Swecha Gonthuka Fine-Tuned Telugu ASR
Overview
This repository contains a fine-tuned Automatic Speech Recognition (ASR) model for the Telugu language.
The model is built on Swecha Gonthuka ASR and further fine-tuned using a custom Telugu speech dataset to improve transcription quality and recognition accuracy for Telugu speech.
This model converts spoken Telugu audio into Telugu Unicode text and can be used in a wide range of speech recognition applications.
Features
- 🎤 Telugu Automatic Speech Recognition (ASR)
- 📝 Speech-to-Text Conversion
- ⚡ Fine-tuned Wav2Vec2ForCTC Model
- 🤗 Compatible with Hugging Face Transformers
- 🎧 Supports 16 kHz WAV Audio
- 📚 Custom Telugu Speech Dataset Fine-tuning
Model Details
| Property | Value |
|---|---|
| Model Name | swecha-gonthuka-finetuned |
| Hugging Face Repository | Vasantha23/swecha-gonthuka-finetuned |
| Model Type | Wav2Vec2ForCTC |
| Framework | Hugging Face Transformers |
| Base Model | swechatelangana/swecha-gonthuka-asr |
| Language | Telugu |
| Task | Automatic Speech Recognition |
| Fine-tuned By | Vasantha |
Base Model
This model is fine-tuned from:
swechatelangana/swecha-gonthuka-asr
Model Lineage
Harveenchadha/wav2vec2-pretrained-clsril-23-10k
↓
swechatelangana/swecha-gonthuka-asr
↓
Vasantha23/swecha-gonthuka-finetuned
After Hugging Face processes the repository metadata, this lineage should automatically appear under the Model Tree section.
Fine-Tuning Dataset
The model was fine-tuned using a custom Telugu speech dataset.
Dataset Characteristics
- Language: Telugu
- Audio Format: WAV
- Sampling Rate: 16 kHz
- Mono Audio
- UTF-8 Telugu Transcriptions
- Task: Automatic Speech Recognition
Training Configuration
| Hyperparameter | Value |
|---|---|
| Architecture | Wav2Vec2ForCTC |
| Framework | Hugging Face Transformers |
| Optimizer | AdamW |
| Learning Rate | 1e-4 |
| Batch Size | 8 |
| Epochs | 10 |
| Mixed Precision | FP16 |
| Loss Function | CTC Loss |
Installation
pip install transformers torchaudio librosa soundfile
Usage
Load Processor and Model
from transformers import AutoProcessor, AutoModelForCTC
processor = AutoProcessor.from_pretrained(
"Vasantha23/swecha-gonthuka-finetuned"
)
model = AutoModelForCTC.from_pretrained(
"Vasantha23/swecha-gonthuka-finetuned"
)
Speech Recognition Pipeline
from transformers import pipeline
pipe = pipeline(
"automatic-speech-recognition",
model="Vasantha23/swecha-gonthuka-finetuned"
)
result = pipe("audio.wav")
print(result["text"])
Manual Inference
import librosa
import torch
from transformers import AutoProcessor, AutoModelForCTC
processor = AutoProcessor.from_pretrained(
"Vasantha23/swecha-gonthuka-finetuned"
)
model = AutoModelForCTC.from_pretrained(
"Vasantha23/swecha-gonthuka-finetuned"
)
speech, sr = librosa.load("audio.wav", sr=16000)
inputs = processor(
speech,
sampling_rate=16000,
return_tensors="pt",
padding=True
)
with torch.no_grad():
logits = model(inputs.input_values).logits
predicted_ids = torch.argmax(logits, dim=-1)
transcription = processor.batch_decode(predicted_ids)[0]
print(transcription)
Example
Input
audio.wav
Output
ఈ రోజు వాతావరణం చాలా బాగుంది.
Applications
This model can be used for:
- Telugu Speech Recognition
- Speech-to-Text Systems
- Voice Assistants
- Audio Transcription
- Subtitle Generation
- Educational Applications
- Research in Telugu ASR
- Accessibility Tools
Performance
The model has been fine-tuned to improve Telugu speech recognition performance over the base model using a custom Telugu dataset.
Typical evaluation metrics include:
- Word Error Rate (WER)
- Character Error Rate (CER)
Limitations
Performance may decrease in the following situations:
- Heavy background noise
- Low-quality recordings
- Multiple speakers
- Strong regional accents
- Very short audio clips
Acknowledgements
This project builds upon the excellent work of:
- Swecha Telangana
- Hugging Face
- Hugging Face Transformers
- PyTorch
- Wav2Vec2
- Mozilla Common Voice Community
Special thanks to the creators of Swecha Gonthuka ASR for making the base Telugu ASR model publicly available.
Citation
If you use this model in your work, please cite:
@misc{swecha_gonthuka_finetuned,
title={Swecha Gonthuka Fine-Tuned Telugu ASR},
author={Vasantha},
year={2026},
publisher={Hugging Face},
url={https://huggingface.co/Vasantha23/swecha-gonthuka-finetuned}
}
License
This model is released under the AGPL-3.0 License, consistent with the licensing of the base model.
Contact
For questions, suggestions, or improvements, please open an issue in the Hugging Face repository or contact the repository maintainer.
Hugging Face Repository