app.py

import gradio as gr
from AinaTheme import theme
from faster_whisper import WhisperModel
import torch

device, torch_dtype = ("cuda", "float32") if torch.cuda.is_available() else ("cpu", "int8")

MODEL_NAME = "suzii/vi-whisper-large-v3-turbo-v1-ct2"
print("Loading model ...")
model = WhisperModel(MODEL_NAME, compute_type=torch_dtype)
print("Loading model done.")

def transcribe(inputs):
    print("transcribe()")
    if inputs is None:
        raise gr.Error("Cap fitxer d'àudio introduit! Si us plau pengeu un fitxer "\
                       "o enregistreu un àudio abans d'enviar la vostra sol·licitud")

    segments, _ = model.transcribe(
        inputs, 
        chunk_length=30,
        task="transcribe",
        word_timestamps=True,
        repetition_penalty=1.1,
        temperature=[0.0, 0.1, 0.2, 0,3, 0.4, 0.6, 0.8, 1.0],
    )

    text = ""
    for segment in segments:
        text += " " + segment.text.strip()
    return text





def clear():
     return (None)


with gr.Blocks(theme=theme) as demo:
    gr.Markdown("CPU - type: int8")
    with gr.Row():
        with gr.Column(scale=1):
            input = gr.Audio(sources=["upload", "microphone"], type="filepath", label="Audio")

        with gr.Column(scale=1):
            output = gr.Textbox(label="Output", lines=8)
    
    with gr.Row(variant="panel"):
            clear_btn = gr.Button("Clear")
            submit_btn = gr.Button("Submit", variant="primary")


    submit_btn.click(fn=transcribe, inputs=[input], outputs=[output])
    clear_btn.click(fn=clear,inputs=[], outputs=[input], queue=False,)


if __name__ == "__main__":
    demo.launch()