import gradio as gr import edge_tts import asyncio import tempfile import os from huggingface_hub import InferenceClient import re from streaming_stt_nemo import Model import torch import random default_lang = "en" engines = { default_lang: Model(default_lang) } def transcribe(audio): client5=InferenceClient("openai/whisper-medium") text=client5.automatic_speech_recognition(audio).text return text HF_TOKEN = os.environ.get("HF_TOKEN", None) def client_fn(model): if "Mixtral" in model: return InferenceClient("mistralai/Mixtral-8x7B-Instruct-v0.1") elif "Llama" in model: return InferenceClient("meta-llama/Meta-Llama-3-8B-Instruct") elif "Mistral" in model: return InferenceClient("mistralai/Mistral-7B-Instruct-v0.3") elif "Phi" in model: return InferenceClient("microsoft/Phi-3-mini-4k-instruct") else: return InferenceClient("microsoft/Phi-3-mini-4k-instruct") def randomize_seed_fn(seed: int) -> int: seed = random.randint(0, 999999) return seed system_instructions1 = "[SYSTEM] Answer as Real Jarvis JARVIS, Made by 'Tony Stark', Keep conversation very short, clear, friendly and concise. The text provided is a request for a specific type of response from you, the virtual assistant. The request asks you to provide friendly responses as if You are the character Jarvis, made by 'Tony Stark.' The expectation is that I will avoid introductions and start answering the query directly, Only answer the question asked by user, Do not say unnecessary things.[USER]" def models(text, model="Mixtral 8x7B", seed=42): seed = int(randomize_seed_fn(seed)) generator = torch.Generator().manual_seed(seed) client = client_fn(model) generate_kwargs = dict( max_new_tokens=300, seed=seed ) formatted_prompt = system_instructions1 + text + "[JARVIS]" stream = client.text_generation( formatted_prompt, **generate_kwargs, stream=True, details=True, return_full_text=False) output = "" for response in stream: if not response.token.text == "": output += response.token.text return output async def respond(audio, model, seed): user = transcribe(audio) reply = models(user, model, seed) communicate = edge_tts.Communicate(reply) with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as tmp_file: tmp_path = tmp_file.name await communicate.save(tmp_path) yield tmp_path DESCRIPTION = """ #
JARVISāš”
###
A personal Assistant of Tony Stark for YOU ###
Voice Chat with your personal Assistant
""" with gr.Blocks(css="style.css") as demo: gr.Markdown(DESCRIPTION) with gr.Row(): select = gr.Dropdown([ 'Mixtral 8x7B', 'Llama 3 8B', 'Mistral 7B v0.3', 'Phi 3 mini', ], value="Mistral 7B v0.3", label="Model" ) seed = gr.Slider( label="Seed", minimum=0, maximum=999999, step=1, value=0, visible=False ) input = gr.Audio(label="User", sources="microphone", type="filepath", waveform_options=False) output = gr.Audio(label="AI", type="filepath", interactive=False, autoplay=True, elem_classes="audio") gr.Interface( batch=True, max_batch_size=10, fn=respond, inputs=[input, select, seed], outputs=[output], live=True) with gr.Row(): select = gr.Dropdown([ 'Mixtral 8x7B', 'Llama 3 8B', 'Mistral 7B v0.3', 'Phi 3 mini', ], value="Mistral 7B v0.3", label="Model" ) seed = gr.Slider( label="Seed", minimum=0, maximum=999999, step=1, value=0, visible=False ) input = gr.Textbox(label="User") output = gr.Textbox(label="AI", interactive=False) gr.Interface( batch=True, max_batch_size=10, fn=models, inputs=[input, select, seed], outputs=[output], live=True) if __name__ == "__main__": demo.queue(max_size=200).launch()