w2vindia

w2vindia is a self-supervised speech representation model based on the Wav2Vec 2.0 Base architecture, trained from scratch on a multilingual corpus of Indian languages.

This model serves as a foundation acoustic model and does not generate text directly. It is intended for fine-tuning on downstream speech tasks such as ASR, phoneme recognition, or language identification.

Load the model

from transformers import Wav2Vec2Model, Wav2Vec2Processor

processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base")
model = Wav2Vec2Model.from_pretrained("balaragavesh/w2vindia")

🔍 Model Description

• Architecture: Wav2Vec 2.0 Base
• Training Type: Self-supervised pretraining
• Languages: Hindi, Tamil, Bengali, Marathi, Telugu, Kannada, Malayalam, Gujarati (and others)
• Sampling Rate: 16 kHz
• Framework: PyTorch / Hugging Face Transformers

Unlike language-specific models, this model was trained on a blind mixture of Indian languages without language identifiers, allowing it to learn shared phonetic and acoustic representations across languages.

📚 Dataset

The model was pre-trained on the IndicTTS dataset collection released by SPRING Lab, available on Hugging Face.

• Source: SPRINGLab / IndicTTS Datasets
• Total Duration: ~200 Hours
• Data Characteristics: Mixed Indian languages (Hindi, Tamil, Marathi, Bengali, etc.)
• Preprocessing:
  • Sampling Rate: 16kHz
  • Audio files were filtered to be between 2s and 15s in length to optimize attention mechanisms.
  • Combined into a unified training set to encourage cross-lingual acoustic transfer.

🚀 Intended Use

✅ Supported Tasks

• Automatic Speech Recognition (ASR)
• Phoneme Recognition
• Low-resource language modeling
• Cross-lingual transfer learning

❌ Not Intended For

• Direct speech-to-text inference without fine-tuning
• Speaker identification without adaptation

🔗 Link

• GitHub Repository:
  👉 https://github.com/balaragavesh/w2vindia

---