Telugu ASR Fine-Tuned Model
Overview
This repository contains a fine-tuned Automatic Speech Recognition (ASR) model for the Telugu language.
The model is based on Swecha Gonthuka ASR and has been fine-tuned using a custom Telugu speech dataset.
The model is capable of converting spoken Telugu audio into Telugu text and can be used for speech transcription and related ASR applications.
Base Model
- Model: swechatelangana/swecha-gonthuka-asr
- Architecture: Wav2Vec2ForCTC
- Framework: Hugging Face Transformers
- Language: Telugu
Fine-Tuning Dataset
A custom Telugu speech dataset was used for fine-tuning.
Dataset details:
- Approximately 752 Telugu audio recordings
- Audio Format: WAV
- Sampling Rate: 16 kHz
- Mono Audio
- Telugu Transcriptions
The dataset contains natural Telugu speech with manually verified transcripts.
Training Configuration
| Parameter | Value |
|---|---|
| Base Model | swechatelangana/swecha-gonthuka-asr |
| Epochs | 10 |
| Learning Rate | 1e-5 |
| Batch Size | 4 |
| Gradient Accumulation | 2 |
| Optimizer | AdamW |
| Framework | Hugging Face Transformers |
| GPU | NVIDIA Tesla T4 |
Evaluation
The model was evaluated using Character Error Rate (CER).
| Metric | Value |
|---|---|
| CER | 0.14 |
| Character Accuracy | 86% |
Character Accuracy = (1 − CER) × 100
Usage
Install dependencies
pip install transformers torch librosa
Load the model
from transformers import pipeline
pipe = pipeline(
"automatic-speech-recognition",
model="Pravalika2711/gonthuka-asr-finetuned",
feature_extractor="Pravalika2711/gonthuka-asr-finetuned",
tokenizer="Pravalika2711/gonthuka-asr-finetuned"
)
Run inference
result = pipe("audio.wav")
print(result["text"])
Applications
- Telugu Speech Recognition
- Voice Assistants
- Speech-to-Text Systems
- Accessibility Applications
- Telugu Language Research
- Educational Applications
Limitations
- Performance may decrease for noisy environments.
- Works best with 16 kHz mono WAV audio.
- Accuracy depends on speaker accent and recording quality.
Acknowledgements
This work is built upon the excellent Swecha Gonthuka ASR model.
Base Model:
https://huggingface.co/swechatelangana/swecha-gonthuka-asr
Citation
If you use this model, please cite both the original Swecha Gonthuka ASR model and this fine-tuned model.
@misc{gonthuka_finetuned,
title={Fine-Tuned Telugu ASR using Swecha Gonthuka ASR},
author={Pravalika Moraboyina},
year={2026},
publisher={Hugging Face}
}