voice-transcript app & microservice

Miniapp and corresponding microservice implementation performing audio/voice transcription (speech-to-text) based on on speechbrain/hugging face deep model (wav2vec2).

Content

Input:

• b64 encoded audio (buffer input) or a file path (file input).

Output:

• text transcription of input audio (french);
• Speech rate in word per minute;
• Emotion related feature computing;

Project structure

• app/ contains application source code.
• eval/ gathers some evaluation/exploratory scripts and jupyter notebooks.
• tests/ is the testing files folder.
• samples/ contains some audio files for testing.
• libs/ contains mainly some audio processing (feature extractions) helpers
• demo/ contains a streamlit interface that runs the service internaly and shows result on a webpage.

Running the demo

• Make sure all dependencies are installed (see Dependencies section),
• Run streamlit demo/interactive_app.py,
• Open in your favourite browser (tested on firefox and Chromium) the generated (localhost) link if not opened automatically.

Runnig the service

Manual installation of the pretrained models

The pretrained models are not in the github. Currently they should be downloaded and put in the folder app/pretrained_models. Two models are needed:

• Speechbrain's wav2vec2-FR-7K-large/: download using

  git lfs clone https://huggingface.co/LeBenchmark/wav2vec2-FR-7K-large
  

• Lebenchmark's asr-wav2vec2-commonvoice-fr/: download using

  git lfs clone https://huggingface.co/speechbrain/asr-wav2vec2-commonvoice-fr
  

1. Locally & Manually

First, create a python virtual environment and check if all dependencies are properly installed.

python -m venv ./.venv && source ./.venv/bin/activate
pip install -r requirements.txt

To run the service, just run the file run.py with python:

python run.py

In this file, the only thing which is done is the uvicorn command to run the service. The api communicate with the following ip adress 127.0.0.1 using port 8000.

2. Using Docker

Be sure to get all the pretrained models installed in their corresponding folders. This should be done manually or added to any automatic deployement of the micro service.

Building the image:

docker build -t voice-transcript .

If no Docker network is established, create network to assign static IP address to the service:

docker network create --subnet=172.20.0.0/16 demo-network

Running the image:

docker run -d --net demo-network --ip 172.20.0.11 -p 8001:8001  -v path/to/local/pretrained_models/folder:/path/to/container/pretrained_models/folder voice-transcript

NOTE: Currently the command is docker run -d --net demo-network --ip 172.20.0.11 -p 8001:8001 -v /root/voice-transcript/pretrained_models:/code/pretrained_models voice-transcript

In this setting, the api communicate with the following ip adress 172.20.0.11 using port 8001.

Communicate with the API using CURL

• Test GET request

curl --request GET \
  --url http://172.20.0.11:8001/ \
  --header 'Content-Type: application/json'

• Transcribe bonjour.wav (POST request)

curl -X 'POST' \
  'http://172.20.0.11:8001/transcribe_file' \
  -H 'accept: application/json' \
  -H 'Content-Type: application/json' \
  -d '{
  "content_format": "wav",
  "name": "bonjour"
}'

Dependencies

to run the API and the streamlit demo.

• see the full dependencies available in requirements.txt

Credits

Author: Ali Saghiran