app.py

import gradio as gr
import torch
from transformers import pipeline
import numpy as np
import time

pipe_base = pipeline("automatic-speech-recognition", model="aitor-medrano/lara-base-pushed")
pipe_small = pipeline("automatic-speech-recognition", model="aitor-medrano/whisper-small-lara")
#pipe_base_2000 = pipeline("automatic-speech-recognition", model="aitor-medrano/whisper-base-lara-2000")

def greet(grabacion):
    inicio = time.time()

    sr, y = grabacion
    # Pasamos el array de muestras a tipo NumPy de 32 bits
    y = y.astype(np.float32)
    y /= np.max(np.abs(y))

    result_base = "base:" + pipe_base({"sampling_rate": sr, "raw": y})["text"]
    result_small = "small:" + pipe_small({"sampling_rate": sr, "raw": y})["text"]
    #result_base_2000 = "base_2000:" + pipe_base_2000({"sampling_rate": sr, "raw": y})["text"]

    fin = time.time()

    #return result_base, result_small, result_base_2000, fin - inicio
    return result_base, result_small, fin - inicio

demo = gr.Interface(fn=greet,
        inputs=[
                gr.Audio(),
        ],
        outputs=[
            gr.Text(label="Salida (Base)"),
            gr.Text(label="Salida (Small)"),
            #gr.Text(label="Salida (Base 2000)"),
            gr.Number(label="Tiempo")
        ])
demo.launch()