searchi/tools/speech-vad/README.md

CPU speech-activity diagnostic

Research tooling, separate from searchi's normal GPU-capable runtime. No model download, transcription, source indexing or product default is changed. Use the frozen source and model hashes in the protocol.

cmake -S tools/speech-vad -B .searchi/speech/build-vad-diagnostic -DWHISPER_SOURCE=/absolute/path/to/pinned/whisper.cpp
cmake --build .searchi/speech/build-vad-diagnostic --config Release --parallel 1

Run searchi-vad-diagnostic MODEL PCM_F32LE, retaining stdout and stderr separately. Independently verify the model SHA-256 before running. Input must be verified mono 16-kHz little-endian float32 PCM, at most ten minutes, with finite samples. Decoder overshoot beyond unity is preserved without clipping. CPU and one inference thread are fixed. Native error logs invalidate detection even when upstream returns success. Probability count must match the pinned 512-sample inference framing.

Output JSON retains every probability and raw centisecond interval, source sample count and load/detection/segmentation milliseconds. Interval padding is already applied. Preserve this raw output; a subsequent harness must explicitly intersect intervals with actual decoded source samples and add the verified source origin exactly once. Do not concatenate disjoint intervals without a piecewise source map. Zero intervals are a valid no-speech decision, distinct from nonzero exit status or missing JSON.

The wrapper fixes the preregistered segmentation parameters: threshold 0.5, minimum speech 250 ms, silence 100 ms, maximum speech FLT_MAX, padding 30 ms, overlap 0.1 seconds. It does not establish model accuracy or a production speech gate. Compare the complete frozen speech/music/sound pool before judging quality; preserve original Google audio retrieval regardless of transcription.

gated_asr.py runs the frozen base transcription experiment from the verified VAD report. It preserves all five competitors, no-speech status, original source intervals and raw ASR logs. compare.py BASELINE GATED reports component timing and segment counts only; it rejects incomplete or invalid pools and does not score retrieval quality. See the protocol document for actual evidence.

Separate frozen VAD + DTW comparison

gated-dtw-policy.json combines the already frozen Silero source-intersection decisions with the separately declared base DTW sentence-point contract. Add --dtw-sentences to gated_asr.py and select a new output, for example .searchi/speech/gated-base-dtw-results-v1.json. All five source rows remain mandatory; explicit no-speech rows use unchanged actual frozen VAD evidence and launch no ASR. Neither relevance labels nor the observed malformed piano output selects a gate or a window. Original raw Google audio retrieval stays unchanged. Original raw recognized segments and ordinary token errors remain recorded. Valid DTW points alone determine the separate aligned candidate times; missing or invalid points make accepted comparison scoring ineligible.

The combined manifest pins unchanged VAD report/worker, base worker, DTW policy, DTW helper and ordinary helper hashes. It preserves the old gated policy in policy, with separate alignment_policy and combination_policy provenance. Rankers must independently validate both policies and reconstruct every frozen VAD window and DTW point before inference/scoring. Existing gated base output is not overwritten. A combination plan or structural tests establish no new speech recognition or timestamp quality.

Run only after the coordinator freezes the plan and grants the exclusive CPU slot. Use --no-gpu --threads 1 --processors 1 --dtw base --no-flash-attn; the harness adds these explicitly. The twelve existing speech windows remain bounded to 30 seconds, 4 MiB raw output and 600 seconds each. Observed gated base processing previously took about 65 seconds including VAD cost; DTW cost and memory have not been measured. Native DTW workspace defaults to an additional 128 MiB, which is not a total memory ceiling. Strict parse failures retain bounded raw bytes and atomically persist incomplete status, then stop instead of skipping a competitor. No full rerun has established this combined policy. DTW native logs are retained byte-exact as bounded base64 plus SHA so malformed recognized bytes in diagnostic stderr cannot cause lossy decoding. Recognized JSON remains strict UTF-8 and cannot be repaired by this log representation.