Swecha Gonthuka ASR Finetuned
Model Details
Model Description
This model is a fine-tuned version of the Swecha Gonthuka ASR model for Telugu Automatic Speech Recognition (ASR). It has been fine-tuned on a custom Telugu speech dataset containing approximately 5,000 audio-transcription pairs.
- Developed by: Venigalla Shamanth Chowdary
- Base model:
swechatelangana/swecha-gonthuka-asr - Model type: Wav2Vec2ForCTC
- Language: Telugu (te)
- License: Apache-2.0 (same as the base model, if applicable)
- Framework: Hugging Face Transformers + PyTorch
Model Sources
- Base Model: https://huggingface.co/swechatelangana/swecha-gonthuka-asr
- Fine-tuned Model Repository: https://huggingface.co/ShamanthChowdary/Swecha_Gonthuka_ASR_Finetuned
Uses
Direct Use
This model is intended for:
- Telugu Speech-to-Text
- Voice assistants
- Automatic transcription
- Telugu speech datasets
- Research in Indian language ASR
- Educational and accessibility applications
Downstream Use
Possible downstream applications include:
- Subtitle generation
- Meeting transcription
- Voice search
- Speech analytics
- Telugu conversational AI
- Voice-enabled applications
Out-of-Scope Use
This model is not intended for:
- Languages other than Telugu
- Speaker identification
- Emotion recognition
- Medical or legal transcription where high accuracy is required
- Safety-critical applications
Bias, Risks and Limitations
Performance depends on:
- Audio quality
- Background noise
- Speaker accent
- Recording device
- Speech speed
The model may perform less accurately on unseen accents or noisy recordings.
How to Get Started
from transformers import AutoProcessor, AutoModelForCTC
processor = AutoProcessor.from_pretrained(
"ShamanthChowdary/Swecha_Gonthuka_ASR_Finetuned"
)
model = AutoModelForCTC.from_pretrained(
"ShamanthChowdary/Swecha_Gonthuka_ASR_Finetuned"
)
Training Details
Training Data
The model was fine-tuned using a custom Telugu speech dataset containing approximately:
- 5000 audio samples
- WAV format
- Sampling rate: 16 kHz
- Human-annotated Telugu transcriptions
Training Procedure
Preprocessing
- Audio resampled to 16 kHz
- Audio-token alignment using Wav2Vec2 Processor
- Text tokenization using the processor tokenizer
Training Hyperparameters
| Parameter | Value |
|---|---|
| Epochs | 10 |
| Learning Rate | 1e-5 |
| Batch Size | 4 |
| Gradient Accumulation | 2 |
| Warmup Steps | 500 |
| Optimizer | AdamW |
| Mixed Precision | FP16 |
Evaluation
Metrics
Word Error Rate (WER)
Results
| Metric | Value |
|---|---|
| Validation Loss | 0.2603 |
| Training Loss | 0.5500 |
| Word Error Rate (WER) | 0.4595 |
Model Architecture
- Architecture: Wav2Vec2ForCTC
- Framework: Transformers
- Backend: PyTorch
Hardware
Training performed using:
- Google Colab
- NVIDIA Tesla T4 GPU
Software
- Python
- PyTorch
- Hugging Face Transformers
- Datasets
- Gradio
More Information
This model was developed as part of a Telugu Automatic Speech Recognition fine-tuning project using the Swecha Gonthuka ASR base model.
Author
Venigalla Shamanth Chowdary
GitHub: https://github.com/shamanth-25