app.py

import gradio as gr
from transformers import pipeline, WhisperProcessor, WhisperForConditionalGeneration
import torch
import librosa
import subprocess
from langdetect import detect_langs
import os
import warnings
from transformers import logging
import math
import json

# Suprimir advertencias
warnings.filterwarnings("ignore")
logging.set_verbosity_error()

# Modelos actualizados por idioma
MODELS = {
    "es": [
        "openai/whisper-large-v3",
        "facebook/wav2vec2-large-xlsr-53-spanish",
        "jonatasgrosman/wav2vec2-xls-r-1b-spanish"
    ],
    "en": [
        "openai/whisper-large-v3",
        "facebook/wav2vec2-large-960h",
        "microsoft/wav2vec2-base-960h"
    ],
    "pt": [
        "facebook/wav2vec2-large-xlsr-53-portuguese",
        "openai/whisper-medium",
        "jonatasgrosman/wav2vec2-xlsr-53-portuguese"
    ]
}

def convert_audio_to_wav(audio_path):
    if os.path.isdir(audio_path):
        raise ValueError(f"La ruta proporcionada es un directorio, no un archivo: {audio_path}")
    wav_path = "converted_audio.wav"
    command = ["ffmpeg", "-i", audio_path, "-ac", "1", "-ar", "16000", wav_path]
    subprocess.run(command, stdout=subprocess.PIPE, stderr=subprocess.PIPE, check=True)
    return wav_path

def detect_language(audio_path):
    try:
        speech, _ = librosa.load(audio_path, sr=16000, duration=30)
    except Exception as e:
        raise ValueError(f"Error al cargar el archivo de audio con librosa: {e}")
    
    processor = WhisperProcessor.from_pretrained("openai/whisper-base")
    model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-base")
    
    input_features = processor(speech, sampling_rate=16000, return_tensors="pt").input_features
    predicted_ids = model.generate(input_features)
    transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
    
    langs = detect_langs(transcription)
    
    es_confidence = next((lang.prob for lang in langs if lang.lang == 'es'), 0)
    pt_confidence = next((lang.prob for lang in langs if lang.lang == 'pt'), 0)
    
    if abs(es_confidence - pt_confidence) < 0.2:
        return 'es'
    
    return max(langs, key=lambda x: x.prob).lang

def transcribe_audio_stream(audio, model_name):
    wav_audio = convert_audio_to_wav(audio)
    speech, rate = librosa.load(wav_audio, sr=16000)
    duration = len(speech) / rate
    
    transcriptions = []

    if "whisper" in model_name:
        processor = WhisperProcessor.from_pretrained(model_name)
        model = WhisperForConditionalGeneration.from_pretrained(model_name)
        
        chunk_duration = 30  # segundos
        
        for i in range(0, int(duration), chunk_duration):
            end = min(i + chunk_duration, duration)
            chunk = speech[int(i * rate):int(end * rate)]
            
            input_features = processor(chunk, sampling_rate=16000, return_tensors="pt").input_features
            predicted_ids = model.generate(input_features)
            transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
            
            progress = min(100, (end / duration) * 100)
            transcriptions.append({
                "start_time": i,
                "end_time": end,
                "text": transcription
            })
            yield transcriptions, progress
    else:
        transcriber = pipeline("automatic-speech-recognition", model=model_name)
        
        chunk_duration = 10  # segundos
        
        for i in range(0, int(duration), chunk_duration):
            end = min(i + chunk_duration, duration)
            chunk = speech[int(i * rate):int(end * rate)]
            result = transcriber(chunk)
            
            progress = min(100, (end / duration) * 100)
            transcriptions.append({
                "start_time": i,
                "end_time": end,
                "text": result["text"]
            })
            yield transcriptions, progress

def detect_and_select_model(audio):
    wav_audio = convert_audio_to_wav(audio)
    language = detect_language(wav_audio)
    model_options = MODELS.get(language, MODELS["en"])
    return language, model_options

def save_transcription(transcriptions, file_format):
    if file_format == "JSON":
        file_path = "transcription.json"
        with open(file_path, 'w') as f:
            json.dump(transcriptions, f, ensure_ascii=False, indent=4)
    elif file_format == "TXT":
        file_path = "transcription.txt"
        with open(file_path, 'w') as f:
            for entry in transcriptions:
                f.write(f"{entry['start_time']:.2f},{entry['end_time']:.2f},{entry['text']}\n")
    return file_path

def combined_interface(audio, file_format):
    try:
        language, model_options = detect_and_select_model(audio)
        selected_model = model_options[0]
        
        # Primer yield: Añadir None para la séptima salida (Archivo de Descarga)
        yield language, model_options, selected_model, "", 0, "Initializing...", None
        
        transcriptions = []
        for partial_transcriptions, progress in transcribe_audio_stream(audio, selected_model):
            transcriptions = partial_transcriptions
            full_transcription = " ".join([t["text"] for t in transcriptions])
            progress_int = math.floor(progress)
            status = f"Transcribing... {progress_int}% complete"
            # Yield con None para el archivo de descarga hasta que esté completo
            yield language, model_options, selected_model, full_transcription.strip(), progress_int, status, None
        
        # Guardar transcripción
        file_path = save_transcription(transcriptions, file_format)
        
        # Limpiar archivos temporales
        os.remove("converted_audio.wav")
        
        # Yield final con el archivo de descarga
        yield language, model_options, selected_model, full_transcription.strip(), 100, f"Transcription complete! Download {file_path}", file_path
        
    except Exception as e:
        # Asegurarse de que el yield de error también devuelva 7 valores
        yield str(e), [], "", "An error occurred during processing.", 0, "Error", ""

iface = gr.Interface(
    fn=combined_interface,
    inputs=[
        gr.Audio(type="filepath", label="Upload Audio File"),
        gr.Radio(choices=["JSON", "TXT"], label="Choose output format")
    ],
    outputs=[
        gr.Textbox(label="Detected Language"),
        gr.Dropdown(label="Available Models", choices=[]),
        gr.Textbox(label="Selected Model"),
        gr.Textbox(label="Transcription", lines=10),
        gr.Slider(minimum=0, maximum=100, label="Progress", interactive=False),
        gr.Textbox(label="Status"),
        gr.File(label="Download Transcription")
    ],
    title="Multilingual Audio Transcriber with Real-time Display and Progress Indicator",
    description="Upload an audio file to detect the language, select the transcription model, and get the transcription in real-time. Optimized for Spanish, English, and Portuguese.",
    live=True
)

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