Swecha Telangana ASR Base Model
Model Description
The Swecha Telangana ASR Base Model is a fine-tuned Automatic Speech Recognition (ASR) model built on OpenAI Whisper Small for Telugu speech recognition.
This model has been fine-tuned using the Mozilla Common Voice Scripted Speech 26.0 Telugu dataset through the Mozilla AI Speech-to-Text Finetuning Blueprint.
Base Model
- openai/whisper-small
Language
- Telugu (te)
Training Dataset
- Mozilla Common Voice Scripted Speech 26.0 – Telugu
- License: CC0-1.0
Training Configuration
| Parameter | Value |
|---|---|
| Base Model | Whisper Small |
| Language | Telugu |
| Learning Rate | 1e-5 |
| Batch Size | 32 |
| Gradient Accumulation Steps | 1 |
| Warmup Steps | 50 |
| Max Steps | 50 |
| Mixed Precision | FP16 |
| Evaluation Metric | Word Error Rate (WER) |
Intended Uses
- Telugu Speech Recognition
- Speech-to-Text Applications
- Voice Assistants
- Accessibility Tools
- Educational Applications
- Automatic Caption Generation
- Research in Indian Language ASR
Limitations
- Trained only on Mozilla Common Voice Telugu speech.
- Performance may decrease for:
- Noisy environments
- Code-mixed speech
- Unseen accents
- Domain-specific vocabulary
Evaluation Results
| Metric | Before Fine-tuning | After Fine-tuning |
|---|---|---|
| WER | 211.71 | 75.92 |
| CER | 174.02 | 46.38 |
The fine-tuned model significantly improves Telugu speech recognition compared to the baseline Whisper Small model.
Citation
If you use this model, please cite:
- Mozilla Common Voice
- Mozilla AI Speech-to-Text Finetuning Blueprint
- OpenAI Whisper
Author
Akhila Vishwanath
Hugging Face: https://huggingface.co/akhila56