app.py

import gradio as gr
import random
import time
from ctransformers import AutoModelForCausalLM
from datetime import datetime
import whisper
from transformers import VitsModel, AutoTokenizer
import torch

params = {
        "max_new_tokens":512,
        "stop":["<end>" ,"<|endoftext|>","[", "<user>"],
        "temperature":0.7,
        "top_p":0.8,
        "stream":True,
        "batch_size": 8}


whisper_model = whisper.load_model("small")
llm = AutoModelForCausalLM.from_pretrained("Aspik101/trurl-2-7b-pl-instruct_GGML", model_type="llama")
tts_model = VitsModel.from_pretrained("facebook/mms-tts-pol")
tokenizer = AutoTokenizer.from_pretrained("facebook/mms-tts-pol")


with gr.Blocks() as demo:
    chatbot = gr.Chatbot()
    audio_input = gr.Audio(source="microphone", type="filepath", show_label=False)
    submit_audio = gr.Button("Submit Audio")
    clear = gr.Button("Clear")
    audio_output = gr.Audio('temp_file.wav', label="Generated Audio (wav)", type='filepath', autoplay=False)
    
    def translate(audio):
        print("__Wysyłam nagranie do whisper!")
        transcription = whisper_model.transcribe(audio, language="pl")
        return transcription["text"]
    
    def read_text(text):
        print("Tutaj jest tekst to przeczytania!", text[-1][-1])
        inputs = tokenizer(text[-1][-1], return_tensors="pt")
        with torch.no_grad():
            output = tts_model(**inputs).waveform.squeeze().numpy()
        sf.write('temp_file.wav', output, tts_model.config.sampling_rate)
        return 'temp_file.wav'
    
    def user(audio_data, history):
        if audio_data:
            user_message = translate(audio_data)
            print("USER!:")
            print("", history + [[user_message, None]])
            return history + [[user_message, None]]

    def parse_history(hist):
        history_ = ""
        for q, a in hist:
            history_ += f"<user>: {q } \n"
            if a:
                history_ += f"<assistant>: {a} \n"
        return history_

    def bot(history):
        print(f"When: {datetime.today().strftime('%Y-%m-%d %H:%M:%S')}")
        prompt = f"Jesteś AI assystentem. Odpowiadaj krótko i po polsku. {parse_history(history)}. <assistant>:"
        stream = llm(prompt, **params)
        history[-1][1] = ""
        answer_save = ""
        for character in stream:
            history[-1][1] += character
            answer_save += character
            time.sleep(0.005)
            yield history

    submit_audio.click(user, [audio_input, chatbot], [chatbot], queue=False).then(bot, chatbot, chatbot).then(read_text, chatbot, audio_output)
    clear.click(lambda: None, None, chatbot, queue=False)

demo.queue()
demo.launch(share=True)