Hear the world. Transcribe it instantly.

Effortlessly transform live audio into text with our fine-tuned Whisper and Wav2Vec2 models. Built for real-time speech recognition, this project brings low-latency transcription to life—perfect for conversations, streaming, or any dynamic audio environment.

Powerful. Simple. Yours.

• Capture live 16kHz mono audio with ease.
• Fine-tune models in real time for unmatched accuracy.
• Switch seamlessly between Whisper and Wav2Vec2.
• Save and load models effortlessly with timestamped precision.
• English transcription, optimized for clarity.

Get Started

Clone the repository and set up in minutes:

git clone https://github.com/bniladridas/speech-model.git
cd speech-model
pip install -r requirements.txt

For Linux, install audio dependencies:

sudo apt-get install libsndfile1

Bring Your Voice to Life

Fine-Tune Models
Train on live audio with a single command:

python main.py --model whisper
# or
python main.py --model wav2vec2

Press Ctrl+C to save your fine-tuned model automatically.

Transcribe Audio
Test your model with real-time transcription:

python test_transcription.py --model whisper
# or
python test_transcription.py --model wav2vec2

Records 5 seconds of audio (customizable) and delivers instant text.

Save Your Work
Models are saved to:
models/speech_recognition_ai_fine_tune_[model]_[timestamp]
Customize the path:

export MODEL_SAVE_PATH="/your/path"
python main.py --model [whisper|wav2vec2]

What You Need

• Python 3.8+
• PyTorch (2.0.1 recommended)
• Transformers (4.35.0 recommended)
• Sounddevice (0.4.6)
• Torchaudio (2.0.1)
  A GPU accelerates fine-tuning. Full requirements in requirements.txt.

Behind the Magic

• Task: Automatic Speech Recognition (ASR)
• Models:
  • Whisper (openai/whisper-small)
  • Wav2Vec2 (facebook/wav2vec2-base-960h)
• Fine-Tuning: Optimized on live 16kHz mono audio with Adam optimizer (learning rate 1e-5).
• Input: 16kHz mono audio
• Output: Precise English transcription

Load models from Hugging Face:

from transformers import WhisperForConditionalGeneration, WhisperProcessor
model = WhisperForConditionalGeneration.from_pretrained("bniladridas/speech-recognition-ai-fine-tune")
processor = WhisperProcessor.from_pretrained("bniladridas/speech-recognition-ai-fine-tune")

Project Structure

• dataset.py: Audio recording and preprocessing
• train.py: Training pipeline
• test_transcription.py: Real-time transcription
• main.py: Core fine-tuning script
• requirements.txt: Dependencies
• README.md: This guide

Troubleshooting

Git Push Issues
For non-fast-forward errors:

git pull origin main --rebase
git push origin main

Stuck Processes
Find and stop processes:

ps aux | grep python
kill -9 <PID>

Large Files
Exclude models from Git:

echo "models/" >> .gitignore
git add .gitignore
git rm -r --cached models/
git commit -m "Update .gitignore"

Your Data, Your Models

No pre-existing dataset needed. Your live audio recordings fuel the fine-tuning process, making every model uniquely yours.

Join the Journey

Licensed under the MIT License. Contribute, refine, and share on GitHub.