Canary ASR API

A FastAPI-based REST API for NVIDIA's family of Canary-1B ASR (Automatic Speech Recognition) models with support for translation.

Features

• Fast and accurate speech-to-text transcription
• Support for both local audio files and URLs
• Optional translation to other languages
• Real-time factor (RTF) calculation
• Processing time metrics
• Simple REST API interface

Installation Options

Option 1: Quick Start with Docker

1. Clone the repository:

git clone https://github.com/zetaphor/nemo-canary-fastapi.git
cd nemo-canary-fastapi

2. Start the service:

docker-compose up --build

The API will be available at http://localhost:8000

Option 2: Local Installation

Prerequisites

• Python 3.10
• NVIDIA GPU with CUDA support
• UV package manager

1. Clone the repository:

git clone https://github.com/zetaphor/nemo-canary-fastapi.git
cd nemo-canary-fastapi

2. Install dependencies:

uv sync
source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Change the model in api.py if desired:

Available models:

• nvidia/canary-1b
• nvidia/canary-1b-flash
• nvidia/canary-180m-flash

canary_model = EncDecMultiTaskModel.from_pretrained('nvidia/canary-1b')

4. Run the API:

uvicorn api:app --host 0.0.0.0 --port 8000

The API will be available at http://localhost:8000

API Usage

Transcribe Endpoint

POST /transcribe

The endpoint accepts JSON with the following parameters:

• audio_source: Path to local audio file or URL (required)
• target_lang: Target language for translation (optional)

Examples

1. Transcribe from local file:

curl -X POST "http://localhost:8000/transcribe" \
    -H "Content-Type: application/json" \
    -d '{"audio_source": "/path/to/audio.wav"}'

2. Transcribe from URL:

curl -X POST "http://localhost:8000/transcribe" \
    -H "Content-Type: application/json" \
    -d '{"audio_source": "https://example.com/audio.wav"}'

3. Translate from local file:

curl -X POST "http://localhost:8000/transcribe" \
    -H "Content-Type: application/json" \
    -d '{"audio_source": "/path/to/audio.wav", "target_lang": "de"}'

Response Format

For ASR:

{
    "text": "Hello, world!",
    "processing_time_seconds": 0.12,
    "audio_duration_seconds": 1.0,
    "rtf": 0.12
}

For translation:

{
    "text": "Hallo, Welt!",
    "processing_time_seconds": 0.15,
    "audio_duration_seconds": 1.0,
    "rtf": 0.15,
    "source_lang": "en",
    "target_lang": "de"
}

Development

To access the interactive API documentation, visit:

• Swagger UI: http://localhost:8000/docs
• ReDoc: http://localhost:8000/redoc