Whisper Small Sinhala

🚀 Whisper Small sinhala - Lingalingeswaran

This model is a fine - tuned version of openai/whisper-small on the Lingalingeswaran/asr-sinhala-dataset_json_v1 dataset, designed for Sinhala speech - related tasks.

🚀 Quick Start

This model is a fine - tuned version of openai/whisper-small on the Lingalingeswaran/asr-sinhala-dataset_json_v1 dataset.

✨ Features

Model description

This Whisper model has been fine - tuned specifically for the Sinhala language using the Common Voice 11.0 dataset. It is designed to handle tasks such as speech - to - text transcription and language identification, making it suitable for applications where Sinhala is a primary language of interest. The fine - tuning process focused on enhancing performance for Sinhala, aiming to reduce the error rate in transcriptions and improve general accuracy.

Intended uses & limitations

Intended Uses

Speech - to - text transcription in Sinhala

Limitations

• May not perform as well on languages or dialects that are not well - represented in the Common Voice dataset.
• Higher Word Error Rate (WER) in noisy environments or with speakers who have heavy accents not covered in the training data.
• The model is optimized for Sinhala; performance in other languages may be suboptimal.

📦 Installation

No specific installation steps are provided in the original document, so this section is skipped.

💻 Usage Examples

Basic Usage

Here is an example of how to use the model for Sinhala speech recognition with Gradio:

import gradio as gr
from transformers import pipeline

# Initialize the pipeline with the specified model
pipe = pipeline(model="Lingalingeswaran/whisper-small-sinhala")

def transcribe(audio):
    # Transcribe the audio file to text
    text = pipe(audio)["text"]
    return text

# Create the Gradio interface

iface = gr.Interface(
    fn=transcribe,
    inputs=gr.Audio(sources=["microphone", "upload"], type="filepath"),
    outputs="text",
    title="Whisper Small Sinhala",
    description="Realtime demo for Sinhala speech recognition using a fine - tuned Whisper small model.",
)

# Launch the interface
if __name__ == "__main__":
    iface.launch()

📚 Documentation

Training and evaluation data

The training data for this model consists of voice recordings in Sinhala from the Mozilla - foundation/Common Voice 11.0 dataset. The dataset is a crowd - sourced collection of transcribed speech, ensuring diversity in terms of speaker accents, age groups, and speech styles.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e - 05
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon = 1e - 08 and optimizer_args = No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• training_steps: 4000
• mixed_precision_training: Native AMP

Framework versions

• Transformers 4.48.1
• Pytorch 2.5.1+cu121
• Datasets 3.2.0
• Tokenizers 0.21.0

🔧 Technical Details

No specific technical details beyond what's already covered are provided, so this section is skipped.

📄 License

This model is released under the apache - 2.0 license.