[wavekat-eason]

Follow-up to #55. The backend landed with plumbing tests — frame sizes, error types, the carry invariant, allocation counts — but nothing that measured whether it actually detects speech, and it was not registered with the accuracy or benchmark harnesses. This adds that, plus coverage for the parts of FrameAdapter the rework left untested.

Accuracy

tests/accuracy.rs now evaluates earshot on the same 30-file TEN-VAD testset as every other backend, and it has a baseline entry:

Backend	Precision	Recall	F1	Frame	Avg inference	RTF
Earshot	0.909	0.949	0.928	256 (16 ms)	~4 µs	0.0003

F1 matches TEN-VAD (0.928) at roughly an order of magnitude less time per frame, with no native code in the build.

earshot is added to the accuracy jobs in ci.yml and vad-accuracy.yml, and to make accuracy, make accuracy-update-baseline, and make bench — so the README benchmark table picks up an Earshot row the next time the VAD Accuracy workflow runs, and the baseline is checked on every push to main.

Backend tests (src/backends/earshot.rs)

Range and finiteness checks hold for any implementation that returns a number, so these pin down behaviour instead:

• a_signal_scores_clearly_above_digital_silence / silence_never_crosses_the_suggested_threshold — the model responds to its input and separates voiced content from silence at the suggested 0.5 threshold.
• rejected_frames_leave_the_stream_state_untouched — the documented "validated before any state is touched" property: interleaving wrong-rate and wrong-length calls between valid frames reproduces the clean score sequence exactly.
• context_carries_across_frame_boundaries — a continuous stream scores differently from the same frames scored in isolation, so the per-stream recurrent context is reaching the model and frame ordering matters.
• default_matches_new, two_fresh_detectors_score_identically — no hidden global or non-deterministic state.

Adapter tests (src/adapter.rs)

• reset_forwards_to_the_inner_detector — reset() clears the carry buffer and the wrapped detector. Only the buffer half was covered.
• timings_come_from_the_inner_detector, accessors_mirror_the_inner_detectors_capabilities — delegation.
• zero_frame_size_is_rejected_by_every_api — the frame_size == 0 branch, across all four entry points, with no input consumed.
• errors_still_report_the_frames_scored_before_the_failure — the documented on_score contract on the error path.
• randomised_chunk_boundaries_preserve_exact_framing — the existing exact-ordering assertions, driven by a seeded LCG across five frame sizes rather than a fixed list of chunk sizes.

End-to-end (tests/earshot_stream.rs)

Public API only, in the shape a transport uses it:

• transport_chunks_score_identically_to_aligned_frames and jittery_chunk_sizes_score_identically_to_aligned_frames — feeding audio as 320-sample packets, or as pseudo-random chunk sizes, produces the exact same score sequence as feeding aligned 256-sample frames directly. This is the property the adapter exists to provide, and it ties the framing fix in feat: add pure-Rust earshot VAD backend #55 to the backend.
• a_minute_of_streaming_stays_in_step — 3000 packets, asserting after each one that every sample fed in has either been scored or is still carried, and that the carry stays under a frame.
• a_voiced_region_scores_above_the_silence_around_it — silence → signal → silence through the adapter.
• reset_restarts_the_stream_through_the_adapter, a_wrong_rate_packet_does_not_disturb_the_stream.

Verification

cargo fmt --all --check, cargo clippy --workspace --all-targets -- -D warnings, and clippy with --no-default-features and --features earshot are clean. Tests pass with no features, with earshot, and on the default workspace set; the full feature combination compiles. The accuracy report was run locally against the downloaded testset to produce the baseline numbers above.

Each new test was checked to fail when the behaviour it describes is removed, so none of them pass by construction.

🤖 Generated with Claude Code

https://claude.ai/code/session_012PF9YKsA6GfchGTPDAWYzL