JARVIS-ChatGPT: A conversational assistant equipped with J.A.R.V.I.S's voice

A voice-based interactive assistant equipped with a variety of synthetic voices (including J.A.R.V.I.S's voice from IronMan)

image by MidJourney AI

Ever dreamed to ask hyper-intelligent system tips to improve your armor? Now you can! Well, maybe not the armor part... This project exploits OpenAI Whisper, OpenAI ChatGPT and IBM Watson.

PROJECT MOTIVATION:

Many times ideas come in the worst moment and they fade away before you have the time to explore them better. The objective of this project is to develop a system capable of giving tips and opinions in quasi-real-time about anything you ask. The ultimate assistant will be able to be accessed from any authorized microphone inside your house or your phone, it should run constantly in the background and when summoned should be able to generate meaningful answers (with a badass voice) as well as interface with the pc or a server and save/read/write files that can be accessed later. It should be able to run research, gather material from the internet (extract content from HTML pages, transcribe Youtube videos, find scientific papers...) and provide summaries that can be used as context to make informed decisions. In addition, it might interface with some external gadgets (IoT) but that's extra.

DEMO:

2023-04-11.23-20-03_Trim.mp4

JULY 14th 2023 UPDATE: Research Mode

I can finnaly share the first draft of the Research Mode. This modality was thought for people often dealing with research papers.

Switch to research mode by saying 'Switch to Research Mode'
⭐ Initialize a new workspace like this: 'Initialize a new workspace about Carbon Fiber Applications in the Spacecraft industry'. A workspace is a folder that collects and organize the results of the research. This protocol is subdivided into 3 sub-routines:
1. Core Paper identification: Use the Semantic Scholar API to identify some strongly relevant papers;
2. Core Expansion: for each paper, finds some suggestions, then keep only the suggestions that appear to be similar to at least 2 paper;
3. Refy Expansion: use the refy suggestion package to enlarge the results;
Find suggestions like: 'find suggestions that are sililar to the paper with title ...'
Download: 'download the paper with title ...'
⭐ Query your database like: 'what is the author of the paper with title ...?' 'what are the experimental conditions set for the paper with title ...?'

PS: This mode is not super stable and needs to be worked on

PPS: This project will be discontinued for some time since I'll be working on my thesis until 2024. However there are already so many things that can be improved so I'll be back!

What you'll need:

DISCLAIMER:
The project might consume your OpenAI credit resulting in undesired billing;
I don't take responsibility for any unwanted charges;
Consider setting limitations on credit consumption at your OpenAI account;

An OpenAI account and API key; (check FAQs below for the alternatives)
PicoVoice account and a free AccessKey; (optional)
ElevenLabs account and free Api Key (optional);
langChain API keys for web surfing (news, weather, serpapi, google-serp, google-search... they are all free)
ffmpeg ;
Python virtual environment (Python>=3.9 and <3.10);
Some credit to spend on ChatGPT (you can get three months of free usage by signing up to OpenAI) (suggested);
CUDA version >= 11.2;
An IBM Cloud account to exploit their cloud-based text-to-speech models (tutorial)(optional);
A (reasonably) fast internet connection (most of the code relies on API so a slower connection might result in a longer time to respond);
mic and speaker;
CUDA capable graphic engine (my Torch Version: 2.0 and CUDA v11.7 pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu117);
Patience 😅

you can rely on the new setup.bat that will do most of the things for you.

GitHub overview

MAIN script you should run: openai_api_chatbot.py if you want to use the latest version of the OpenAI API Inside the demos folder you'll find some guidance for the packages used in the project, if you have errors you might check these files first to target the problem. Mostly is stored in the Assistant folder: get_audio.py stores all the functions to handle mic interactions, tools.py implements some basic aspects of the Virtual Assistant, voice.py describes a (very) rough Voice class. Agents.py handle the LangChain part of the system (here you can add or remove tools from the toolkits of the agents)
The remaining scripts are supplementary to the voice generation and should not be edited.

INSTALLATION TUTORIAL

Automatic installation

You can run setup.bat if you are running on Windows/Linux. The script will perform every step of the manual installation in sequence. Refer to those in case the procedure should fail.
The automatic installation will also run the Vicuna installation (Vicuna Installation Guide)

Manual Installation

Step 1: installation, accounts, APIs...

Environment

Make a new, empty virtual environment with Python 3.8 and activate it (.\venv_name\Scripts\activate );
pip install -r venv_requirements.txt; This might take some time; if you encounter conflicts on specific packages, install them manually without the ==<version>;
install manually PyTorch according to your CUDA VERSION;
Copy and paste the files you'll find in the folder whisper_edits to the whisper folder of your environment (.\venv\lib\site-packages\whisper\ ) these edits will add just an attribute to the whisper model to access its dimension more easily;
install TTS;
Run their script and check everything is working (it should download some models) (you can alternatively run demos/tts_demo.py);
Rename or delete the TTS folder and download the Assistant and other scripts from this repo
Install Vicuna following the instructions on the Vicuna folder or by running:
cd Vicuna
call vicuna.ps1
Manual instructions will instruct you to follow the Vicuna Installation Guide
paste all your keys in the env.txt file and rename it to .env (yes, remove the txt extension)
Check everything works (following)

Checks

Verify your graphic engine and CUDA version are compatible with PyTorch by running torch.cuda.is_available() and torch.cuda.get_device_name(0) inside Pyhton; .
run tests.py. This file attempt to perform basic operations that might raise errors;
[WARNING] Check the FAQs below if you have errors;
You can check the sources of error by running demos in the demos folder;

Step 2: Language support

To have answers spoken in your language you should first check if your language is supported by the speech generator at https://cloud.ibm.com/docs/text-to-speech?topic=text-to-speech-voices;
If it's supported, add or change the languages inside VirtualAssistant.__init__() ;

Remember: The loaded Whisper is the medium one. If it performs badly in your language, upgrade to the larger one in the __main__() at whisper_model = whisper.load_model("large"); but I hope your GPU memory is large likewise.

Step 3: Running (`openai_api_chatbot.py`):

When running, you'll see much information being displayed. I'm constantly striving to improve the readability of the execution, the whole project is a huge beta, forgive slight variations from the screens below. Anyway, this is what happens in general terms when you hit 'run':

Preliminary initializations take place, you should hear a chime when the Assistant is ready;
When awaiting for triggering words is displayed you'll need to say Jarvis to summon the assistant. At this point, a conversation will begin and you can speak in whatever language you want (if you followed step 2). The conversation will terminate when you 1) say a stop word 2) say something with one word (like 'ok') 3) when you stop making questions for more than 30 seconds

After the magic word is said, the word listening... should then appear. At this point, you can make your question. When you are done just wait (3 seconds) for the answer to be submitted;
The script will convert the recorded audio to text using Whisper;
The text will be analyzed and a decision will be made. If the Assistant believes it needs to take some action to respond (like looking for a past conversation) the langchain agents will make a plan and use their tool to answer.
Elsewise, the script will then expand the chat_history with your question, it will send a request with the API and it will update the history as soon as it receives a full answer from ChatGPT (this may take up to 5-10 seconds, consider explicitly asking for a short answer if you are in a hurry);
The say() function will perform voice duplication to speak with Jarvis/Someone's voice; if the argument is not in English, IBM Watson will send the response from one of their nice text-to-speech models. If everything fails, the functions will rely on pyttsx3 which is a fast yet not as cool alternative;

When any of the stop keywords are said, the script will ask ChatGPT to give a title to the conversation and will save the chat in a .txt file with the format 'CurrentDate_Title.txt';
The assistant will then go back to sleep;

I made some prompts and closed the conversation

Keywords:

to stop or save the chat, just say 'THANKS' at some point;
To summon JARVIS voice just say 'JARVIS' at some point;

not ideal I know but works for now

History:

currently working on:

Extend doc processing tools
Find a free alternative for LangChain Agents

following:

fixing chat length bug (when the chat is too long it can't be processed by ChatGPT 3.5 Turbo)
expanding Memory
crash reports
Refine capabilities

waiting for ChatGPT4 to:

add multimodal input (i.e. "Do you think 'this' [holding a paper plane] could fly" -> camera -> ChatGPT4 -> "you should improve the tip of the wings" )
Extend project memory to images, pdfs, papers...

Check the UpdateHistory.md of the project for more insights.

Have fun!

ERRORS and FAQs

categories: Install, General, Runtime

INSTALL: I have conflicting packages while installing venv_requirements.txt, what should I do?

Make sure you have the right Python version (3.7) on the .venv (>python --version with the virtual environment activated).
Try to edit the venv_requirements.txt and remove the version requirements of the incriminated dependencies.
Straight remove the package from the txt file and install them manually afterward.

INSTALL: I meet an error when running openai_api_chatbot.py saying: TypeError: LoadLibrary( ) argument 1 must be str, not None what's wrong?

The problem is concerning Whisper. You should re-install it manually with pip install whisper-openai

INSTALL: I can't import 'openai.embeddings_utils'

Try to pip install --upgrade openai.
This happens because openai elevated their minimum requirements. I had this problem and solved by manually downloading embeddings_utils.py inside ./<your_venv>/Lib/site-packages/openai/

3. If the problem persists with ```datalib``` raise an issue and I'll provide you the missing file 4. upgrade to Python 3.8 (create new env and re-install TTS, requirements)

INSTALL: I encounter the error ModuleNotFoundError: No module named '<some module>'

Requirements are not updated every commit. While this might generate errors you can quickly install the missing modules, at the same time it keeps the environment clean from conflicts when I try new packages (and I try LOTS of them)

RUN TIME: I encounter some OOM memory when loading the Whisper model, what does it mean?

It means the model you selected is too big for your CUDA device memory. Unfortunately, there is not much you can do about it except load a smaller model. If the smaller model does not satisfy you, you might want to speak 'clearer' or make longer prompts to let the model predict more accurately what you are saying. This sounds inconvenient but, in my case, greatly improved my English-speaking :)

RUN TIME: Max length tokens for ChatGPT-3.5-Turbo is 4096 but received... tokens.

This is a bug still present, don't expect to have ever long conversations with your assistant as it will simply have enough memory to remember the whole conversation at some point. A fix is in development, it might consist of adopting a 'sliding windows' approach even if it might cause repetition of some concepts.

GENERAL: I finished my OPENAI credit/demo, what can I do?

Go online only. The price is not that bad and you might end up paying a few dollars a month since pricing depends on usage (with heavy testing I ended up consuming the equivalent of ~4 dollars a month during my free trial). You can set limits on your monthly tokens consumption.
Use a Hybrid mode where the most credit-intensive tasks are executed locally for free and the rest is done online.
Install Vicuna and run OFFLINE mode only with limited performance.

GENERAL: For how long will this project be updated?

Right now (April 2023) I'm working almost non-stop on this. I will likely take a break in the summer because I'll be working on my thesis.

If you have questions you can contact me by raising an Issue and I'll do my best to help as soon as possible.

Gianmarco Guarnier

bytjn1416124/JARVIS-ChatGPT