app.py

import gradio as gr
from audio_processing import process_audio, print_results
from transformers import pipeline, AutoTokenizer, AutoModelForSeq2SeqLM, AutoModelForQuestionAnswering
import spaces
import torch

# Check if CUDA is available
cuda_available = torch.cuda.is_available()
device = "cuda" if cuda_available else "cpu"

# Initialize the summarization and question-answering models
summarizer_model = AutoModelForSeq2SeqLM.from_pretrained("facebook/bart-large-cnn").to(device)
summarizer_tokenizer = AutoTokenizer.from_pretrained("facebook/bart-large-cnn")

qa_model = AutoModelForQuestionAnswering.from_pretrained("distilbert-base-cased-distilled-squad").to(device)
qa_tokenizer = AutoTokenizer.from_pretrained("distilbert-base-cased-distilled-squad")

@spaces.GPU(duration=120)
def transcribe_audio(audio_file, translate, model_size):
    language_segments, final_segments = process_audio(audio_file, translate=translate, model_size=model_size)
    
    output = "Detected language changes:\n\n"
    for segment in language_segments:
        output += f"Language: {segment['language']}\n"
        output += f"Time: {segment['start']:.2f}s - {segment['end']:.2f}s\n\n"

    output += f"Transcription with language detection and speaker diarization (using {model_size} model):\n\n"
    full_text = ""
    for segment in final_segments:
        output += f"[{segment['start']:.2f}s - {segment['end']:.2f}s] ({segment['language']}) {segment['speaker']}:\n"
        output += f"Original: {segment['text']}\n"
        if translate:
            output += f"Translated: {segment['translated']}\n"
            full_text += segment['translated'] + " "
        else:
            full_text += segment['text'] + " "
        output += "\n"
    
    return output, full_text

@spaces.GPU(duration=120)
def summarize_text(text):
    inputs = summarizer_tokenizer(text, max_length=1024, truncation=True, return_tensors="pt").to(device)
    summary_ids = summarizer_model.generate(inputs["input_ids"], max_length=150, min_length=50, do_sample=False)
    summary = summarizer_tokenizer.decode(summary_ids[0], skip_special_tokens=True)
    return summary

@spaces.GPU(duration=120)
def answer_question(context, question):
    inputs = qa_tokenizer(question, context, return_tensors="pt").to(device)
    outputs = qa_model(**inputs)
    answer_start = torch.argmax(outputs.start_logits)
    answer_end = torch.argmax(outputs.end_logits) + 1
    answer = qa_tokenizer.decode(inputs["input_ids"][0][answer_start:answer_end])
    return answer

@spaces.GPU(duration=120)
def process_and_summarize(audio_file, translate, model_size):
    transcription, full_text = transcribe_audio(audio_file, translate, model_size)
    summary = summarize_text(full_text)
    return transcription, summary

@spaces.GPU(duration=120)
def qa_interface(audio_file, translate, model_size, question):
    _, full_text = transcribe_audio(audio_file, translate, model_size)
    answer = answer_question(full_text, question)
    return answer

# Main interface
with gr.Blocks() as iface:
    gr.Markdown("# WhisperX Audio Transcription, Translation, Summarization, and QA (with ZeroGPU support)")
    
    with gr.Tab("Transcribe and Summarize"):
        audio_input = gr.Audio(type="filepath")
        translate_checkbox = gr.Checkbox(label="Enable Translation")
        model_dropdown = gr.Dropdown(choices=["tiny", "base", "small", "medium", "large", "large-v2", "large-v3"], label="Whisper Model Size", value="small")
        transcribe_button = gr.Button("Transcribe and Summarize")
        transcription_output = gr.Textbox(label="Transcription")
        summary_output = gr.Textbox(label="Summary")
        
        transcribe_button.click(
            process_and_summarize,
            inputs=[audio_input, translate_checkbox, model_dropdown],
            outputs=[transcription_output, summary_output]
        )
    
    with gr.Tab("Question Answering"):
        qa_audio_input = gr.Audio(type="filepath")
        qa_translate_checkbox = gr.Checkbox(label="Enable Translation")
        qa_model_dropdown = gr.Dropdown(choices=["tiny", "base", "small", "medium", "large", "large-v2", "large-v3"], label="Whisper Model Size", value="small")
        question_input = gr.Textbox(label="Ask a question about the audio")
        qa_button = gr.Button("Get Answer")
        answer_output = gr.Textbox(label="Answer")
        
        qa_button.click(
            qa_interface,
            inputs=[qa_audio_input, qa_translate_checkbox, qa_model_dropdown, question_input],
            outputs=answer_output
        )

    gr.Markdown(
        """
        ## ZeroGPU Support
        This application supports ZeroGPU for Hugging Face Spaces pro users. 
        GPU-intensive tasks are automatically optimized for better performance.
        """
    )

iface.launch()