app.py

import os
import gradio as gr
from transformers import pipeline
import librosa
import numpy as np

# Ensure offline mode is used
os.environ["TRANSFORMERS_OFFLINE"] = "1"

# Model name (must already be cached)
MODEL_NAME = "Rezuwan/regional_asr_weights"

# Load the ASR pipeline from local cache only
transcriber = pipeline(
    "automatic-speech-recognition",
    model=MODEL_NAME,
)

# Transcription function
def transcribe_audio(audio_path):
    try:
        audio_data, sample_rate = librosa.load(audio_path, sr=16000)
        audio_data = librosa.to_mono(audio_data) if audio_data.ndim > 1 else audio_data
        audio_data = audio_data.astype(np.float32)
        audio_data /= np.max(np.abs(audio_data))
        result = transcriber(audio_data)
        return result["text"]
    except Exception as e:
        return f"Error: {str(e)}"

# Gradio UI
iface = gr.Interface(
    fn=transcribe_audio,
    inputs=gr.Audio(type="filepath", label="Upload or Record Audio"),
    outputs=gr.Textbox(label="Transcription"),
    title="Bengali Speech-to-Text with Regional Dialects",
    description=(
        f"""
        Model Card: [{MODEL_NAME}](https://huggingface.co/{MODEL_NAME}) and 🤗 Transformers to transcribe audio files of arbitrary length.
        
        Instructions:
        1. Record or upload an audio file using the left panel.
        2. Click 'Submit' after waveform appears.
        3. Wait for processing and see the result on the right.
        
        Notes:
        - This model handles Bengali speech with regional dialects.
        - Accuracy may vary due to limited training data.
        - Offline mode is enabled for isolated environments.
        """
    )
)

# Launch the Gradio app
iface.launch()