app.py

import gradio as gr
import torch
import numpy as np
from transformers import Wav2Vec2Processor, Wav2Vec2ForSequenceClassification

# 
model_name = "results"
processor = Wav2Vec2Processor.from_pretrained(model_name)
model = Wav2Vec2ForSequenceClassification.from_pretrained(model_name, from_tf=False, from_flax=False, from_safetensors=True)

def classify_accent(audio):
    if audio is None:
        return "Error: No se recibió audio"

    # Entrada
    print(f"Tipo de entrada de audio: {type(audio)}")

    # O áudio formato
    print(f"Entrada de audio recibida: {audio}")

    try:
        audio_array = audio[1]  # 
        sample_rate = audio[0]  # 

        print(f"Forma del audio: {audio_array.shape}, Frecuencia de muestreo: {sample_rate}")

        # 
        audio_array = audio_array.astype(np.float32)

        # 
        if sample_rate != 16000:
            import librosa
            audio_array = librosa.resample(audio_array, orig_sr=sample_rate, target_sr=16000)

        input_values = processor(audio_array, return_tensors="pt", sampling_rate=16000).input_values

        # Infer
        with torch.no_grad():
            logits = model(input_values).logits
        predicted_ids = torch.argmax(logits, dim=-1).item()

        # 
        labels = ["Español", "Otro"]
        return labels[predicted_ids]

    except Exception as e:
        return f"Error al procesar el audio: {str(e)}"

# 
description_html = """
<p>Prueba con grabación o cargando un archivo de audio. Para probar, recomiendo una palabra.</p>
<p>Ramon Mayor Martins, Ph.D.: <a href="https://rmayormartins.github.io/" target="_blank">Website</a> | <a href="https://huggingface.co/rmayormartins" target="_blank">Spaces</a></p>
"""

# 
interface = gr.Interface(
    fn=classify_accent,
    inputs=gr.Audio(type="numpy"),
    outputs="label",
    title="Clasificador de Sotaques (Español vs Otro)",
    description=description_html
)

interface.launch()