DeepSpeech Frontend

A flask app that transcribes files served to it via HTTP POST, and redirects the user to the text we were able to get from their audio.

Installation

Install the dependencies, for Debian that would look like this:

apt update && apt install python3-pip git ffmpeg python3-virtualenv wget
git clone https://git.callpipe.com/fusionpbx/deepspeech_frontend
cd deepspeech_frontend

Creating a virtual environment is recommended:

python3 /usr/lib/python3/dist-packages/virtualenv.py -p python3 env
. env/bin/activate

Followed by installing the python packages needed:

pip install -r requirements.txt

At this point, Mozilla's DeepSpeech needs a language model:

mkdir models && cd models && wget https://github.com/mozilla/DeepSpeech/releases/download/v0.7.3/deepspeech-0.7.3-models.pbmm && wget https://github.com/mozilla/DeepSpeech/releases/download/v0.7.3/deepspeech-0.7.3-models.scorer

Now, lets take this for a test spin!

python run.py

Navigate to http://localhost:5000/ in a web browser and upload your audio file. Once transcribed, you'll be redirected to the transcription.

Production Use

Next, you'll probably want to install something to serve it with, rather than the flask built in web server. gunicorn is what I use:

apt update && apt install gunicorn3 python3-pip git ffmpeg wget 
mkdir /var/lib/deepspeech/ && /var/lib/deepspeech/models && cd /var/lib/deepspeech/models
wget https://github.com/mozilla/DeepSpeech/releases/download/v0.7.3/deepspeech-0.7.3-models.pbmm && wget https://github.com/mozilla/DeepSpeech/releases/download/v0.7.3/deepspeech-0.7.3-models.scorer
pip3 install git+https://git.callpipe.com/fusionpbx/deepspeech_frontend.git
pip3 install -r requirements.txt

Finally, run gunicorn to start the server:

gunicorn -b 0.0.0.0 deepspeech_frontend:app

You may want this running as a daemon and probably managed by your init system, take a look at deepspeech_frontend.service in this directory for a sample systemd unit file, you will likely need to modify it to your liking.

Configuration

Configuration is done in the beginning of deepspeech_frontend/__init__.py. You can modify the directory where uploaded and not yet fully processed files are temporarily stored, change the language model and more.

Adding an api_keys.txt to the base directory will restrict the API to authenticated users with valid API keys. The file format is simple with <API_Key>, anything you want on each line. To use a key, the API expects to be accessed with something equivalent to curl -X POST <server>/api/v1/process -H 'Authorization: Bearer <api_key>' -F file=@<your_audio_file>.

References Used

Thanks to the following people and resources, this project exists:

thefinn93 - https://github.com/thefinn93/webhook-receiver
- Helped me with Python Q&A
- Let me use his README & systemd unit file as a template
Ashraff Hathibelagal - https://progur.com/2018/02/how-to-use-mozilla-deepspeech-tutorial.html
- Provided great examples on how to work with Mozilla's DeepSpeech
Mozilla's DeepSpeech - https://github.com/mozilla/DeepSpeech
Flask's Documentation Writers - http://flask.pocoo.org/docs/1.0/patterns/fileuploads/
FFmpeg-Python - https://github.com/kkroening/ffmpeg-python/

Things to improve

Queue for recordings to be processed
Add endpoints that act like standard proprietary HTTP voice endpoints (making this a drop in replacement)
Add GPU support