app.py

from transformers import TTSForConditionalGeneration, AutoModelForSeq2SeqLM, AutoModelForCausalLM
from datasets import load_dataset
from gradio import Interface


tts_model = TTSForConditionalGeneration.from_pretrained("facebook/s2s-wav2vec2-ft")
whisper_model = AutoModelForSeq2SeqLM.from_pretrained("openai/whisper-medium")
gpt2_xl_model = AutoModelForCausalLM.from_pretrained("openai-community/gpt2-xl")


def voice_chat(user_voice):
    messages = [{"role": "system", "content": "You are a kind helpful assistant."}]
    user_message = whisper_model.generate(user_voice)["generated_text"]
    messages.append({"role": "user", "content": user_message})
    print(messages)

    # Utilize gpt2-xl locally for chat completion
    with gpt2_xl_model.no_grad():
        input_ids = gpt2_xl_model.tokenizer(messages, return_tensors="pt").input_ids
        outputs = gpt2_xl_model.generate(input_ids, max_length=100)
        chat_response = gpt2_xl_model.tokenizer.decode(outputs[0], skip_special_tokens=True)

    reply = chat_response
    messages.append({"role": "assistant", "content": reply})
    audio = tts_model.generate(text=reply, return_tensors="pt").tolist()[0][0]
    return reply, audio


text_reply = Interface.Textbox(label="ChatGPT Text")
voice_reply = Interface.Audio(type="audio/wav")

Interface(
    title="AI Voice Assistant with ChatGPT AI",
    fn=voice_chat,
    inputs=[Interface.Audio(source="microphone", type="filepath")],
    outputs=[text_reply, voice_reply],
    live=True,
).launch(debug=True)