---
license: mit
datasets:
- mozilla-foundation/common_voice_17_0
language:
- en
- ta
metrics:
- wer
base_model:
- openai/whisper-small
pipeline_tag: automatic-speech-recognition
library_name: transformers
tags:
- language-identification
- speech-to-text 



---

# Whisper-small-ta
This model is trainned for voice to text trancription for tamil language

## Model Overview
This model is fine-tuned from `openai/whisper-small` using the [Mozilla Common Voice 17.0](https://huggingface.co/datasets/mozilla-foundation/common_voice_17_0) dataset for language identification and transcription in **Tamil** . The model is designed to accurately transcribe spoken audio into text and identify whether the language is Tamil .

### Key Features:
- **Languages**: Tamil
- **Base Model**: Whisper-small from OpenAI
- **Dataset**: Mozilla Common Voice 17.0

## Intended Use
The model is designed for automatic speech recognition (ASR) in Tamil, making it suitable for transcription and language identification in real-time applications.

## Training Details
This model was fine-tuned using a subset of the Mozilla Common Voice dataset. The dataset contains '53,468 ' samples  

### Fine-tuning Process:
- The fine-tuning was performed on `Whisper-small`, a smaller version of OpenAI's Whisper model, for reduced latency and higher accuracy for low-resource languages.
- The model was trained for `2` epochs on a `Google Colab Pro` environment.

## Performance
The model achieved a **Word Error Rate (WER)** of `34%` , using a validation dataset with `8` hours of audio.
We expect further improvements with continued training.

## Usage
You can use this model with the following code:

```python
from transformers import WhisperForConditionalGeneration, WhisperProcessor
import torch

model = WhisperForConditionalGeneration.from_pretrained("Lingalingeswaran/whisper-small-ta")
processor = WhisperProcessor.from_pretrained("Lingalingeswaran/whisper-small-ta")

# Example audio input
audio = "path_to_audio_file"

inputs = processor(audio, return_tensors="pt", padding="longest")
with torch.no_grad():
    predicted_ids = model.generate(inputs.input_ids)
    
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)
print(transcription)