app.py

import torch
import torchaudio
import numpy as np
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
from pydub import AudioSegment
import os
import gradio as gr

# Load the model and processor
model_id = "hackergeek98/whisper-fa-tinyyy"
device = "cuda" if torch.cuda.is_available() else "cpu"

model = AutoModelForSpeechSeq2Seq.from_pretrained(model_id).to(device)
processor = AutoProcessor.from_pretrained(model_id)

# Create ASR pipeline
pipe = pipeline(
    "automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    device=0 if torch.cuda.is_available() else -1,
)

# Convert audio to WAV format
def convert_to_wav(audio_path):
    audio = AudioSegment.from_file(audio_path)
    audio = audio.set_channels(1)  # Ensure mono audio
    wav_path = "converted_audio.wav"
    audio.export(wav_path, format="wav")
    return wav_path

# Split long audio into chunks
def split_audio(audio_path, chunk_length_ms=30000):  # Default: 30 sec per chunk
    audio = AudioSegment.from_wav(audio_path)
    chunks = [audio[i:i+chunk_length_ms] for i in range(0, len(audio), chunk_length_ms)]
    chunk_paths = []
    
    for i, chunk in enumerate(chunks):
        chunk_path = f"chunk_{i}.wav"
        chunk.export(chunk_path, format="wav")
        chunk_paths.append(chunk_path)
    
    return chunk_paths

# **🔹 Fixed: Convert Stereo to Mono Before Processing**
def transcribe_audio_chunk(chunk_path):
    waveform, sampling_rate = torchaudio.load(chunk_path)  # Load audio
    if waveform.shape[0] > 1:  # If stereo (more than 1 channel)
        waveform = torch.mean(waveform, dim=0, keepdim=True)  # Convert to mono
    waveform = waveform.numpy()  # Convert to numpy
    result = pipe({"raw": waveform, "sampling_rate": sampling_rate})  # Pass raw data
    return result["text"]

# Transcribe a long audio file
def transcribe_long_audio(audio_path):
    wav_path = convert_to_wav(audio_path)
    chunk_paths = split_audio(wav_path)
    transcription = ""
    
    for chunk in chunk_paths:
        transcription += transcribe_audio_chunk(chunk) + "\n"
        os.remove(chunk)  # Remove processed chunk
    
    os.remove(wav_path)  # Cleanup original file
    
    return transcription

# Gradio interface
def transcribe_interface(audio_file):
    if not audio_file:
        return "No file uploaded."
    return transcribe_long_audio(audio_file)

iface = gr.Interface(
    fn=transcribe_interface,
    inputs=gr.Audio(type="filepath"),
    outputs="text",
    title="Whisper ASR - Transcription",
    description="Upload an audio file, and the model will transcribe it."
)

if __name__ == "__main__":
    iface.launch()