ASR Fusion

ASR Fusion is a Python package for audio transcription using multiple ASR (Automatic Speech Recognition) models. It provides a unified API interface that supports various ASR engines including Faster Whisper, FunASR, and SenseVoice.

Features

• Support for multiple ASR engines (Faster Whisper, FunASR, SenseVoice)
• RESTful API server with OpenAI-compatible endpoints
• File and streaming transcription support
• Multiple output formats (JSON, text, SRT, VTT)
• Easy model switching via model identifiers
• Configuration-based model management

Installation

# Clone the repository
git clone <repository-url>
cd asr-fusion

# Install dependencies
pip install -e .

Quick Start

1. Start the API Server

python main.py --host localhost --port 8603

Or using uvicorn directly:

uvicorn asr_fusion.api.server:app --host localhost --port 8603 --reload

2. Test with Client

# Test with the provided client
python client/faster_whisper_test.py

Or use the general test client:

python client/test_client.py --file path/to/audio.wav --model faster-whisper/large-v3

API Usage

The API follows the OpenAI speech recognition API format:

File Transcription

curl http://localhost:8603/v1/audio/transcriptions \
  -H "Content-Type: multipart/form-data" \
  -F file="@audio.wav" \
  -F model="faster-whisper/large-v3"

Streaming Transcription

To enable streaming, add the stream=true parameter:

curl http://localhost:8603/v1/audio/transcriptions \
  -H "Content-Type: multipart/form-data" \
  -F file="@audio.wav" \
  -F model="faster-whisper/large-v3" \
  -F stream="true"

Parameters

• file: Audio file to transcribe
• model: Model identifier (e.g., "faster-whisper/large-v3")
• language: Language code (optional)
• prompt: Initial prompt for the transcription (optional)
• response_format: Output format ("json", "text", "srt", "verbose_json", "vtt")
• temperature: Sampling temperature (default: 0.0)
• timestamp_granularities: Timestamp granularities (optional)

Supported Models

• Faster Whisper: faster-whisper/model-name
• FunASR: funasr/model-name
• SenseVoice: sensevoice/model-name

Configuration

The config.yaml file controls server and model settings:

server:
  host: localhost
  port: 8603
  model:
    - faster-whisper:
       - large-v3:   
           device: cpu
           compute_type: int8
           path: "."

SDK Usage

from asr_fusion import ASRFusionClient

# Initialize client
client = ASRFusionClient(base_url="http://localhost:8603")

# Transcribe file
result = client.transcribe_file(
    file_path="audio.wav",
    model="faster-whisper/large-v3"
)

print(result)

# Streaming transcription
stream_result = client.transcribe_file(
    file_path="audio.wav",
    model="faster-whisper/large-v3",
    stream=True
)

# Process streaming results
for chunk in stream_result:
    if chunk["type"] == "transcript.text.delta":
        print(f"Partial result: {chunk['delta']}")
    elif chunk["type"] == "transcript.text.done":
        print(f"Final result: {chunk['text']}")

## Development

### Project Structure

asr-fusion/ ├── asr_fusion/ │ ├── api/ # API server implementation │ ├── routers/ # API route definitions │ ├── models/ # Model implementations │ ├── config/ # Configuration handling │ └── sdk/ # Client SDK ├── client/ # Test clients ├── config.yaml # Configuration file └── main.py # Entry point

### Adding New Models

1. Create a new model implementation in `asr_fusion/models/`
2. Update `asr_fusion/models/model_manager.py` to support the new model
3. Add configuration to `config.yaml`

## License

MIT License