app.py

import cohere
import gradio as gr
import os, json
from loguru import logger
import random
from transformers import pipeline
import torch

co = cohere.Client(os.environ.get('COHERE_API_KEY'))

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

whisper_model = pipeline(
    task="automatic-speech-recognition",
    model="openai/whisper-large-v2",
    chunk_length_s=30,
    device=device,
)

all_special_ids = whisper_model.tokenizer.all_special_ids
transcribe_token_id = all_special_ids[-5]
translate_token_id = all_special_ids[-6]

def translate_or_transcribe(audio, task):
    whisper_model.model.config.forced_decoder_ids = [[2, transcribe_token_id if task=="Transcribe in Spoken Language" else translate_token_id]]
    text = whisper_model(audio)["text"]
    return text
    
def get_response_from_chatbot(text):
    try:
        resp = api.send_message(text)    
        response = resp['message']
      # logger.info(f"response_: {response}")
    except:
        response = "Sorry, the chatGPT queue is full. Please try again in some time"
    return response

def chat(message, chat_history):      
    out_chat = []
    if chat_history != '':
        out_chat = json.loads(chat_history)
    response = get_response_from_chatbot(message)
    out_chat.append((message, response))
    chat_history = json.dumps(out_chat)
    logger.info(f"out_chat_: {len(out_chat)}")
    return out_chat, chat_history

start_work = """async() => {
    function isMobile() {
        try {
            document.createEvent("TouchEvent"); return true;
        } catch(e) {
            return false; 
        }
    }
	function getClientHeight()
	{
	  var clientHeight=0;
	  if(document.body.clientHeight&&document.documentElement.clientHeight) {
		var clientHeight = (document.body.clientHeight<document.documentElement.clientHeight)?document.body.clientHeight:document.documentElement.clientHeight;
	  } else {
		var clientHeight = (document.body.clientHeight>document.documentElement.clientHeight)?document.body.clientHeight:document.documentElement.clientHeight;
	  }
	  return clientHeight;
	}
 
    function setNativeValue(element, value) {
      const valueSetter = Object.getOwnPropertyDescriptor(element.__proto__, 'value').set;
      const prototype = Object.getPrototypeOf(element);
      const prototypeValueSetter = Object.getOwnPropertyDescriptor(prototype, 'value').set;
      
      if (valueSetter && valueSetter !== prototypeValueSetter) {
            prototypeValueSetter.call(element, value);
      } else {
            valueSetter.call(element, value);
      }
    }
    var gradioEl = document.querySelector('body > gradio-app').shadowRoot;
    if (!gradioEl) {
        gradioEl = document.querySelector('body > gradio-app');
    }
    
    if (typeof window['gradioEl'] === 'undefined') {
        window['gradioEl'] = gradioEl;
       
        const page1 = window['gradioEl'].querySelectorAll('#page_1')[0];
        const page2 = window['gradioEl'].querySelectorAll('#page_2')[0]; 
    
        page1.style.display = "none";
        page2.style.display = "block"; 

        window['div_count'] = 0;
        window['chat_bot'] = window['gradioEl'].querySelectorAll('#chat_bot')[0];
        window['chat_bot1'] = window['gradioEl'].querySelectorAll('#chat_bot1')[0];   
        chat_row = window['gradioEl'].querySelectorAll('#chat_row')[0]; 
        prompt_row = window['gradioEl'].querySelectorAll('#prompt_row')[0]; 
        window['chat_bot1'].children[1].textContent = '';
        
        clientHeight = getClientHeight();
        new_height = (clientHeight-300) + 'px';
        chat_row.style.height = new_height;
        window['chat_bot'].style.height = new_height;
        window['chat_bot'].children[2].style.height = new_height;
        window['chat_bot1'].style.height = new_height;
        window['chat_bot1'].children[2].style.height = new_height;
        prompt_row.children[0].style.flex = 'auto';
        prompt_row.children[0].style.width = '100%';
        
        window['checkChange'] = function checkChange() {
            try {
                if (window['chat_bot'].children[2].children[0].children.length > window['div_count']) {
                    new_len = window['chat_bot'].children[2].children[0].children.length - window['div_count'];
                    for (var i = 0; i < new_len; i++) { 
                        new_div = window['chat_bot'].children[2].children[0].children[window['div_count'] + i].cloneNode(true);
                        window['chat_bot1'].children[2].children[0].appendChild(new_div);
                    }
                    window['div_count'] = chat_bot.children[2].children[0].children.length;
                }
                if (window['chat_bot'].children[0].children.length > 1) {
                     window['chat_bot1'].children[1].textContent = window['chat_bot'].children[0].children[1].textContent;
                } else {
                    window['chat_bot1'].children[1].textContent = '';
                }
              
            } catch(e) {
            }        
        }
        window['checkChange_interval'] = window.setInterval("window.checkChange()", 500);         
    }
   
    return false;
}"""


with gr.Blocks(title='Talk to catGPT') as demo:
    gr.Markdown("## Talk to catGPT with your voice in your native language ! ##")
    with gr.Group(elem_id="page_1", visible=True) as page_1:
        with gr.Box():            
            with gr.Row():
                start_button = gr.Button("Let's talk to chatGPT!", elem_id="start-btn", visible=True) 
                start_button.click(fn=None, inputs=[], outputs=[], _js=start_work)
                
    with gr.Group(elem_id="page_2", visible=False) as page_2:        
        with gr.Row(elem_id="chat_row"):
            chatbot = gr.Chatbot(elem_id="chat_bot", visible=False).style(color_map=("green", "blue"))
            chatbot1 = gr.Chatbot(elem_id="chat_bot1").style(color_map=("green", "blue"))    
        with gr.Row():
            prompt_input_audio = gr.Audio(
                                        source="microphone",
                                        type="filepath",
                                        label="Record Audio Input",
                                        
                                    )
            translate_btn = gr.Button("Check Whisper first ? 👍")
            
        whisper_task = gr.Radio(["Translate to English", "Transcribe in Spoken Language"], value="Translate to English", show_label=False)
        with gr.Row(elem_id="prompt_row"):
            prompt_input = gr.Textbox(lines=2, label="Input text",show_label=True)
            chat_history = gr.Textbox(lines=4, label="prompt", visible=False)
            submit_btn = gr.Button(value = "Send to catgpt",elem_id="submit-btn").style(
                    margin=True,
                    rounded=(True, True, True, True),
                    width=100
                )
            
             

        translate_btn.click(fn=translate_or_transcribe, 
                         inputs=[prompt_input_audio,whisper_task], 
                         outputs=prompt_input
                        )
        
          
        submit_btn.click(fn=chat, 
                             inputs=[prompt_input, chat_history], 
                             outputs=[chatbot, chat_history],
                            )
    gr.HTML('''
         <p>Note: Please be aware that audio records from iOS devices will not be decoded as expected by Gradio. For the best experience, record your voice from a computer instead of your smartphone ;)</p>
        <div class="footer">
                    <p>Whisper Model by <a href="https://github.com/openai/whisper" style="text-decoration: underline;" target="_blank">OpenAI</a> -   
                    <a href="https://chat.openai.com/chat" target="_blank">chatGPT</a> by <a href="https://openai.com/" style="text-decoration: underline;" target="_blank">OpenAI</a> 
                    </p>
        </div>
        ''')
    #gr.Markdown("![visitor badge](https://visitor-badge.glitch.me/badge?page_id=RamAnanth1.chatGPT_voice)")

demo.launch(debug = True)