Whisper is a general-purpose speech recognition model. It is trained on a large dataset of diverse audio and is also a multi-task model that can perform multilingual speech recognition as well as speech translation and language identification. For more details: github.com/openai/whisper

Run (Development Environment)

Enable venv:

python3.9 -m venv venv
source venv/bin/activate

Install poetry with following command:

pip3 install poetry==1.2.0

Install packages:

poetry install

Starting the Webservice:

poetry run whisper_asr

Quick start

After running the docker image or poetry run whisper_asr interactive Swagger API documentation is available at localhost:9000/docs

There are two endpoints available:

/asr
/detect-language

Automatic Speech recognition service /asr

If you choose the transcribe task, transcribes the uploaded sound file. You can provide the language or it will be automatically recognized. If you choose the translate task it will provide an English transcript no matter which language was spoken.

Returns a json with following fields:

text : Contains the full transcript
segments : Contains an entry per segment. Each entry provides time stamps, transcript, token ids and other metadata
language: detected or provided language (as a language code)

Language detection service /detect-language

Detects the language spoken in the uploaded sound file. For longer files it only processes first 30 seconds.

Returns a json with following fields:

detected_language
langauge_code

Build

Run

poetry build

Configuring the Model

export ASR_MODEL=base

Docker

Build Image

docker build -t whisper-asr-webservice .

Run Container

docker run -d -p 9000:9000 whisper-asr-webservice
# or
docker run -d -p 9000:9000 -e ASR_MODEL=base whisper-asr-webservice

TODO

Detailed README file
Github pipeline
Unit tests
CUDA version of Docker image
Hosted Swagger documentation with descriptions