ViStreamASR 0.1.3

Vietnamese Streaming Automatic Speech Recognition Library

🎙️ ViStreamASR - Real-Time Vietnamese Speech Recognition

ViStreamASR is a simple Vietnamese Streaming Automatic Speech Recognition library for real-time audio processing.

Features

• 🎯 Streaming ASR: Real-time audio processing with configurable chunk sizes
• 🇻🇳 Vietnamese Optimized: Specifically designed for Vietnamese speech recognition
• 📦 Simple API: Easy-to-use interface with minimal setup
• ⚡ High Performance: CPU/GPU support

Installation

pip install ViStreamASR

Installation from Source

For development or to use the latest version:

# Clone the repository
git clone https://github.com/nguyenvulebinh/ViStreamASR.git
cd ViStreamASR

# Install dependencies
pip install -r requirements.txt

# Option 1: Use directly from source
python test_library.py  # Test the installation

# Option 2: Install in development mode
pip install -e .

Using from Source

When using from source, import the modules directly:

import sys
sys.path.insert(0, 'src')
from streaming import StreamingASR

# Initialize and use
asr = StreamingASR()
for result in asr.stream_from_file("audio.wav"):
    print(result['text'])

Quick Start

Python API

from ViStreamASR import StreamingASR

# Initialize ASR
asr = StreamingASR()

# Process audio file
for result in asr.stream_from_file("audio.wav"):
    if result['partial']:
        print(f"Partial: {result['text']}")
    if result['final']:
        print(f"Final: {result['text']}")

Command Line

# Basic transcription
vistream-asr transcribe audio.wav

API Reference

StreamingASR

from ViStreamASR import StreamingASR

# Initialize with options
asr = StreamingASR(
    chunk_size_ms=640,           # Chunk size in milliseconds
    auto_finalize_after=15.0,    # Auto-finalize after seconds
    debug=False                  # Enable debug logging
)

# Stream from file
for result in asr.stream_from_file("audio.wav"):
    # result contains:
    # - 'partial': True for partial results
    # - 'final': True for final results
    # - 'text': transcription text
    # - 'chunk_info': processing information
    pass

Advanced Usage

For low-level control:

from ViStreamASR import ASREngine

engine = ASREngine(chunk_size_ms=640, debug_mode=True)
engine.initialize_models()

# Process audio chunks directly
result = engine.process_audio(audio_chunk, is_last=False)

Model Information

• Language: Vietnamese
• Architecture: U2-based streaming ASR
• Model Size: ~2.7GB (cached after first download)
• Sample Rate: 16kHz (automatically converted)
• Optimal Chunk Size: 640ms

How U2 Streaming Works

The following picture shows how U2 (Unified Streaming and Non-streaming) architecture works:

The U2 model enables both streaming and non-streaming ASR in a unified framework, providing low-latency real-time transcription while maintaining high accuracy.

Performance

• RTF: ~0.34x (faster than real-time)
• Latency: ~640ms with default settings
• GPU Support: Automatic CUDA acceleration when available

Limitations

• Audio Input Assumption: The system assumes audio input is speech. Non-speech audio may produce unexpected results.
• Production Recommendation: For practical use, it's recommended to add VAD (Voice Activity Detection) before running streaming ASR to reduce ASR streaming load and improve efficiency.

CLI Commands

# Transcription
vistream-asr transcribe <file>                    # Basic transcription
vistream-asr transcribe <file> --chunk-size 640   # Custom chunk size
vistream-asr transcribe <file> --no-debug         # Clean output

# Information
vistream-asr info                                  # Library info
vistream-asr version                               # Version

Requirements

System Requirements

• RAM: Minimum 5GB RAM
• CPU: Minimum 2 cores
• Performance: RTF 0.3-0.4x achievable on CPU-only systems meeting above specs
• GPU: Supports GPU acceleration for better performance, but CPU-only operation still achieves RTF 0.3-0.4x

Software Requirements

• Python 3.8+
• PyTorch 2.5+
• TorchAudio 2.5+
• NumPy 1.19.0+
• Requests 2.25.0+
• flashlight-text
• librosa

License

MIT License