Gonthuka Telugu ASR Fine-Tuned Model
Overview
This repository contains a fine-tuned Wav2Vec2 Automatic Speech Recognition (ASR) model for the Telugu language.
The model has been fine-tuned on a custom Telugu speech dataset collected as part of the Gonthuka Project. It converts spoken Telugu audio into Telugu text and can be used for transcription, subtitle generation, accessibility applications, and speech-enabled systems.
Base Model
- Model: swechatelangana/swecha-gonthuka-asr
- Architecture: Wav2Vec2ForCTC
- Framework: Hugging Face Transformers
- Language: Telugu
Fine-Tuning Dataset
The model was fine-tuned using a custom Telugu speech dataset.
Dataset Details
| Property | Value |
|---|---|
| Language | Telugu |
| Audio Format | WAV |
| Sampling Rate | 16 kHz |
| Channels | Mono |
| Task | Automatic Speech Recognition |
Training Details
| Parameter | Value |
|---|---|
| Architecture | Wav2Vec2ForCTC |
| Framework | Hugging Face Transformers |
| Task | Speech-to-Text |
| Language | Telugu |
Evaluation
The model was evaluated using standard Automatic Speech Recognition metrics.
| Metric | Description |
|---|---|
| WER | Word Error Rate |
| CER | Character Error Rate |
Usage
Install
pip install transformers torchaudio librosa
Load the Model
from transformers import Wav2Vec2Processor
from transformers import Wav2Vec2ForCTC
processor = Wav2Vec2Processor.from_pretrained(
"suryateja2109/swecha_gonthuka_finetuned_asr"
)
model = Wav2Vec2ForCTC.from_pretrained(
"suryateja2109/swecha_gonthuka_finetuned_asr"
)
Inference
import torch
import librosa
speech, sr = librosa.load("sample.wav", sr=16000)
inputs = processor(
speech,
sampling_rate=16000,
return_tensors="pt",
padding=True
)
with torch.no_grad():
logits = model(**inputs).logits
predicted_ids = torch.argmax(logits, dim=-1)
transcription = processor.batch_decode(predicted_ids)
print(transcription[0])
Applications
- Telugu Speech Recognition
- Voice Assistants
- Speech-to-Text Systems
- Subtitle Generation
- Audio Transcription
- Accessibility Applications
- Educational Tools
Limitations
- Best suited for Telugu speech.
- Performance may decrease with noisy recordings.
- Mixed-language speech may reduce transcription accuracy.
- Larger and more diverse datasets can further improve performance.
Future Work
- Increase dataset size
- Improve Word Error Rate (WER)
- Support noisy environments
- Add multilingual support
- Optimize for real-time inference
Citation
If you use this model in your research or applications, please cite this repository.
@misc{gonthuka_asr,
author = {Surya Teja},
title = {Gonthuka Telugu ASR Fine-Tuned Model},
year = {2026},
publisher = {Hugging Face},
url = {https://huggingface.co/suryateja2109/swecha_gonthuka_finetuned_asr}
}
License
This project is released under the Apache-2.0 License.
Author
Surya Teja
B.Tech Data Science
Fine-tuned as part of the Gonthuka Telugu ASR Project.