Have been looking at what it would take to make a small voice assistant similar to Amazons Alexa. We have one (several, actually) but all we use them for is streaming music and the occasional kitchen timer. I was wondering if there were some pieces out there that could be put together on, say, a Raspberry Pi to do something similar. This is very early stages but wanted to put these bit out now, mostly for future reference. I realize there are a few similar projects out there but I didnt see any that looked like what I wanted. But they are a good place to start getting ideas. Some obvious bits:

Voice Recognition: I dont have a microphone on my desktop Raspberry Pi 4 but I found ancient USB camera that has a built-in mic. It sees to do the job. I looked at a few packages for voice recognition, but they tend to be expensive and complex. The one that looks the most interesting at this time is OpenAI Whisper. It as some quirks, like having to pad everything to 30 second sound clips, but it will be fun to play with. It also seems like a stable project likely to be around a while.

Text to Speech: we will also need a way for the system to communicate back without a permanent terminal / keyboard interface. I havent done much in this area either but this used to require specialized hardware i.e. “soundcards”. I know, that was a long time ago. Looking around I found eSpeak ready for install on Ubuntu. It has a bunch of different voices and was easy to get it up and running and it seems solid. My wife said it sounded too robotic though (not her words).

Sound Output: The Raspberry Pi 5 doesnt have a dedicated sound output like the old versions, but I have a bluetooth speaker I use for playing music. It works just fine with eSpeak.

Text Search: comparing a text string to other text strings, especially for non-exact matches is not something I want to code up myself. Again, lots of stuff out there, but I want something simple and stable. I also probably want something directly in python. FuzzyWuzzy looks good and popular as does FuzzySet and the more standard SequenceMatcher. All seem to use the difflib and Levenshtein distance. Sounds good to me. There are probably some newer AI-based approaches but I want to keep it simple.

Wake Word: One problem is “waking up” the assistant. There is usually a “wake word”. This more or less means listening 24 / 7 until this word is spoken. Will have to see how this works. A simple loop will probably be good for starters, but maybe a bit wasteful. But that might not matter.

Commands: I figure a table of expected voice commands mapped to actual Linux command line commands is the easiest. Need to figure out how best to get to things like Spotify, but that is another problem. I assume running some streams will be set up using things like ffmpeg and will require a little knowledge, but it should also be pretty simple.

Music Search: I expect to have my MP3 library either locally of remotely hosted. Getting an actual song or album from test to a file(s) location might require a scan and some sort of simple database (maybe just a table). Havent thought this one out much yet.

Thats about it for now. Will probably do this in pieces on my desktop and then eventually deploy on a smaller dedicated system.