6.3 KiB
Independent English/German speech preparation v1
No ASR, VAD or Google inference has run on this fixture. The ten topic/absence queries and two source recordings were frozen before preparation. The sources are human LibriVox readings of Article 1 of the Universal Declaration of Human Rights, unrelated to the earlier NASA topics. They share translated content and are correlated; this is not a representative multilingual speech benchmark.
| Source | Human reader | Recording permission | Original bytes |
|---|---|---|---|
| English Article 1 | David Barnes | LibriVox public-domain dedication / copyrighted free use | 168186 |
| German Article 1 | Roman Buettner | LibriVox public-domain dedication / copyrighted free use | 151136 |
The publisher pages explicitly identify the human readers, document and language. The primary reference text is UN Article 1 in English and UNRIC Article 1 in German. The manifest records reference keywords and topic paraphrases, never model captions or recognition-derived ground truth. Full primary text is externally linked; it must not be supplied as an ASR prompt or indexed as if recognized.
Only 319322 new bytes were downloaded. Sources are checked against official Commons API size/SHA-1 and acquired SHA-256. The existing independently licensed US Navy drum cadence supplies the instrumental control, with its previously locked source bytes. Including that reused source, original inputs total 1893869 bytes. Generated silence is exact zero PCM, released CC0.
Four-clip pool and honest timestamp scope
Prepared audio is mono 16-kHz s16 WAV. Each human recording remains complete, with five seconds of exact leading silence; no reference words are cropped. The resulting English/German durations are 22.512 and 16.219 seconds. Music and silence controls each cover exactly thirty seconds. Core WAVs total 3159568 bytes. The music waveform is the first thirty seconds of the already locked drum source; no descriptive tags are used to gate recognition.
Actual source audio stream starts are verified as zero, and the five-second padding/PCM source mapping is exact. The publishers do not provide phrase-level timed captions. Neither the padding boundary nor the whole article bounds prove the first phoneme or each sentence's start. Strict utterance-point scoring is explicitly unavailable; do not count a start anywhere in the article as precise localization success. Independent listening/timed annotations remain required.
Six positive EN/DE queries describe equal dignity/rights, reason/conscience and brotherly treatment. Either language's recording is relevant: translated topic equivalence is intentional. Four absent queries ask about Mars funding and salary/ paid leave. Silence and instrumental music remain in every candidate pool, and any recognized hallucination remains an explicit measured failure/competitor. Per-language recognized concept coverage is separate from semantic file recall.
Separate seam primitives
seam-plan.json freezes two artificial repeat controls, outside the four-clip
retrieval pool. English source samples repeat twice and German samples three
times after the same known padding, producing 40.024 and 38.657 seconds. Their
2517880 additional WAV bytes and exact piecewise source maps stress a 30-second
window with two-second overlap. These joins are artificial repetitions, not
natural uninterrupted speech. Do not extrapolate their quality to long talks.
Independent byte checks verified each derived piece equals the original decoded source PCM and the prefix is all zeros. Compare full-source VAD on these short controls with bounded independent windows, recording interval disagreement, source coverage, seam omissions/duplicates and processing cost. Restarted VAD need not reproduce a continuously initialized VAD's probabilities. Preserve differences rather than changing thresholds or treating an unverified reference as infallible. The same labels must not choose gate intervals or discard a seam.
Preparation and acceptance
python -m unittest discover -s tools/speech-eval -p test_prepare_validation.py
python tools/speech-eval/prepare_validation.py fixtures/speech-validation-v1/fixture.json .searchi/speech-validation-v1-seams .searchi/media-runtime/bin/ffmpeg.exe .searchi/media-runtime/bin/ffprobe.exe
Preparation already completed; existing files are preserved, so use a fresh
owned directory for another preparation. Five structural tests passed. Exact
padding and every seam's source pieces were checked independently without a
model. No recognition/quality score follows. Hashes are frozen in inputs.lock.json:
- Fixture:
e3a850d7a2385e250408f57fe51f4620ce80a20f8625ced42205f02cc46d517d. - Seam plan:
5da4400814c4965fc12a87bb00c36665300730a702a69acd65a72b9f5201ac9e. - Prepared
.searchi/speech-validation-v1-seams/prepared.json:085f34f0af863656a28e7a905746973a54d1123bb07053541e09c8af0a58bb3f.
The earlier four-clip-only plan
4635e772834b0eed847459f5f4d47d7c99a591f0adba2c60ce75712557092279
is preserved under its original ignored directory and was never inferred.
Before accepted scoring, require every source/window and all ten queries, strict UTF-8 and bounded times, zero excluded failed competitors, unchanged worker/model/ decoder hashes and exact source-byte provenance. Report genuine German recognition and all preregistered reference concepts, source-level recall/MRR, absent-return rate and gate false positives separately. Both negative controls must retain zero accepted recognized speech segments; raw Google audio search stays available. Measure decode/VAD/ASR/alignment/Google/storage, cold model load, foreground impact and warm reuse separately. ASR remains bounded to thirty seconds even for seam controls. Pause between work units for foreground priority; do not claim native preemption during one inference call.
This plan does not adopt a provider, change production search or close fine timestamp, long-source, accent, representative-noise or human-listening gates.