app.py

import torch
import json
import os
from transformers import pipeline, VitsModel, VitsTokenizer, SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor
import numpy as np
os.system("pip install git+https://github.com/openai/whisper.git")
import gradio as gr
import whisper
import requests

MODEL = "gpt-3.5-turbo"
API_URL = os.getenv("API_URL")
OPENAI_API_KEY = os.getenv("OPENAI_API_KEY")
NUM_THREADS = 2

model = whisper.load_model("small")
device = "cuda:0" if torch.cuda.is_available() else "cpu"


def parse_codeblock(text):
    lines = text.split("\n")
    for i, line in enumerate(lines):
        if "```" in line:
            if line != "```":
                lines[i] = f'<pre><code class="{lines[i][3:]}">'
            else:
                lines[i] = '</code></pre>'
        else:
            if i > 0:
                lines[i] = "<br/>" + line.replace("<", "&lt;").replace(">", "&gt;")
    return "".join(lines)
    
def inference(audio):
    audio = whisper.load_audio(audio)
    print("loading finished")
    audio = whisper.pad_or_trim(audio)
    print("audio trimed")
    mel = whisper.log_mel_spectrogram(audio).to(model.device)
    print("spectro finished")
    _, probs = model.detect_language(mel)
    print("lang detected")
    options = whisper.DecodingOptions(fp16 = False)

    result = whisper.decode(model, mel, options)
    
    print(result.text)
    return result.text

    
#Load Whisper-small
pipe = pipeline("automatic-speech-recognition",
                model="openai/whisper-small",
                device=device
)
#pipe = pipeline(model="Sleepyp00/whisper-small-Swedish") 

model2 = VitsModel.from_pretrained("facebook/mms-tts-eng")
tokenizer = VitsTokenizer.from_pretrained("facebook/mms-tts-eng")


# Define a function to translate an audio, in english here
def translate(audio):
    # return inference(audio)
    outputs = pipe(audio, max_new_tokens=256,
                   generate_kwargs={"task": "translate"})
    return outputs["text"]


# Define function to generate the waveform output
def synthesise(text):
    inputs = tokenizer(text, return_tensors="pt")
    input_ids = inputs["input_ids"]

    with torch.no_grad(): 
      outputs = model2(input_ids)
    
    return outputs.audio[0]


def gpt_predict(inputs, request:gr.Request=gr.State([]), top_p = 1, temperature = 1, chat_counter = 0,history =[]):
    payload = {
        "model": MODEL,
        "messages": [{"role": "user", "content": f"{inputs}"}],
        "temperature" : 1.0,
        "top_p":1.0,
        "n" : 1,
        "stream": True,
        "presence_penalty":0,
        "frequency_penalty":0,
    }

    headers = {
        "Content-Type": "application/json",
        "Authorization": f"Bearer {OPENAI_API_KEY}",
    }

    # print(f"chat_counter - {chat_counter}")
    if chat_counter != 0 :
        messages = []
        for i, data in enumerate(history):
            if i % 2 == 0:
                role = 'user'
            else:
                role = 'assistant'
            message = {}
            message["role"] = role
            message["content"] = data
            messages.append(message)
        
        message = {}
        message["role"] = "user" 
        message["content"] = inputs
        messages.append(message)
        payload = {
            "model": MODEL,
            "messages": messages,
            "temperature" : temperature,
            "top_p": top_p,
            "n" : 1,
            "stream": True,
            "presence_penalty":0,
            "frequency_penalty":0,
        }

    chat_counter += 1

    history.append(inputs)
    token_counter = 0 
    partial_words = "" 
    counter = 0

    try:
        # make a POST request to the API endpoint using the requests.post method, passing in stream=True
        print('in the try')
        response = requests.post(API_URL, headers=headers, json=payload, stream=True)
        response_code = f"{response}"
        print(response_code)
        #if response_code.strip() != "<Response [200]>":
        #    #print(f"response code - {response}")
        #    raise Exception(f"Sorry, hitting rate limit. Please try again later. {response}")
        out = []
        for chunk in response.iter_lines():
            #Skipping first chunk
            if counter == 0:
                counter += 1
                continue
                #counter+=1
            # check whether each line is non-empty
            if chunk.decode() :
                chunk = chunk.decode()
                # decode each line as response data is in bytes
                if len(chunk) > 12 and "content" in json.loads(chunk[6:])['choices'][0]['delta']:
                    partial_words = partial_words + json.loads(chunk[6:])['choices'][0]["delta"]["content"]
                    print(partial_words)
                    print(response)
                    if token_counter == 0:
                        history.append(" " + partial_words)
                    else:
                        history[-1] = partial_words
                    token_counter += 1
                    out.append([(parse_codeblock(history[i]), parse_codeblock(history[i + 1])) for i in range(0, len(history) - 1, 2) ], history, chat_counter, response, gr.update(interactive=False), gr.update(interactive=False)  # resembles {chatbot: chat, state: history}  
    except Exception as e:
        print (f'error found: {e}')
    print(json.dumps({"chat_counter": chat_counter, "payload": payload, "partial_words": partial_words, "token_counter": token_counter, "counter": counter}))
    return response["choices"]["message"]["content"]
    return [(parse_codeblock(history[i]), parse_codeblock(history[i + 1])) for i in range(0, len(history) - 1, 2) ], history, chat_counter, response, gr.update(interactive=True), gr.update(interactive=True)


# Define the pipeline
def speech_to_speech_translation(audio):
    translated_text = translate(audio)
    synthesised_speech = synthesise(translated_text)
    synthesised_speech = (
        synthesised_speech.numpy() * 32767).astype(np.int16)
    return [translated_text, (16000, synthesised_speech)]

def predict(transType, language, audio, audio_mic = None):
        print("debug1:", audio,"debug2", audio_mic)
        if not audio and audio_mic:
            audio = audio_mic
        st = gr.State([])
        return gpt_predict("What time is it?",st)
        if transType == "Text":
            return translate(audio), None
        if transType == "GPT answer":
            req = translate(audio)
            return gpt_predict(req)
        if transType == "Audio":
            return speech_to_speech_translation(audio)
            
# Define the title etc
title = "Swedish STSOT (Speech To Speech Or Text)"
description="Use Whisper pretrained model to convert swedish audio to english (text or audio)"


supportLangs = ["Swedish", "French (in training)"]
transTypes = ["Text", "Audio", "GPT answer"]

#examples = [
#    ["Text", "Swedish", "./ex1.wav", None],
#    ["Audio", "Swedish", "./ex2.wav", None]
#]

examples =[]
demo = gr.Interface(
    fn=predict,
    inputs=[
        gr.Radio(label="Choose your output format", choices=transTypes),
        gr.Radio(label="Choose a source language", choices=supportLangs, value="Swedish"),
        gr.Audio(label="Import an audio", sources="upload", type="filepath"),
        #gr.Audio(label="Import an audio", sources="upload", type="numpy"),
        gr.Audio(label="Record an audio", sources="microphone", type="filepath"),
        
    ],
    outputs=[
        gr.Text(label="Text translation or gpt answer"),gr.Audio(label="Audio translation",type = "numpy")
    ],
    title=title,
    description=description,
    article="",
    examples=examples,
)


demo.launch()