Project 2501 is an open-source AI assistant, written in C++.
Currently it can tell you the time, the weather, or any information it knows about from having read basically the entire internet :)
You can speak to it and it will speak back to you.
Project 2501 is a fun project to experiment with a partially-offline AI assistant. There's lots of stuff I want to add to it:
-
Store all previous conversations locally, and access them with associative memory. This way Project 2501 can remember the details of previous conversations.
-
Add more internet capabilities - being able to search the web for information etc.
-
Allow time-based thinking, for example being able to set timers, or to remind the user to do something. e.g. "Hey nick, you should take in your washing, it's about to rain outside"
-
Try one of the offline/local large-language models.
Project 2501 can currently tell you the time, tell you the weather outside (current location hard-coded to Wellington, NZ), tell you the current system volume, allow you to set the volume, and tell you about anything it knows about (more or less the entire internet up the Chat GPT Training data).
Adding a new capability is a matter of writing a little C++ code to query the system, or querying an internet API, and presenting the results as a string to the chat system.
I use Windows Speech API (SAPI) for Wake-word detection, as SAPI uses very little CPU, unlike Whisper.cpp.
Once the wake-word/phrase ("Hey Twenty") is detected, I use Whisper.cpp (https://github.com/ggerganov/whisper.cpp) for speech-to-text. This works pretty well but is a bit slow. This should be faster when some mulithreading issues in Whisper.cpp are fixed (see ggerganov/whisper.cpp#200 (comment))
Once the voice query has been converted to text, it's sent off to OpenAI's chat API (similar to ChatGPT).
The response from the chat API is then spoken with Windows Speech API text-to-speech system.
Project 2501 is written in C++, and currently works only on Windows (due to use of Windows Speech API).
Windows and Visual Studio 2022.
"C++ ATL for latest xx build tools" should be selected/installed in the Visual Studio Installer.
You will need Ruby installed (for build_libressl.rb script)
For running the program, you will need a microphone that your computer can detect.
Check out https://github.com/glaretechnologies/glare-core to a directory
Define the GLARE_CORE_LIBS environment variable, to something like c:\programming. This is the directory the LibreSSL directory will be made in.
Build LibreSSL by changing into the scripts directory in glare-core (whereever you checked it out), then running build_libressl.rb:
cd glare-core/trunk
ruby build_libressl.rb
This should build files into e.g. C:\programming\LibreSSL\libressl-3.5.2-x64-vs2022-install-debug and C:\programming\LibreSSL\libressl-3.5.2-x64-vs2022-install
Get the latest Simple DirectMedia Layer (SDL) code from https://www.libsdl.org/ and build it.
Make a build directory, and run cmake, using the locations where you checked out the project-2501 and glare-core code and where you built SDL.
mkdir project_2509_build_dir
cd project_2509_build_dir
cmake N:\project-2501\trunk\ -DGLARE_CORE_TRUNK=n:/glare-core/trunk -DSDL_BUILD_DIR=C:/programming/SDL/sdl_build_vs2022
Now open aibot.sln and build it.
You will need an OpenAI API key, which is free and comes with some free credit.
Place in e.g. project_2509_build_dir\openai_API_key.txt.
The program will complain if it can't find it and tell you where to put it.
This file contains the parameters for the Whisper neural network.
Download ggml-base.en.bin from https://huggingface.co/ggerganov/whisper.cpp/tree/main and place in project_2509_build_dir.