stt-listen 4.0.1b2

Transcribe long audio files with ASR or use the streaming interface

Listen: STT Services

This program is composed of two parts:

• A server aimed to be runned as a background service to serve ASR models within the bounds of a socket.
• A client to query the models to transcribe audio from files or directly from a live microphone stream.

The outputed wav file can be stored for later use.

You can then use the data.helper script to verify the transcription of every wav file and update the CSV training register before you start training a model.

Requirements

• python-pyaudio

Installation

Once you have a working pyaudio for your version of python, install listen.

pip install stt-listen
# Or from source
pip install git+https://gitlab.com/waser-technologies/technologies/listen.git

Usage

❯ listen --help
usage: listen [-h] [-f FILE] [--aggressive {0,1,2,3}] [-d MIC_DEVICE]
                   [-w SAVE_WAV]

Transcribe long audio files using webRTC VAD or use the streaming interface
from a microphone

options:
  -h, --help            show this help message and exit
  -f FILE, --file FILE  Path to the audio file to run (WAV format)
  --aggressive {0,1,2,3}
                        Determines how aggressive filtering out non-speech is.
                        (Integer between 0-3)
  -d MIC_DEVICE, --mic_device MIC_DEVICE
                        Device input index (Int) as listed by
                        pyaudio.PyAudio.get_device_info_by_index(). If not
                        provided, falls back to PyAudio.get_default_device().
  -w SAVE_WAV, --save_wav SAVE_WAV
                        Path to directory where to save recorded sentences
  --debug               Show debug info

Start the server

To use listen, you need a socket with STT models at the ready.

Example to enable as service.

cp ./listen.service.example /usr/lib/systemd/user/listen.service
systemctl --user enable --now listen.service

Models for faster-whisper will be downloaded the first time your run the server.

Or manually with uvicorn.

uvicorn listen.Whisper.as_service:app --reload --port 5063

Get authorization to listen

You need to authorize the system to listen first. Change the service configuration as follows.

# ~/.assistant/stt.toml
...
[stt]
is_allowed = true
...

Then start the server and use listen to start transcribing audio.

Use the client

Transcribe a file

You can quickly transcribe a wav file.

❯ listen -f audio.wav
Filename                       Duration(s)         
audio.wav                      3.580               

❯ cat audio.txt
───────┬───────────────────────────────────────────────────────────────────────────────────────────────────────────────
       │ File: audio.txt
───────┼───────────────────────────────────────────────────────────────────────────────────────────────────────────────
   1   │ Bonjour.
───────┴───────────────────────────────────────────────────────────────────────────────────────────────────────────────

Transcribe from a live microphone stream

You can also query the models in real time from a microphone.

❯ listen
You can speak now.
Bonjour.
^C
Stopped listening.

Supported languages

By default, the server uses the system's language according to the environment variable $LANG.

You can manually specify a supported language for the server to use.

LANG="en_US.UTF-8" uvicorn listen.Whisper.as_service:app --reload --port 5063

This will look for a good model for this language.

You can also directly specify a model to load.

ASR_MODEL_ID="bofenghuang/whisper-large-v3-french" uvicorn listen.Whisper.as_service:app --reload --port 5063