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_1600 = pipeline("automatic-speech-recognition", model="aitor-medrano/whisper-base-lara-1600")

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))

    inicio_base = time.time()
    result_base = "base:" + pipe_base({"sampling_rate": sr, "raw": y})["text"]
    fin_base = time.time()

    inicio_small = time.time()
    result_small = "small:" + pipe_small({"sampling_rate": sr, "raw": y})["text"]
    fin_small = time.time()

    inicio_1600 = time.time()
    result_base_1600 = "base_2000:" + pipe_base_1600({"sampling_rate": sr, "raw": y})["text"]
    fin_1600 = time.time()

    fin = time.time()

    return result_base, fin_base - inicio_base, result_small, fin_small - inicio_small, result_base_1600, fin_1600 - inicio_1600, fin - inicio
    #return result_base, result_small, fin - inicio

demo = gr.Interface(fn=greet,
        inputs=[
                gr.Audio(),
        ],
        outputs=[
            gr.Text(label="Salida (Base)"),
            gr.Number(label="Tiempo (Base)"),
            gr.Text(label="Salida (Small)"),
            gr.Number(label="Tiempo (Small)"),
            gr.Text(label="Salida (Base 1600)"),
            gr.Number(label="Tiempo (1600)"),
            gr.Number(label="Tiempo")
        ])
demo.launch()