app.py

import gradio as gr
import numpy as np
import torch
import torchaudio
from transformers import AutoProcessor, AutoModelForSpeechSeq2Seq, WhisperTokenizer

# Define paths to the model and processor
model_name = "userdata/whisper-largeV2-03-ms-v11-LORA-Merged"

# Load the processor and model
processor = AutoProcessor.from_pretrained(model_name)
tokenizer = WhisperTokenizer.from_pretrained(model_name)
model = AutoModelForSpeechSeq2Seq.from_pretrained(model_name)

# Check and set the device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# Function to chunk the audio
def chunk_audio(audio, chunk_length):
    num_chunks = len(audio) // chunk_length + (1 if len(audio) % chunk_length > 0 else 0)
    return [audio[i * chunk_length:(i + 1) * chunk_length] for i in range(num_chunks)]

# Function to transcribe an audio file
def transcribe(audio_path, chunk_length=16000 * 30):  # 30 seconds chunks
    # Load audio
    speech_array, sampling_rate = torchaudio.load(audio_path)
    # Resample to 16kHz
    resampler = torchaudio.transforms.Resample(sampling_rate, 16000)
    speech = resampler(speech_array).squeeze().numpy()

    # Chunk the audio if it's too long
    chunks = chunk_audio(speech, chunk_length)

    # Transcribe each chunk
    transcriptions = []
    for chunk in chunks:
        # Process the audio
        inputs = processor(chunk, sampling_rate=16000, return_tensors="pt")
        inputs = {key: value.to(device).to(torch.float16) for key, value in inputs.items()}  # Convert to float16

        # Generate token IDs
        with torch.no_grad():
            generated_ids = model.generate(inputs["input_features"], max_length=448)

        # Decode the token IDs to text
        transcription = tokenizer.decode(generated_ids[0], skip_special_tokens=True)
        transcriptions.append(transcription)

    # Combine transcriptions
    full_transcription = ' '.join(transcriptions)
    return full_transcription

# Create the Gradio interface
iface = gr.Interface(
    fn=transcribe,  # Update to match the function name
    inputs=gr.Audio(type="filepath"),
    outputs=gr.Textbox(),
    title="Audio Transcription App",
    description="Upload an audio file to get a transcription."
)

# Launch the Gradio interface
iface.launch()