public-speech.json records four LibriSpeech dev-clean clips.
OpenSLR SLR12 licenses LibriSpeech under CC BY 4.0.
OpenSLR identifies the corpus as 16 kHz read English speech derived from LibriVox audiobooks.
Transcript rows were inspected through the Hugging Face datasets-server endpoint recorded per clip.
Four individual FLAC clips were downloaded into approved OS-temp storage, totaling 409,296 bytes.
Checksums identify those exact Hugging Face-served files, not an independently downloaded OpenSLR archive.
Durations, 16 kHz sample rate and mono channel count were verified from each FLAC STREAMINFO header.
No audio binary is committed; tests remain offline.
Metadata inspection is not listening, STT accuracy, native fixture or hardware evidence.
FFmpeg n9.0.2 decoded all four clips to finite mono 16 kHz Float32LE samples within [-1, 1].
Separate .fixture.f32le variants add only the zero-padding needed for 160-frame fixture quanta; source PCM remains unchanged.
fixture_pcm records their filenames, exact frame counts, padding and checksums; binary files remain in OS temp.
voice-journeys.json is logical fake UI data only.
Its book-derived captions are not evidence of speech recognition.
Its conversational utterances are separate synthetic acceptance text, not attributed recordings.
recorded-speech.json contains measured 100 ms RMS/16-bin peak envelopes derived from the first recording.
The fake replays only these small JSON measurements and supplied text at real time, without decoding PCM or running a speech model.
Preview word timing is uniform synthetic pacing, not measured alignment.
Synthesis/playback events remain logical fake evidence, not audible generated speech.
# Voice fixture provenance
`public-speech.json` records four LibriSpeech `dev-clean` clips.
OpenSLR SLR12 licenses LibriSpeech under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
OpenSLR identifies the corpus as 16 kHz read English speech derived from LibriVox audiobooks.
Transcript rows were inspected through the Hugging Face datasets-server endpoint recorded per clip.
Four individual FLAC clips were downloaded into approved OS-temp storage, totaling 409,296 bytes.
Checksums identify those exact Hugging Face-served files, not an independently downloaded OpenSLR archive.
Durations, 16 kHz sample rate and mono channel count were verified from each FLAC STREAMINFO header.
No audio binary is committed; tests remain offline.
Metadata inspection is not listening, STT accuracy, native fixture or hardware evidence.
FFmpeg n9.0.2 decoded all four clips to finite mono 16 kHz Float32LE samples within [-1, 1].
Separate `.fixture.f32le` variants add only the zero-padding needed for 160-frame fixture quanta; source PCM remains unchanged.
`fixture_pcm` records their filenames, exact frame counts, padding and checksums; binary files remain in OS temp.
`voice-journeys.json` is logical fake UI data only.
Its book-derived captions are not evidence of speech recognition.
Its conversational utterances are separate synthetic acceptance text, not attributed recordings.
`recorded-speech.json` contains measured 100 ms RMS/16-bin peak envelopes derived from the first recording.
The fake replays only these small JSON measurements and supplied text at real time, without decoding PCM or running a speech model.
Preview word timing is uniform synthetic pacing, not measured alignment.
Synthesis/playback events remain logical fake evidence, not audible generated speech.