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
from pyannote.audio import Pipeline

# Suppress warnings
warnings.filterwarnings("ignore")
logging.set_verbosity_error()

# Updated models by language
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-large-xlsr-53-portuguese"
    ]
}

def convert_audio_to_wav(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)
    return wav_path

def detect_language(audio_path):
    speech, _ = librosa.load(audio_path, sr=16000, duration=30)
    
    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 diarize_audio(wav_audio):
    pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization")
    diarization = pipeline(wav_audio)
    return diarization

def transcribe_audio_stream(audio, model_name, diarization):
    wav_audio = convert_audio_to_wav(audio)
    speech, rate = librosa.load(wav_audio, sr=16000)
    duration = len(speech) / rate
    
    if "whisper" in model_name:
        processor = WhisperProcessor.from_pretrained(model_name)
        model = WhisperForConditionalGeneration.from_pretrained(model_name)
        
        chunk_duration = 30  # seconds
        
        transcriptions = []
        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)
            timestamp = i
            transcriptions.append((timestamp, transcription))
            yield transcriptions, progress
    else:
        transcriber = pipeline("automatic-speech-recognition", model=model_name)
        
        chunk_duration = 10  # seconds
        
        transcriptions = []
        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)
            timestamp = i
            transcriptions.append((timestamp, result["text"]))
            yield transcriptions, progress

    # Merge diarization results with transcription
    speaker_transcriptions = []
    for segment in diarization.itertracks(yield_label=True):
        start, end, speaker = segment
        start_time = start / rate
        end_time = end / rate
        text_segment = ""
        for ts, text in transcriptions:
            if start_time <= ts <= end_time:
                text_segment += text + " "
        speaker_transcriptions.append((start_time, end_time, speaker, text_segment.strip()))

    return speaker_transcriptions

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 == "txt":
        with open("transcription.txt", "w") as f:
            for start, end, speaker, text in transcriptions:
                f.write(f"[{start}-{end}] {speaker}: {text}\n")
        return "transcription.txt"
    elif file_format == "json":
        with open("transcription.json", "w") as f:
            json.dump(transcriptions, f)
        return "transcription.json"

def combined_interface(audio):
    try:
        language, model_options = detect_and_select_model(audio)
        selected_model = model_options[0]
        
        yield language, model_options, selected_model, [], 0, "Initializing..."
        
        wav_audio = convert_audio_to_wav(audio)
        diarization = diarize_audio(wav_audio)
        transcriptions = []
        for partial_transcriptions, progress in transcribe_audio_stream(audio, selected_model, diarization):
            transcriptions = partial_transcriptions
            transcriptions_text = "\n".join([f"[{start}-{end}] {speaker}: {text}" for start, end, speaker, text in transcriptions])
            progress_int = math.floor(progress)
            status = f"Transcribing... {progress_int}% complete"
            yield language, model_options, selected_model, transcriptions_text, progress_int, status
        
        # Clean up temporary files
        os.remove("converted_audio.wav")
        
        yield language, model_options, selected_model, transcriptions_text, 100, "Transcription complete!"
        
    except Exception as e:
        yield str(e), [], "", "An error occurred during processing.", 0, "Error"

iface = gr.Interface(
    fn=combined_interface,
    inputs=gr.Audio(type="filepath"),
    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 (TXT)", type="filepath", interactive=True, value="transcription.txt"),
        gr.File(label="Download Transcription (JSON)", type="filepath", interactive=True, value="transcription.json")
    ],
    title="Multilingual Audio Transcriber with Real-time Display, Timestamps, and Speaker Diarization",
    description="Upload an audio file to detect the language, select the transcription model, and get the transcription with timestamps and speaker labels in real-time. Download the transcription as TXT or JSON. Optimized for Spanish, English, and Portuguese.",
    live=True
)

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