karicotiza/acuity

Simple and reliable REST API service for audio recognition

Acuity

A simple and reliable REST API service for audio recognition

Features

🎙️ Recognizes English with wav2vec2-large-xlsr-53-english

🧩 Accepts files and base64

📄 Has support for Swagger and Redoc

💾 Logs hashes instead of sensitive information

🔥 Uses caching, queues and data validation

🔐 Uses JWT for authentication

🛡️ DOS protected

🏎️ GPU accelerated

Run

Steps:

1. Make sure you have docker installed
2. Download xlsr-53 and put in ./nn_model folder
3. Run docker compose up --build in the terminal

GPU acceleration

Steps:

1. Make sure that the nvidia graphics card is physically installed on your device
2. Run docker compose -f .\docker-compose-gpu.yml up --build

Endpoints

• api/token/ - JWT
• api/v1/schema/swagger/ or api/v1/schema/redoc/ - Documentation
• api/v1/ - DRF browsable API

Develop

Steps:

1. Make sure you have python 3.12 installed
2. Make sure you have docker installed
3. Download xlsr-53 and put in ./nn_model folder
4. Run python -m pip install -r .\src\requirements\dev.txt in the terminal to install dependencies
5. Write some new code
6. Run python manage.py migrate in the terminal from src folder to apply migrations
7. Run python manage.py createsuperuser in the terminal from src folder to create user
8. Run python manage.py runserver in the terminal from src folder to start django dev server
9. Run python manage.py celery dev in the terminal from src folder to start celery dev server
10. Run python -m pytest . in the terminal from . folder to run tests

Customize

Change the values in the ./prod.env file

PostgreSQL

• POSTGRES_HOST - Host
• POSTGRES_NAME- Table prefixes
• POSTGRES_PASSWORD - Password
• POSTGRES_USER - User
• POSTGRES_DB - Database name
• POSTGRES_PORT - Port

Redis

• REDIS_ADDRESS - Address
• REDIS_TIMEOUT - Cache lifetime in seconds

RabbitMQ

• RABBITMQ_ADDRESS - Address including vhost
• RABBITMQ_DEFAULT_USER - User
• RABBITMQ_DEFAULT_PASS - Password
• RABBITMQ_DEFAULT_VHOST - VHost

Neural Network Model

• NN_CONVERTER_FORMAT - Format to which the audio will be converted
• NN_CONVERTER_BITRATE - Bitrate of converted audio
• NN_CONVERTER_MONO - Convert audio to mono
• NN_MODEL_PATH - Path to the neural network model inside docker
• NN_MAX_LENGTH - Maximum length of audio to be processed
• NN_SAMPLE_RATE - Sample rate of audio coming into the neural network

Django

• DJANGO_SUPERUSER_USERNAME - Username
• DJANGO_SUPERUSER_PASSWORD - Password
• DJANGO_SUPERUSER_EMAIL - Email
• DJANGO_SECRET_KEY - Secret Key
• DJANGO_ALLOWED_HOSTS - Allowed hosts

Example

from pathlib import Path
from requests import get, post, Response
from time import sleep


def authenticate(url: str, login: str, password: str) -> dict[str, str]:
    response: Response = post(
        url, data={'username': login, 'password': password}
    )

    data: dict = response.json()
    jwt: str = data.get('access', '')
    headers: dict[str, str] = {'Authorization': f'Bearer {jwt}'}

    return headers


def recognize(
    url: str, file: Path, wait: float, headers: dict[str, str],
) -> str:
    response: Response = post(
        url, files={'file': open(file, 'rb')}, headers=headers
    )

    data: dict = response.json()
    link_to_recognized_text: str = data.get('link', '')

    while True:
        redirect: Response = get(link_to_recognized_text, headers=headers)
        redirect_data: dict = redirect.json()
        ready: bool = redirect_data.get('ready', False)

        if ready:
            text: str = redirect_data.get('text', '')
            return text

        else:
            sleep(wait)


auth_url: str = 'http://localhost:8000/api/token/'
recognition_url: str = 'http://localhost:8000/api/v1/file/'
login: str = 'admin'
password: str = 'admin'


headers = authenticate(auth_url, login, password)
file: Path = Path('tests', 'data', 'audio', 'audio.wav')
recognized_text: str = recognize(recognition_url, file, 0.2, headers)

print(recognized_text)