API Processor

Русский · English

A REST API that turns audio and documents into text:

speech is recognized offline with a Vosk model;
text, tables and images are extracted from PDF and DOCX.

The result can be forwarded over gRPC to a text processing service. Speech recognition targets Russian.

Status: personal project, completed

Stack: Python 3.12 · Django 5.2 · Django REST Framework · Vosk · FFmpeg · pdfplumber · python-docx · gRPC · Docker

Features

Audio → text. Any format FFmpeg reads (OGG, MP3, M4A, WAV…). Audio is converted to 16 kHz mono with noise reduction and loudness normalization first. Recognition is offline, no external services.
Document → HTML. PDF and DOCX: paragraphs in <p>, tables as CSV in <pre>, images as base64. The document text is escaped, so the result is safe to render in a browser.
gRPC. The text is sent to the TextProcessor service (proto/text_service.proto), and its reply is returned in grpc_response. If no service address is set, the step is skipped.
Optional access token and a file size limit (50 MB by default).

API

curl -F audio=@examples/sample.ogg http://localhost:8000/api/audio-to-text/
curl -F document=@report.pdf http://localhost:8000/api/document-to-text/

{
  "text": "раз два три проверка перевода голоса текст насколько качественно она работает один два три четыре пять шесть семь восемь девять десять",
  "grpc_response": null
}

With API_TOKEN set, add Authorization: Bearer <token>. Errors come as {"error": "..."}:

Code	When
400	no file or unsupported document format
403	wrong token
413	file over the limit
422	audio could not be read

Running

git clone https://github.com/EDeev/api_processor.git && cd api_processor
cp .env.example .env      # set DJANGO_SECRET_KEY
docker compose up -d      # API at http://localhost:8000

The Vosk model and FFmpeg are inside the image. Prebuilt image: docker pull ghcr.io/edeev/api_processor or docker pull git.deev.su/edeev/api_processor.

Without Docker you need:

FFmpeg in PATH;
the vosk-model-small-ru-0.22 model unpacked into models/.

Then:

pip install -r requirements.txt
DJANGO_DEBUG=True python manage.py migrate
DJANGO_DEBUG=True python manage.py runserver

Variable	Purpose
`DJANGO_SECRET_KEY`	secret key, required unless `DJANGO_DEBUG=True`
`DJANGO_ALLOWED_HOSTS`	comma-separated domains
`API_TOKEN`	if set, access requires the token
`GRPC_SERVER`, `GRPC_TIMEOUT`	TextProcessor address and timeout, s
`MAX_UPLOAD_SIZE_MB`	file size limit
`VOSK_MODEL_PATH`, `FFMPEG_BINARY`	paths to the model and ffmpeg

How it works

flowchart LR
    C[Client] -->|multipart| V[DRF APIView]
    V -->|audio| F[FFmpeg: 16 kHz mono, denoise] --> K[Vosk]
    V -->|PDF / DOCX| S[pdfplumber / python-docx]
    K --> T[text]
    S --> T
    T -->|gRPC ProcessText| G[TextProcessor]
    T --> R[JSON response]

The Vosk model is loaded once per process. Each request gets its own temporary file. Uploaded files and results are stored in SQLite and media/.

Development

pip install -r requirements-dev.txt
ruff check --select E9,F,B --exclude proto . && pytest

What the tests cover:

recognition of the whole sample, with every phrase;
OGG conversion;
PDF and DOCX extraction with escaping;
limits and the token;
a real gRPC server and an unavailable one.

The Docker image is built on v* tags and published to GitHub Packages and git.deev.su.

License

MIT — see LICENSE. The Vosk model is licensed under Apache 2.0.

Author

Egor Deev — GitHub · Telegram · egor@deev.space

_{⭐ If you find this project useful, give it a star on GitHub!}

_{Made with ❤️ — deev.space}

4.5 KiB Raw Permalink Blame History