README.md

metadata

title: asr-multi-model
emoji: 😊
colorFrom: blue
colorTo: blue
sdk: gradio
sdk_version: 4.44.0
app_file: app.py
pinned: false

🎤 Multi-Model ASR Speech Recognition

A comprehensive Automatic Speech Recognition (ASR) application with support for multiple models and Word Error Rate (WER) analysis.

✨ Features

• 🤖 Multiple Models: Support for Wav2Vec2 and Whisper models
• 🎤 Audio Recording: Direct microphone recording
• 📁 File Upload: Support for various audio formats
• 📊 WER Analysis: Calculate Word Error Rate with detailed breakdown
• 💾 Memory Efficient: Dynamic model loading and cleanup
• 🌍 Multilingual: Whisper models support multiple languages

🚀 Live Demo

This application is hosted on Hugging Face Spaces. You can access it at: [Your HF Spaces URL]

🤖 Available Models

Model	Type	Speed	Accuracy	Memory	Best For
Wav2Vec2 Base (100h)	Wav2Vec2	⚡ Fast	🟡 Good	~300MB	Basic tasks
Wav2Vec2 Base (960h)	Wav2Vec2	🟡 Balanced	🟢 Better	~1GB	General use
Wav2Vec2 Large (960h)	Wav2Vec2	🐌 Slower	🔴 High	~3GB	Difficult audio
Whisper Large V3 Turbo	Whisper	🐌 Slower	🔴 Best	~5GB	Multilingual

📖 How to Use

1. Select Model: Choose from available Wav2Vec2 and Whisper models
2. Load Model: Click 'Load Model' to load the selected model
3. Record/Upload: Record audio or upload an audio file
4. Transcribe: Click 'Transcribe' or wait for auto-transcription
5. WER Analysis: Enter reference text to calculate Word Error Rate
6. Copy Text: Use 'Copy Text' to copy the result

🔧 Technical Details

Models Used

• Wav2Vec2: Facebook's self-supervised speech recognition models
• Whisper: OpenAI's multilingual speech recognition model

Audio Processing

• Automatic resampling to 16kHz
• Mono conversion for stereo audio
• Audio normalization
• Support for various formats (MP3, WAV, M4A, FLAC)

WER Calculation

• Uses edit distance for accurate alignment
• Normalizes text (lowercase, no punctuation)
• Provides detailed breakdown of insertions, deletions, and substitutions

🛠️ Local Development

Prerequisites

• Python 3.8+
• CUDA-compatible GPU (optional, for faster inference)

Installation

git clone [your-repo-url]
cd [your-repo-name]
pip install -r requirements.txt
python app.py

Requirements

gradio>=4.44.0
torch>=2.6.0
torchaudio>=2.6.0
transformers>=4.36.2
librosa>=0.10.1
soundfile>=0.12.1
numpy>=1.24.3
editdistance>=1.0.11

📊 WER Analysis

The application provides detailed Word Error Rate analysis:

• Word Error Rate: Percentage of errors
• Error Breakdown: Insertions, deletions, substitutions
• Word Statistics: Correct words, total words, accuracy
• Normalized Texts: Shows processed texts for verification

🎯 Performance Tips

• Clear Speech: Speak clearly for better accuracy
• Quiet Environment: Minimize background noise
• Good Microphone: Use quality audio input
• Model Selection: Choose based on your needs (speed vs accuracy)

🤝 Contributing

Feel free to submit issues, feature requests, or pull requests to improve this application.

📝 License

This project is open source and available under the MIT License.

🙏 Acknowledgments

• Hugging Face for the transformers library
• Facebook for Wav2Vec2 models
• OpenAI for Whisper models
• Gradio for the web interface framework