{ "cells": [ { "cell_type": "code", "execution_count": null, "id": "3bedf0dc-8d8e-4ede-a9e6-b8f35136aa00", "metadata": {}, "outputs": [], "source": [ "#|default_exp app" ] }, { "cell_type": "code", "execution_count": null, "id": "667802a7-0f36-4136-a381-e66210b20462", "metadata": {}, "outputs": [], "source": [ "#| export\n", "#tts_openai_secrets.py content:\n", "#import os\n", "#os.environ['OPENAI_API_KEY'] = 'sk-XXXXXXXXXXXXXXXXXXXXXX'\n", "import os\n", "secret_import_failed = False\n", "try:\n", " _ = os.environ['OPENAI_API_KEY']\n", " print('OPENAI_API_KEY environment variable was found.')\n", "except:\n", " print('OPENAI_API_KEY environment variable was not found.')\n", " secret_import_failed = True\n", "try:\n", " GRADIO_PASSWORD = os.environ['GRADIO_PASSWORD']\n", " print('GRADIO_PASSWORD environment variable was found.')\n", "except:\n", " print('GRADIO_PASSWORD environment variable was not found.')\n", " secret_import_failed = True\n", "\n", "if secret_import_failed == True:\n", " import tts_openai_secrets\n", " GRADIO_PASSWORD = os.environ['GRADIO_PASSWORD']\n", " print('import tts_openai_secrets succeeded')" ] }, { "cell_type": "code", "execution_count": null, "id": "4d9863fc-969e-409b-8e20-b9c3cd2cc3e7", "metadata": {}, "outputs": [], "source": [ "#| hide\n", "try:\n", " import nbdev\n", "except:\n", " print('to convert this notebook to app.py you need to pip install nbdev')" ] }, { "cell_type": "code", "execution_count": null, "id": "4f486d3a", "metadata": {}, "outputs": [], "source": [ "#| export\n", "import gradio as gr\n", "import openai\n", "from pydub import AudioSegment\n", "import io\n", "from datetime import datetime\n", "from math import ceil\n", "from multiprocessing.pool import ThreadPool\n", "from functools import partial\n", "from pathlib import Path\n", "import uuid\n", "from tenacity import (\n", " retry,\n", " stop_after_attempt,\n", " wait_random_exponential,\n", ") # for exponential backoff" ] }, { "cell_type": "code", "execution_count": null, "id": "ecb7f207-0fc2-4d19-a313-356c05776832", "metadata": {}, "outputs": [], "source": [ "#| export\n", "TEMP = os.environ['TEMP']\n", "TEMP_DIR = Path(TEMP)\n", "print('TEMP Dir:', TEMP_DIR)" ] }, { "cell_type": "code", "execution_count": null, "id": "0ffd33b4-cb9b-4c01-bff6-4c3102854ab6", "metadata": {}, "outputs": [], "source": [ "#| export\n", "try:\n", " tts_models = [o.id for o in openai.models.list().data if 'tts' in o.id]\n", " print('successfully got tts model list:', tts_models)\n", "except:\n", " tts_models = ['tts-1']" ] }, { "cell_type": "code", "execution_count": null, "id": "2ddbca5d-4b04-43ab-afaf-430802980e78", "metadata": {}, "outputs": [], "source": [ "#| export\n", "tts_voices = ['alloy', 'echo', 'fable', 'onyx', 'nova', 'shimmer']" ] }, { "cell_type": "code", "execution_count": null, "id": "8eb7e7d5-7121-4762-b8d1-e5a9539e2b36", "metadata": {}, "outputs": [], "source": [ "#| export\n", "clean_text_prompt = \"\"\"Your job is to clean up text that is going to be fed into a text to speech (TTS) model. You must remove parts of the text that would not normally be spoken such as reference marks `[1]`, spurious citations such as `(Reddy et al., 2021; Wu et al., 2022; Chang et al., 2022; Kondratyuk et al., 2023)` and any other part of the text that is not normally spoken. Please also clean up sections and headers so they are on new lines with proper numbering. You must also clean up any math formulas that are salvageable from being copied from a scientific paper. If they are garbled and do not make sense then remove them. You must carefully perform the text cleanup so it is translated into speech that is easy to listen to however you must not modify the text otherwise. It is critical that you repeat all of the text without modifications except for the cleanup activities you've been instructed to do. Also you must clean all of the text you are given, you may not omit any of it or stop the cleanup task early.\"\"\"\n" ] }, { "cell_type": "code", "execution_count": null, "id": "52d373be-3a79-412e-8ca2-92bb443fa52d", "metadata": {}, "outputs": [], "source": [ "#| export\n", "#Number of threads created PER USER REQUEST. This throttels the # of API requests PER USER request. This is in ADDITION to the Gradio threads.\n", "OPENAI_CLIENT_TTS_THREADS = 10 " ] }, { "cell_type": "code", "execution_count": null, "id": "24674094-4d47-4e48-b591-55faabcff8df", "metadata": {}, "outputs": [], "source": [ "#| export\n", "def split_text(input_text, max_length=4000, lookback=1000):\n", " # If the text is shorter than the max_length, return it as is\n", " if len(input_text) <= max_length:\n", " return [input_text]\n", "\n", " chunks = []\n", " while input_text:\n", " # Check if the remaining text is shorter than the max_length\n", " if len(input_text) <= max_length:\n", " chunks.append(input_text)\n", " break\n", "\n", " # Define the split point, initially set to max_length\n", " split_point = max_length\n", "\n", " # Look for a newline in the last 'lookback' characters\n", " newline_index = input_text.rfind('\\n', max_length-lookback, max_length)\n", " if newline_index != -1:\n", " split_point = newline_index + 1 # Include the newline in the current chunk\n", "\n", " # If no newline, look for a period followed by space\n", " elif '. ' in input_text[max_length-lookback:max_length]:\n", " # Find the last '. ' in the lookback range\n", " period_index = input_text.rfind('. ', max_length-lookback, max_length)\n", " split_point = period_index + 2 # Split after the space\n", "\n", " # Split the text and update the input_text\n", " chunks.append(input_text[:split_point])\n", " input_text = input_text[split_point:]\n", "\n", " return chunks" ] }, { "cell_type": "code", "execution_count": null, "id": "e6224ae5-3792-42b2-8392-3abd42998a50", "metadata": {}, "outputs": [], "source": [ "#| export\n", "def concatenate_mp3(mp3_files):\n", " if len(mp3_files) == 1:\n", " return mp3_files[0]\n", " else:\n", " # Initialize an empty AudioSegment object for concatenation\n", " combined = AudioSegment.empty()\n", " \n", " # Write out audio file responses as individual files for debugging\n", " # for idx, mp3_data in enumerate(mp3_files):\n", " # with open(f'./{idx}.mp3', 'wb') as f:\n", " # f.write(mp3_data)\n", "\n", " # Loop through the list of mp3 binary data\n", " for mp3_data in mp3_files:\n", " # Convert binary data to an audio segment\n", " audio_segment = AudioSegment.from_file(io.BytesIO(mp3_data), format=\"mp3\")\n", " # Concatenate this segment to the combined segment\n", " combined += audio_segment\n", "\n", " #### Return Bytes Method\n", " # # Export the combined segment to a new mp3 file\n", " # # Use a BytesIO object to handle this in memory\n", " # combined_mp3 = io.BytesIO()\n", " # combined.export(combined_mp3, format=\"mp3\")\n", "\n", " # # Seek to the start so it's ready for reading\n", " # combined_mp3.seek(0)\n", "\n", " # return combined_mp3.getvalue()\n", "\n", " #### Return Filepath Method\n", " filepath = TEMP_DIR/(str(uuid.uuid4())+'.mp3')\n", " combined.export(filepath, format=\"mp3\")\n", " return str(filepath)" ] }, { "cell_type": "code", "execution_count": null, "id": "4691703d-ed0f-4481-8006-b2906289b780", "metadata": {}, "outputs": [], "source": [ "#| export\n", "def create_speech_openai(chunk_idx, input, model='tts-1', voice='alloy', speed=1.0, **kwargs):\n", " client = openai.OpenAI()\n", " \n", " @retry(wait=wait_random_exponential(min=1, max=180), stop=stop_after_attempt(6))\n", " def _create_speech_with_backoff(**kwargs):\n", " return client.audio.speech.create(**kwargs)\n", " \n", " response = _create_speech_with_backoff(input=input, model=model, voice=voice, speed=speed, **kwargs)\n", " client.close()\n", " return chunk_idx, response.content" ] }, { "cell_type": "code", "execution_count": null, "id": "e34bb4aa-698c-4452-8cda-bd02b38f7122", "metadata": {}, "outputs": [], "source": [ "#| export\n", "def create_speech2(input_text, model='tts-1', voice='alloy', profile: gr.OAuthProfile|None=None, progress=gr.Progress(), **kwargs):\n", " print('cs2-profile:',profile)\n", " assert authorized(profile) is not None,'Unauthorized M'\n", " start = datetime.now()\n", " # Split the input text into chunks\n", " chunks = split_text(input_text)\n", "\n", " # Initialize the progress bar\n", " progress(0, desc=f\"Started processing {len(chunks)} text chunks using {OPENAI_CLIENT_TTS_THREADS} threads. ETA is ~{ceil(len(chunks)/OPENAI_CLIENT_TTS_THREADS)} min.\")\n", "\n", " # Initialize a list to hold the audio data of each chunk\n", " audio_data = []\n", "\n", " # Process each chunk\n", " with ThreadPool(processes=OPENAI_CLIENT_TTS_THREADS) as pool:\n", " results = pool.starmap(\n", " partial(create_speech_openai, model=model, voice=voice, **kwargs), \n", " zip(range(len(chunks)),chunks)\n", " )\n", " audio_data = [o[1] for o in sorted(results)]\n", "\n", " # Progress\n", " progress(.9, desc=f\"Merging audio chunks... {(datetime.now()-start).seconds} seconds to process.\")\n", " \n", " # Concatenate the audio data from all chunks\n", " combined_audio = concatenate_mp3(audio_data)\n", "\n", " # Final update to the progress bar\n", " progress(1, desc=f\"Processing completed... {(datetime.now()-start).seconds} seconds to process.\")\n", " \n", " print(f\"Processing time: {(datetime.now()-start).seconds} seconds.\")\n", "\n", " return combined_audio\n" ] }, { "cell_type": "code", "execution_count": null, "id": "5388e860", "metadata": {}, "outputs": [], "source": [ "#| export\n", "def create_speech(input_text, model='tts-1', voice='alloy', profile: gr.OAuthProfile|None=None, progress=gr.Progress()):\n", " assert authorized(profile) is not None,'Unauthorized M'\n", " # Split the input text into chunks\n", " chunks = split_text(input_text)\n", "\n", " # Initialize the progress bar\n", " progress(0, desc=\"Starting TTS processing...\")\n", "\n", " # Initialize a list to hold the audio data of each chunk\n", " audio_data = []\n", "\n", " # Create a client instance for OpenAI\n", " client = openai.OpenAI()\n", "\n", " # Calculate the progress increment for each chunk\n", " progress_increment = 1.0 / len(chunks)\n", "\n", " # Process each chunk\n", " for i, chunk in enumerate(chunks):\n", " response = client.audio.speech.create(\n", " model=model,\n", " voice=voice,\n", " input=chunk,\n", " speed=1.0\n", " )\n", " # Append the audio content of the response to the list\n", " audio_data.append(response.content)\n", "\n", " # Update the progress bar\n", " progress((i + 1) * progress_increment, desc=f\"Processing chunk {i + 1} of {len(chunks)}\")\n", "\n", " # Close the client connection\n", " client.close()\n", "\n", " # Concatenate the audio data from all chunks\n", " combined_audio = concatenate_mp3(audio_data)\n", "\n", " # Final update to the progress bar\n", " progress(1, desc=\"Processing completed\")\n", "\n", " return combined_audio\n" ] }, { "cell_type": "code", "execution_count": null, "id": "236dd8d3-4364-4731-af93-7dcdec6f18a1", "metadata": {}, "outputs": [], "source": [ "#| export\n", "def get_input_text_len(input_text):\n", " return len(input_text)" ] }, { "cell_type": "code", "execution_count": null, "id": "0523a158-ee07-48b3-9350-ee39d4deee7f", "metadata": {}, "outputs": [], "source": [ "#| export\n", "def get_generation_cost(input_text, tts_model_dropdown):\n", " text_len = len(input_text)\n", " if tts_model_dropdown.endswith('-hd'):\n", " cost = text_len/1000 * 0.03\n", " else:\n", " cost = text_len/1000 * 0.015\n", " return \"${:,.3f}\".format(cost)" ] }, { "cell_type": "code", "execution_count": null, "id": "b5b29507-92bc-453d-bcc5-6402c17e9a0d", "metadata": {}, "outputs": [], "source": [ "#| export\n", "def authorized(profile: gr.OAuthProfile=None) -> str:\n", " print('Profile:', profile)\n", " if profile is not None and profile.username in [\"matdmiller\"]:\n", " return f\"{profile.username}\"\n", " else:\n", " print('Unauthorized',profile)\n", " return None" ] }, { "cell_type": "code", "execution_count": null, "id": "e4fb3159-579b-4271-bc96-4cd1e2816eca", "metadata": {}, "outputs": [], "source": [ "#| export\n", "with gr.Blocks(title='OpenAI TTS', head='OpenAI TTS') as app:\n", " gr.Markdown(\"# OpenAI TTS\")\n", " gr.Markdown(\"\"\"Start typing below and then click **Go** to create the speech from your text. The current limit is 4,000 characters. \n", "For requests longer than 4,000 chars they will be broken into chunks of 4,000 or less chars automatically. [Spaces Link](https://matdmiller-tts-openai.hf.space/)\"\"\")\n", " with gr.Row():\n", " input_text = gr.Textbox(max_lines=100, label=\"Enter text here\")\n", " with gr.Row():\n", " tts_model_dropdown = gr.Dropdown(value='tts-1',choices=tts_models, label='Model')\n", " tts_voice_dropdown = gr.Dropdown(value='alloy',choices=tts_voices,label='Voice')\n", " input_text_length = gr.Label(label=\"Number of characters\")\n", " generation_cost = gr.Label(label=\"Generation cost\")\n", " output_audio = gr.Audio()\n", " input_text.input(fn=get_input_text_len, inputs=input_text, outputs=input_text_length)\n", " input_text.input(fn=get_generation_cost, inputs=[input_text,tts_model_dropdown], outputs=generation_cost)\n", " tts_model_dropdown.input(fn=get_generation_cost, inputs=[input_text,tts_model_dropdown], outputs=generation_cost)\n", " go_btn = gr.Button(\"Go\")\n", " go_btn.click(fn=create_speech2, inputs=[input_text, tts_model_dropdown, tts_voice_dropdown], outputs=[output_audio])\n", " clear_btn = gr.Button('Clear')\n", " clear_btn.click(fn=lambda: '', outputs=input_text)\n", "\n", " gr.LoginButton()\n", " m = gr.Markdown('')\n", " app.load(authorized, None, m)\n", " " ] }, { "cell_type": "code", "execution_count": null, "id": "a00648a1-891b-470b-9959-f5d502055713", "metadata": {}, "outputs": [], "source": [ "#| export\n", "# launch_kwargs = {'auth':('username',GRADIO_PASSWORD),\n", "# 'auth_message':'Please log in to Mat\\'s TTS App with username: username and password.'}\n", "launch_kwargs = {}\n", "queue_kwargs = {'default_concurrency_limit':10}" ] }, { "cell_type": "code", "execution_count": null, "id": "4b534fe7-4337-423e-846a-1bdb7cccc4ea", "metadata": {}, "outputs": [], "source": [ "#| hide\n", "#Notebook launch\n", "app.queue(**queue_kwargs)\n", "app.launch(**launch_kwargs)" ] }, { "cell_type": "code", "execution_count": null, "id": "cb886d45", "metadata": {}, "outputs": [], "source": [ "#| export\n", "#.py launch\n", "if __name__ == \"__main__\":\n", " app.queue(**queue_kwargs)\n", " app.launch(**launch_kwargs)" ] }, { "cell_type": "code", "execution_count": null, "id": "28e8d888-e790-46fa-bbac-4511b9ab796c", "metadata": {}, "outputs": [], "source": [ "#| hide\n", "app.close()" ] }, { "cell_type": "code", "execution_count": null, "id": "afbc9699-4d16-4060-88f4-cd1251754cbd", "metadata": {}, "outputs": [], "source": [ "#| hide\n", "gr.close_all()" ] }, { "cell_type": "code", "execution_count": 53, "id": "0420310d-930b-4904-8bd4-3458ad8bdbd3", "metadata": {}, "outputs": [], "source": [ "#| hide\n", "nbdev.export.nb_export('app.ipynb',lib_path='.')" ] }, { "cell_type": "code", "execution_count": null, "id": "9869749d-bc7c-4e24-9dbc-403f665d6200", "metadata": {}, "outputs": [], "source": [] } ], "metadata": { "kernelspec": { "display_name": "gradio1", "language": "python", "name": "gradio1" }, "language_info": { "codemirror_mode": { "name": "ipython", "version": 3 }, "file_extension": ".py", "mimetype": "text/x-python", "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", "version": "3.11.8" } }, "nbformat": 4, "nbformat_minor": 5 }