app.py

import os
import gradio as gr
# import openai
from numpy._core.defchararray import endswith, isdecimal
from openai import OpenAI
from dotenv import load_dotenv
from pathlib import Path
from time import sleep
import audioread
import queue
import threading
from glob import glob
import copy

load_dotenv(override=True)
key = os.getenv('OPENAI_API_KEY')
users = os.getenv('LOGNAME')
unames = users.split(',')
pwds = os.getenv('PASSWORD')
pwdList = pwds.split(',')

site = os.getenv('SITE')
if site == 'local':
    dp = Path('./data')
    dp.mkdir(exist_ok=True)
    dataDir = './data/'
else:
    dp = Path('/data')
    dp.mkdir(exist_ok=True)
    dataDir = '/data/'

speak_file = dataDir + "speek.wav"

client = OpenAI(api_key = key)

#digits = ['zero: ','one: ','two: ','three: ','four: ','five: ','six: ','seven: ','eight: ','nine: ']

abbrevs = {'St. ' : 'Saint ', 'Mr. ': 'mister ', 'Mrs. ':'mussus ', 'Mr. ':'mister ', 'Ms. ':'mizz '}

def genUsageStats(do_reset=False):
    result = []
    ttotal4o_in = 0
    ttotal4o_out = 0
    ttotal4mini_in = 0
    ttotal4mini_out = 0
    totalAudio = 0
    totalSpeech = 0
    for user in unames:
        tokens4o_in = 0
        tokens4o_out = 0
        tokens4mini_in = 0
        tokens4mini_out = 0
        fp = dataDir + user + '_log.txt'
        if os.path.exists(fp):
            accessOk = False
            for i in range(3):
                try:
                    with open(fp) as f:
                        dataList = f.readlines()
                    if do_reset:
                        os.remove(fp)
                    else:
                        for line in dataList:
                            (u, t) = line.split(':')
                            (t, m) = t.split('-')
                            (tin, tout) = t.split('/')
                            incount = int(tin)
                            outcount = int(tout)
                            if 'mini' in m:
                                tokens4mini_in += incount
                                tokens4mini_out += outcount
                                ttotal4mini_in += incount
                                ttotal4mini_out += outcount
                            else:
                                tokens4o_in += incount
                                tokens4o_out += outcount
                                ttotal4o_in += incount
                                ttotal4o_out += outcount
                    accessOk = True
                    break
                except:
                    sleep(3)
            if not accessOk:
                return f'File access failed reading stats for user: {user}'
        userAudio = 0
        fp = dataDir + user + '_audio.txt'
        if os.path.exists(fp):
            accessOk = False
            for i in range(3):
                try:
                    with open(fp) as f:
                        dataList = f.readlines()
                    if do_reset:
                        os.remove(fp)
                    else:
                        for line in dataList:
                            (dud, len) = line.split(':')
                            userAudio += int(len)
                        totalAudio += int(userAudio)
                    accessOk = True
                    break
                except:
                    sleep(3)
            if not accessOk:
                return f'File access failed reading audio stats for user: {user}'
        userSpeech = 0
        fp = dataDir + user + '_speech.txt'
        if os.path.exists(fp):
            accessOk = False
            for i in range(3):
                try:
                    with open(fp) as f:
                        dataList = f.readlines()
                    if do_reset:
                        os.remove(fp)
                    else:
                        for line in dataList:
                            (dud, len) = line.split(':')
                            userSpeech += int(len)
                        totalSpeech += int(userSpeech)
                    accessOk = True
                    break
                except:
                    sleep(3)
            if not accessOk:
                return f'File access failed reading speech stats for user: {user}'
        result.append([user, f'{tokens4mini_in}/{tokens4mini_out}', f'{tokens4o_in}/{tokens4o_out}', f'audio:{userAudio}',f'speech:{userSpeech}'])
    result.append(['totals', f'{ttotal4mini_in}/{ttotal4mini_out}', f'{ttotal4o_in}/{ttotal4o_out}', f'audio:{totalAudio}',f'speech:{totalSpeech}'])
    return result       

def new_conversation(user):
    clean_up(user)
    return [None, [], None, []]

def updatePassword(txt):
    password = txt.lower().strip()
    return [password, "*********"]

# def setModel(val):
#     return val

def chat(prompt, user_window, pwd_window, past, response, gptModel):
    user_window = user_window.lower().strip()
    isBoss = False
    if user_window == unames[0] and pwd_window == pwdList[0]:
        isBoss = True
        if prompt == 'stats':
            response = genUsageStats()
            list_permanent_files()
            return [past, response, None, gptModel]
        if prompt == 'reset':
            response = genUsageStats(True)
            return [past, response, None, gptModel]
        if prompt.startswith('gpt4'):
            gptModel = 'gpt-4o'
            prompt = prompt[5:]
        if prompt.startswith("clean"):
            user = prompt[6:]
            response = f'cleaned all .wav files for {user}'
            final_clean_up(user)
            return [past, response, None, gptModel]
        if prompt.startswith('files'):
            (log_cnt, wav_cnt, other_cnt, others) = list_permanent_files()
            response = f'{log_cnt} log files\n{wav_cnt} .wav files\n{other_cnt} Other files:\n{others}'
            return [past, response, None, gptModel]
    if user_window in unames and pwd_window == pwdList[unames.index(user_window)]:
        past.append({"role":"user", "content":prompt})
        completion = client.chat.completions.create(model=gptModel,
                                            messages=past)
        reply = completion.choices[0].message.content
        tokens_in = completion.usage.prompt_tokens
        tokens_out = completion.usage.completion_tokens
        tokens = completion.usage.total_tokens
        response += "\n\nYOU: " + prompt + "\nGPT: " + reply
        if isBoss:
            response += f"\n{gptModel}: tokens in/out = {tokens_in}/{tokens_out}"
        if tokens > 40000:
            response += "\n\nTHIS DIALOG IS GETTING TOO LONG. PLEASE RESTART CONVERSATION SOON."
        past.append({"role":"assistant", "content": reply})
        accessOk = False
        for i in range(3):
            try:
                dataFile = new_func(user_window)
                with open(dataFile, 'a') as f:
                    m = '4o'
                    if 'mini' in gptModel:
                        m = '4omini'
                    f.write(f'{user_window}:{tokens_in}/{tokens_out}-{m}\n')
                accessOk = True
                break
            except Exception as e:
                sleep(3)
        if not accessOk:
            response += f"\nDATA LOG FAILED, path = {dataFile}"
        return [past, response , None, gptModel]
    else:
        return [[], "User name and/or password are incorrect", prompt, gptModel]

def new_func(user):
    dataFile = dataDir + user + '_log.txt'
    return dataFile

def transcribe(user, pwd, fpath):
    user = user.lower().strip()
    pwd = pwd.lower().strip()
    if not (user in unames and pwd in pwdList):
        return 'Bad credentials'
    with audioread.audio_open(fpath) as audio:
        duration = int(audio.duration)
        if duration > 0:
            with open(dataDir + user + '_audio.txt','a') as f:
                f.write(f'audio:{str(duration)}\n')
    with open(fpath,'rb') as audio_file:
        transcript = client.audio.transcriptions.create(
            model='whisper-1', file = audio_file ,response_format = 'text' )
    reply = transcript
    return str(reply)

def pause_message():
    return "Audio input is paused.  Resume or Stop as desired"

# def gen_output_audio(txt):
#     if len(txt) < 10:
#         txt = "This dialog is too short to mess with!"
#     response = client.audio.speech.create(model="tts-1", voice="fable", input=txt)
#     with open(speak_file, 'wb') as fp:
#         fp.write(response.content)
#     return speak_file


def set_speak_button(txt):
    vis = False
    if len(txt) > 2:
        vis = True
    return gr.Button(visible=vis)

def update_user(user_win):
    user_win = user_win.lower().strip()
    user = 'unknown'
    for s in unames:
        if user_win == s:
            user = s
            break
    return [user, user]

def speech_worker(chunks=[],q=[]):
    for chunk in chunks:
        fpath = q.pop(0)
        response = client.audio.speech.create(model="tts-1", voice="fable", input=chunk, speed=0.85, response_format='wav')
        with open(fpath, 'wb') as fp:
            fp.write(response.content)

def gen_speech_file_names(user, cnt):
    rv = []
    for i in range(0, cnt):
        rv.append(dataDir + f'{user}_speech{i}.wav')
    return rv

def final_clean_up(user):
    if user.strip().lower() == 'all':
        flist = glob(dataDir + '*_speech*.wav')
    else:
        flist = glob(dataDir + f'{user}_speech*.wav')
    for fpath in flist:
        try:
            os.remove(fpath)
        except:
            continue


def list_permanent_files():
    flist = os.listdir(dataDir)
    others = []
    log_cnt = 0
    wav_cnt = 0
    other_cnt = 0
    for fpath in flist:
        if fpath.endswith('.txt'):
            log_cnt += 1
        elif fpath.endswith('.wav'):
            wav_cnt += 1
        else:
            others.append(fpath)
    other_cnt = len(others)
    return (str(log_cnt), str(wav_cnt), str(other_cnt), str(others))


with gr.Blocks() as demo:
    history = gr.State([])
    password = gr.State("")
    user = gr.State("unknown")
    model = gr.State("gpt-4o-mini")
    q = gr.State([])
    qsave = gr.State([])

    def clean_up(user):
        flist = glob(dataDir + f'{user}_speech*.wav')
        for fpath in flist:
            try:
                os.remove(fpath)
            except:
                continue

    def initial_audio_output(txt, user):
        global digits
        global abbrevs
        if not user in unames:
            return [gr.Audio(sources=None), []]
        clean_up(user)
        q = []
        if len(txt.strip()) < 5:
            return ['None', q]
        for s,x in abbrevs.items():
            txt = txt.replace(s, x)
        words_in = txt.replace('**', '').splitlines(False)
        words_out = []
        for s in words_in:
            s = s.lstrip('- *@#$%^&_=+-')
            if len(s) > 0:
                loc = s.index(' ')
                if loc > 1:
                    val = s[0:loc]
                    isnum = val.replace('.','0').isdecimal()
                    if isnum:
                        if val.endswith('.'):
                            val = val[:-1].replace('.',' point ') + '., '
                        else:
                            val = val.replace('.', ' point ') + ', '
                        s = 'num'+ val + s[loc:]
                words_out.append(s)
        chunklist = []
        for chunk in words_out:
            if chunk.strip() == '':
                continue
            isnumbered = chunk.startswith('num')
            number = ''
            loc = 0
            if isnumbered:
                chunk = chunk[3:]
                loc = chunk.index(',')
                number = chunk[0:loc]
                chunk = chunk[loc:]
            locs = []
            for i in range(1,len(chunk)-1):
                (a, b, c) = chunk[i-1:i+2]
                if a.isdecimal() and b == '.' and c.isdecimal():
                    locs.append(i)
            for i in locs:
                chunk = chunk[:i] + ' point ' + chunk[i+1:]
            if len(chunk) > 50:
                finechunks = chunk.split('.')
                for fchunk in finechunks:
                    if isnumbered:
                        fchunk = number + fchunk
                        isnumbered = False
                    if len(fchunk) > 0:
                        if fchunk != '"':
                            chunklist.append(fchunk)
            else:
                line = number + chunk
                if line != '"':
                    chunklist.append(line)
        total_speech = 0
        for chunk in chunklist:
            total_speech += len(chunk)
        with open(dataDir + user + '_speech.txt','a') as f:
            f.write(f'speech:{str(total_speech)}\n')
        chunk = chunklist[0]
        if chunk.strip() == '':
            return gr.Audio(sources=None)
        fname_list = gen_speech_file_names(user, len(chunklist))
        q = fname_list.copy()
        qsave = fname_list.copy()
        fname = q.pop(0)
        if len(chunklist) > 0:
            threading.Thread(target=speech_worker, daemon=True, args=(chunklist[1:],fname_list[1:])).start()
        response = client.audio.speech.create(model="tts-1", voice="fable", input=chunk, speed=0.85, response_format='wav')
        with open(fname, 'wb') as fp:
            fp.write(response.content)
        return [fname, q]

    def gen_output_audio(q, user):
        try:
            fname = q.pop(0)
        except:
            final_clean_up(user)
            return [None, gr.Audio(sources=None)]
        return [fname, q]


    gr.Markdown('# GPT Chat')
    gr.Markdown('Enter user name & password then enter prompt and click submit button.  Restart conversation if topic changes.  ' +
      'You can enter prompts by voice.  Tap "Record", speak, then tap "Stop".  ' +
      'Tap "Reset Voice Entry" to enter more voice. Tap "Speak Dialog" to hear dialog.  ' + 
               'Note: first voice response may take a longer time.')
    with gr.Row():
        user_window = gr.Textbox(label = "User Name")
        user_window.blur(fn=update_user, inputs=user_window, outputs=[user, user_window])
        pwd_window = gr.Textbox(label = "Password")
        pwd_window.blur(updatePassword, inputs = pwd_window, outputs = [password, pwd_window])
    with gr.Row():
        audio_widget = gr.Audio(type='filepath', format='wav',waveform_options=gr.WaveformOptions(
           show_recording_waveform=True), sources=['microphone'], scale = 3, label="Prompt/Question Voice Entry", max_length=120)
        reset_button = gr.ClearButton(value="Reset Voice Entry", scale=1) #new_func1()
    with gr.Row():
        clear_button = gr.Button(value="Restart Conversation")
        # gpt_chooser=gr.Radio(choices=[("GPT-3.5","gpt-3.5-turbo"),("GPT-4o","gpt-4o-mini")],
        #                      value="gpt-3.5-turbo", label="GPT Model", interactive=True)
        submit_button = gr.Button(value="Submit Prompt/Question")
        speak_output = gr.Button(value="Speak Dialog", visible=False)
    prompt_window = gr.Textbox(label = "Prompt or Question")
    output_window = gr.Textbox(label = "Dialog")
    submit_button.click(chat, inputs=[prompt_window, user_window, password, history, output_window, model],
                         outputs=[history, output_window, prompt_window, model])
    clear_button.click(fn=new_conversation, inputs=user_window, outputs=[prompt_window, history, output_window])
    audio_widget.stop_recording(fn=transcribe, inputs=[user_window, password, audio_widget],
                                outputs=[prompt_window])
    audio_widget.pause_recording(fn=pause_message, outputs=[prompt_window])
    reset_button.add(audio_widget)
    audio_out = gr.Audio(autoplay=True, visible=False)
    audio_out.stop(fn=gen_output_audio, inputs=[q, user_window], outputs = [audio_out, q])
    speak_output.click(fn=initial_audio_output, inputs=[output_window, user_window], outputs=[audio_out, q])
    output_window.change(fn=set_speak_button, inputs=output_window,outputs=speak_output)
    # demo.unload(final_clean_up(user))
demo.launch(share=True)