Fabio
c36510c21e
Some checks failed
Linux Build / Linux cpu (push) Has been cancelled
Linux Build / Linux cpu-clang (push) Has been cancelled
Linux Build / Linux vulkan (push) Has been cancelled
mac-build / macOS CPU (push) Has been cancelled
Nix Build / Nix (cpu) (push) Has been cancelled
Nix Build / Nix (vulkan) (push) Has been cancelled
windows-build / Windows CPU (push) Has been cancelled
35 lines
1.2 KiB
Markdown
35 lines
1.2 KiB
Markdown
# ASR Validation Assets
|
|
|
|
Small, public ASR/STT fixtures for smoke validation. These are not training data
|
|
and are intentionally tiny.
|
|
|
|
## LibriSpeech
|
|
|
|
Path: `assets/asr_validation/librispeech/`
|
|
|
|
The LibriSpeech examples are taken from the `openslr/librispeech_asr` Hugging Face
|
|
dataset mirror, with upstream source OpenSLR SLR12. LibriSpeech is distributed
|
|
under CC BY 4.0.
|
|
|
|
Each example includes:
|
|
|
|
- a 16 kHz mono WAV file
|
|
- a `.txt` reference transcript
|
|
- one row in `manifest.jsonl`
|
|
|
|
The manifest rows include source metadata, split, speaker/chapter ids, duration,
|
|
audio path, transcript path, and the reference text.
|
|
|
|
## Common Voice (Japanese)
|
|
|
|
Path: `assets/asr_validation/common_voice_ja/`
|
|
|
|
The Common Voice example is `common_voice_ja_20461197` from the Common Voice 8.0
|
|
Japanese test split, taken from the `japanese-asr/ja_asr.common_voice_8_0`
|
|
Hugging Face dataset mirror (row 264), with upstream source Mozilla Common
|
|
Voice. Common Voice is distributed under CC0 1.0. The 48 kHz MP3 was decoded
|
|
and resampled to 16 kHz. The mirror ends each transcription with a `.`, which
|
|
is kept as it is.
|
|
|
|
Each example includes the same files and manifest fields as LibriSpeech, with
|
|
the mirror's row in place of the speaker and chapter ids.
|