deep-speech

Dataset

Librispeech: 1000 hours of read English speech with sampling rate of 16 kHz.

train split: "train-clean-100" has 28539 examples (251 speakers)

- speaker_id
  - chapter id
    - speaker_id_chapter_id_000.flac
    - speaker_id_chapter_id_000.flac
    - ...
    - speaker_id_chapter_id.trans.txt
- speaker_id
...  

test split: "test-clean" has 2620 examples. (40 speakers)

Experiment

$ python3 experiment.py

Inference

$ python3 transcribe.py

Citation

• https://www.assemblyai.com/blog/end-to-end-speech-recognition-pytorch/
• CTC decoder gist: https://gist.github.com/awni/56369a90d03953e370f3964c826ed4b0