soebb/pydeepspeech

Easy setup for Mozillas Deepspeech transcriber

pydeepspeech

• The simpliest way to use AI to generate transcriptions from a wav file.
• This project uses the Mozilla DeepSpeech engine built from the included demo:
  • https://github.com/mozilla/DeepSpeech-examples/tree/r0.9/vad_transcriber

Why you need this

Mozilla's deep speech can't process long voice samples. pydeepspeech fixes this by "chunking" the input sound into seperate wav files that are then individualy processed. Wav files are cut along periods of detected silence, controlled by the aggressive parameter.

Besides this, pydeepspeech is probably better to use anyway because it's much simpler to install than Mozilla's Deepspeech because the required data models needed for pydeepspeech are automatically downloaded and installed on first use.

Quick start

Console api:

$ pip install pydeepspeech
$ pydeepspeech --wav_file <WAVE_FILE> --aggressive 1 --out_file <TEXT_FILE>

-or-

$ pip install pydeepspeech
$ pydeepspeech --wav_file <WAVE_FILE> --out_file <TEXT_FILE> --model_dir <MY_PBMM_AND_SCORER_FILES>

-or-

$ pip install pydeepspeech
$ pydeepspeech_installmodels --pbmm <PBMM_FILE_OR_URL> --scorer <SCORER_FILE_OR_URL>
$ pydeepspeech --wav_file <WAVE_FILE> --out_file <TEXT_FILE>

Or in python

from pydeepspeech.transcribe import transcribe
transcribe(...)

Optional: Create a virtual python package

Download and install virtual env:

# Download
curl -X GET https://raw.githubusercontent.com/zackees/make_venv/main/make_venv.py -o make_env.py
python make_env.py  # Make the environment
source activate.sh  # Enter environment
$ pip install pydeepspeech

To get back into the environment execute source activate.sh (if windows, you must be using git-bash)

Testing

Testing and linting is very simple. Just run tox (link).

$ pip install tox
$ tox