MUCKYWORX // PROJECT
Voice Chaser
PROJECT

Voice Chaser

My absolute most important tool: the Voice Chaser. Without this thing, I basically can't get anything done.

So, the Voice Chaser. I built this thing because I work with speech an extreme amount, and it basically saves me a ton of work, especially the typing. The basic idea was: I speak into the microphone and the transcription is typed right at the cursor position.

It actually started with me wanting that as the main function, and then the possibility of inserting up to three individual AIs in between was added. Depending on what prompt I use for the individual AI stages, the text is modified. Most of the time, it's just for fun to change the style somehow – either making it inspiring or using an archaic way of speaking. But very often it serves for translation: I speak something in, point 1: the text is roughly translated, then it's proofread, and then it comes out again exactly where my cursor is.

The next function was that the main audio files are saved – in my case, on the NAS. This has created a dataset of over 63 hours. After distillation, it's basically the dataset for MuckyTTS. Furthermore, the system includes my RAM disk handler; essentially a program that uploads the required AI models and everything else to the RAM disk based on my config files.

As I said, the Voice Chaser is my primary tool that I use pretty much all the time. By the way, this text right here is being created this way. And other than that, the Voice Chaser also includes the option to assign the AI-4 slot to a separate AI that is then accessible via API. And what else do we have? Well, you're welcome to check out all the underlying infrastructure stuff in the source code. If you have any questions about it, feel free to fire away.

Now, since I'm such a lunatic and have more than four monitors in my setup, I found an old GTX 1050 with 2 GB VRAM in my cabinet. And I had nothing better to do than to misuse this little graphics card to provide the fallback for the Voice Chaser. Should the big 3090 be busy, there is an automatic fallback to the small card, where a "Whisper Small" or whatever is currently current and I find great, is running.

Additionally, the whole system has a stream function so that I don't have to press the button every single time. By the way, I misused an old USB keypad that I bought for 8 euros or so on Amazon to basically build a Stream Deck Lite out of it. I modded this USB keypad so that my keys run there, which I now use to start and end the dictation, as I have it right now. So: press button, blabber away, press button, wait a moment, text appears exactly at the cursor position.

And that's where I put the stream button. When I have the stream active, I can talk on and on without stopping, then it runs through the entire chain and eventually comes out at the cursor position. And that also works with the downstream AI systems.