app.py

import gradio as gr
import torch
from transformers import pipeline
import torchaudio

# Check for CUDA availability and set device
if torch.cuda.is_available():
    device = "cuda"
else:
    device = "cpu"

# Load the Whisper pipeline
whisper_pipeline = pipeline("automatic-speech-recognition", model="openai/whisper-large-v3", device=device)

def transcribe_audio(audio_file):
    if audio_file is None:
        return "Please upload or record an audio file."

    try:
        # Load audio using torchaudio to handle various formats and long files
        audio, sample_rate = torchaudio.load(audio_file)

        # Resample if necessary (Whisper often expects 16kHz)
        if sample_rate != 16000:
            resampler = torchaudio.transforms.Resample(sample_rate, 16000)
            audio = resampler(audio)

        # Convert to Mono
        if audio.shape[0] > 1:  # Check if multi-channel
            audio = torch.mean(audio, dim=0, keepdim=True) # Average channels

        # Long-Form Transcription with Timestamps
        transcription = whisper_pipeline(audio.squeeze().numpy(), return_timestamps=True)

        # Format the output with timestamps (Improved)
        formatted_transcription = ""
        for segment in transcription["chunks"]:
            start = segment["timestamp"][0]
            end = segment["timestamp"][1]
            text = segment["text"]
            formatted_transcription += f"[{start:.2f} - {end:.2f}] {text}\n"  # Nicer formatting

        return formatted_transcription

    except Exception as e:
        return f"An error occurred: {e}"


with gr.Blocks() as demo:
    with gr.Row():
        audio_input = gr.Audio(type="filepath", label="Upload or Record Audio")

    transcribe_button = gr.Button("Transcribe")
    transcription_output = gr.Textbox(label="Transcription")

    transcribe_button.click(transcribe_audio, inputs=audio_input, outputs=transcription_output)

demo.launch()