app.py

import gradio as gr
import numpy as np
import torch
from peft import PeftModel, PeftConfig
from transformers import WhisperForConditionalGeneration, WhisperTokenizer, WhisperProcessor, AutomaticSpeechRecognitionPipeline

peft_model_id = "mfidabel/Modelo_1_Whisper_Large_V3"
language = "guarani"
task = "transcribe"
peft_config = PeftConfig.from_pretrained(peft_model_id)
model = WhisperForConditionalGeneration.from_pretrained(
    peft_config.base_model_name_or_path, load_in_8bit=True, device_map="auto"
)
model = PeftModel.from_pretrained(model, peft_model_id)
tokenizer = WhisperTokenizer.from_pretrained(peft_config.base_model_name_or_path, language=language, task=task)
processor = WhisperProcessor.from_pretrained(peft_config.base_model_name_or_path, language=language, task=task)
feature_extractor = processor.feature_extractor
forced_decoder_ids = processor.get_decoder_prompt_ids(language="english", task=task)

pipeline = AutomaticSpeechRecognitionPipeline(model=model, tokenizer=tokenizer, feature_extractor=feature_extractor)

def transcribe(audio):
    if audio is None:
        return "Espera a que la grabación termine de subirse al servidor !! Intentelo de nuevo en unos segundos"
        
    sr, y = audio
    y = y.astype(np.float32)
    y /= np.max(np.abs(y))
    with torch.cuda.amp.autocast():
        return pipeline({"sampling_rate": sr, "raw": y}, generate_kwargs={"forced_decoder_ids": forced_decoder_ids}, max_new_tokens=255)["text"]

gr.Interface(fn=transcribe, inputs="microphone", outputs="text").launch(share=True)