Splits mp3s into chunks, converts them to wavs in mono at 16Hz, then processes in parallel using coqui stt. ~2h of audio in ~15min. The more CPU's the faster it will go. set RAYON_NUM_THREADS for control
- Requires stt c libraries to be available in the path.
- Requires coqui stt models and scorers
- Variables in
src/constants.rs
build with cargo build --release
ffmpeg -i originals/sn-906.20s.mp3 -acodec pcm_s16le -ac 1 -ar 16000 -aframes 10000 test-audio.wav
stt --model models/coqui-model.tflite --scorer models/coqui-huge-vocabulary.scorer --audio test-audio.wav --json > out-stt.json
# Notes
python -m pip install coqui-stt-model-manager
python -m pip install stt
# FFMPEG
ffmpeg -i 800.mp3 -acodec pcm_s16le -ac 1 -ar 16000 -aframes 10000 800.wav
ffprobe -v error -show_entries format=duration -of default=noprint_wrappers=1:nokey=1
ffmpeg -ss 60 -i input-audio.aac -t 15 -c copy output.aac
ffmpeg -i files/episodes/sn0904.mp3 -acodec pcm_s16le -ac 1 -ar 16000 files/output/sn0904.wav
ffmpeg -i originals/sn-906.mp3 -ss 0 -t 20 -c copy originals/sn-906.20s.mp3
# DS/STT
pip3 install stt
pip3 install deepspeech
deepspeech --model models/deepspeech-0.9.3-models.pbmm --scorer models/deepspeech-0.9.3-models.scorer --audio sandbox/800.wav --json > out.json
stt --model models/coqui-model.tflite --scorer models/coqui-huge-vocabulary.scorer --audio originals/sn-906.20s.mp3 --json > out-stt.json
# Deepspeech
curl -LO https://github.com/mozilla/DeepSpeech/releases/download/v0.9.3/deepspeech-0.9.3-models.pbmm
curl -LO https://github.com/mozilla/DeepSpeech/releases/download/v0.9.3/deepspeech-0.9.3-models.scorer
# Coqui STT
https://github.com/coqui-ai/STT/releases/download/v1.4.0/libstt.tflite.Linux.zip
https://coqui.gateway.scarf.sh/english/coqui/v1.0.0-huge-vocab/model.tflite
https://coqui.gateway.scarf.sh/english/coqui/v1.0.0-huge-vocab/huge-vocabulary.scorer
# Assets
https://www.grc.com/sn/sn-904.txt
https://media.grc.com/sn/sn-904.mp3