for the parity binary lives at tests/parity/Cargo.lock and is gitignored via the existing Cargo.lock rule. The parity crate is excluded from cargo package because nested Cargo.toml workspaces are skipped automatically.

test(parity): python runner + scorer for upstream silero-vad

Adds the upstream Python silero-vad reference side of the parity harness:

• python/pyproject.toml: pins silero-vad >= 5.1 and `onnxruntime

  = 1.18(theload_silero_vad(onnx=True)` path needs onnxruntime).

• python/silero_vad_runner.py: same CLI / same JSON schema as the Rust runner. Defaults to --backend onnx so both runners feed byte-identical bytes to ORT — same model file, same backend — and any IoU disagreement is segmenter logic, not the inference runtime. --backend jit available for measuring runtime drift separately. Audio loading is an ffmpeg shell-out matching WhisperX's load_audio (pcm_s16le -ac 1 -ar 16000 → np.float32 / 32768.0), same byte path as the Rust loader.
• python/score.py: sequence-position pairing, per-segment IoU, median + p10/p90 + worst-N report, JSON summary on stdout / --out. Pass/fail: median IoU >= 0.95 AND segment counts match.

test(parity): driver script and README

run.sh brings up the uv venv, runs both runners, and pipes the JSONs through score.py. Accepts a fixture directory (uses clip_16k.wav inside) or a direct WAV path.

run.sh passes --min-silence-ms 132 to the Rust side as a parity override (NOT a crate-default change). The silero crate's SpeechSegmenter::push_probability computes
silence_samples = current_sample - silence_start AFTER the current frame's increment, while upstream Python silero-vad computes it BEFORE — a one-frame (32 ms at 16 kHz / 512-sample windows) off-by-one that causes the crate to close segments after 4 low-prob frames vs Python's 5. Bumping the override by exactly one frame on the crate side restores byte-identical segment counts (verified on all 5 dia parity fixtures: median IoU 1.0000 across the board).

README documents the layout, prerequisites (cargo + uv + ffmpeg, no ORT_DYLIB_PATH needed because ort 2.0.0-rc.12's default features include download-binaries + copy-dylibs), the canonical fixture set (dia's parity fixtures, intentionally not copied), the parameter alignment table (named defaults match upstream silero-vad 6.2.1 exactly), and the off-by-one silence threshold finding in detail.

fix(detector): silence threshold counter matches upstream silero-vad

The silence-counter in SpeechSegmenter::push_probability evaluated silence_samples = current_sample - silence_start AFTER the current frame's contribution had been added to current_sample, while upstream Python silero-vad evaluates the equivalent sil_dur_now = cur_sample - temp_end BEFORE the current frame is consumed. The crate's counter therefore fired one model frame (32 ms at 16 kHz / 512-sample windows) early — at the default
min_silence_duration_ms = 100, the crate closed a segment after 4 consecutive low-probability frames where Python tolerates the dip and closes after 5.

Switch the comparator to frame_start.saturating_sub(silence_start), mirroring Python's cur_sample - temp_end evaluated before the frame is consumed. Same correction applies to the
min_silence_at_max_speech_samples comparator on the same line block, which used the same off-by-one counter.

Audit existing tests:

• middle_band_frames_do_not_reset_tentative_end and min_speech_duration_is_checked_before_padding extended their trailing silence runs by one frame so the close still fires via push_probability under the corrected counter.
• force_split_during_silence_closes_without_restarting raised its max_speech_duration ceiling by one frame so the max-speech split still fires after max_split_end has been recorded by the silence-counter logic.
• All updated tests document the change in their docstrings, citing the parity harness in tests/parity/ as motivation.

Add two new regression tests:

• four_frame_silence_dip_does_not_close_segment_at_default_min_silence pins that a 4-frame (128 ms) silence dip is now tolerated.
• five_frame_silence_dip_closes_segment_at_default_min_silence pins that 5 consecutive low-prob frames still close, matching upstream.

Discovered by the parity harness in commits dd64c35 / 003e8b6 / da8c0de.

test(parity): drop now-unneeded min-silence override

The --min-silence-ms 132 workaround in run.sh was compensating for an off-by-one in SpeechSegmenter::push_probability's silence counter (silero v0.2.x evaluated silence_samples = current_sample - silence_start after the current frame's increment, while Python evaluates the equivalent cur_sample - temp_end before the current frame is consumed). The crate fix in the previous commit aligns the two semantics, so the runner now uses upstream silero-vad defaults verbatim and parity numbers are unchanged.

Verified on all 5 short fixtures (01_dialogue, 02_pyannote_sample, 03_dual_speaker, 04_three_speaker, 05_four_speaker): median IoU 1.0000 and segment counts match exactly (51/51, 4/4, 14/14, 6/6, 14/14) without the override — same numbers the override produced pre-fix.

The --min-silence-ms flag remains on the runner CLI for advanced users who want to override per-run; only run.sh no longer applies it. README updated to mark the off-by-one finding as fixed in v0.3.0 and preserve the previous analysis as historical context.

chore: bump to 0.3.0 with CHANGELOG entry for the silence threshold fix

The silence-counter fix in SpeechSegmenter::push_probability is a behaviour change for any caller that hand-tuned
min_silence_duration_ms against v0.2.x's response curve. Bumping the minor version (0.2.x → 0.3.0) signals that even though it's strictly a bug fix, the new response curve may require re-tuning at the call site. Default callers do not need to change anything.

CHANGELOG entry covers what changed (silence-counter semantics now match upstream Python silero-vad), why (parity harness uncovered an off-by-one), and the migration note (subtract ~32 ms from hand-tuned min_silence_duration_ms overrides if you want to keep the v0.2.x effective behaviour).

[Copilot]

Pull request overview

Adds a manual parity harness (Rust + Python) to compare silero’s VAD segmentation against upstream silero-vad, and aligns the Rust segmenter’s silence counter semantics with upstream Python (shipping as v0.3.0).

Changes:

• Introduce tests/parity/ runner tooling: Rust silero-parity-runner, Python reference runner + IoU scorer, plus a driver script and README.
• Fix an off-by-one in SpeechSegmenter::push_probability silence accounting and update/add regression tests.
• Bump crate version to 0.3.0 and document the behavior change in CHANGELOG.md.

Reviewed changes

Copilot reviewed 10 out of 11 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
tests/parity/src/main.rs	New Rust parity runner: ffmpeg decode → detect_speech → JSON output
tests/parity/Cargo.toml	New standalone Cargo package for the parity runner
tests/parity/run.sh	Driver script to run Rust + Python runners and score results
tests/parity/python/pyproject.toml	Python env definition for upstream silero-vad reference runner
tests/parity/python/silero_vad_runner.py	Python reference runner emitting the same JSON schema as Rust
tests/parity/python/score.py	IoU scoring and pass/fail logic for runner outputs
tests/parity/README.md	Harness documentation, parameter alignment, and historical off-by-one notes
src/detector.rs	Silence-counter semantic fix + updated and new regression tests
Cargo.toml	Version bump to 0.3.0
CHANGELOG.md	Release notes for the silence-counter behavior change
.gitignore	Ignore parity harness outputs and Python venv artifacts

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.