This project wraps OpenAI's whisper models with a simple HTTP API. It is assumed that the service is used by exactly one consumer, so a pre-shared API key is used as authentication.

The API design takes inspiration from the rev.ai async speech-to-text API. Transcription jobs are submitted via a HTTP POST, returning an internal reference, which can later be used to retrieve the transcription results. Results are stored in an internal database until retrieved, and can optionally be deleted afterwards.

OpenAPI documentation for the service is available at <service_url>/docs.

Deploy

0. Choose model & instance size

Whisper provides several sizes of their model, where model size is a trade-off between model accuracy, resource usage and transcription speed. Smaller models are generally faster and lighter, but more inaccurate, especially for certain languages and translation tasks.

Whisper inference can be run on both CPU and GPU, and this project supports both via slightly altered docker compose configuration. (see GPU support section) CPU inference is a lot slower, but easier to host. CPU inference generally scales well with CPU speed.

Another consideration when choosing an instance is disk size. In order to transcribe audio, it needs to be downloaded to a temporary folder before processing, so the HDD needs to have enough free space to allow for that. For some hosting environments (e.g. Digital Ocean), it can make sense to mount an additional disk in the VM instead of choosing a larger instance.

As a baseline, the small model can run on a 4GB Digital Ocean droplet, achieving roughly a 1-2x speedup over original audio when transcribing.

1. Prepare host environment

This project is intended to be run via docker compose. To get started:

install docker engine.
Clone this repository to the machine.

2. Configure service

Create an .env file from .env.example and configure it:

API_SECRET: the API key used to authenticate against the API.
WHISPER_MODEL: the whisper model size you want to use.
TRAEFIK_DOMAIN: the domain you want to access the service from. Its A records need to point to the host IP.
TRAEFIK_SSLEMAIL: an email which is used to verify domain ownership before a TLS certificate is issued.

3. Run service

Run make run to start the server.4. To launch at system startup, wrap it in a systemd launch service.

Develop

docker compose is required for local development.

It is recommended to setup a virtual environment for python tooling. To install dependencies in your virtual env, run pip install -e .[tooling,web,worker].

Copy .env.test to .env to configure the service.

Start

make dev

Builds and starts the docker containers.

# Bindings
http://localhost:5555                        => Celery dashboard
http://whisperbox-transcribe.localhost       => API
http://whisperbox-transcribe.localhost/docs  => API docs
./whisperbox-transcribe.sqlite               => Database

Destroy

This removes all containers and attached volumes.

make clean

Test

make test

Lint

make lint

Format

make fmt