Hello everyone,
For some time now, I’ve been teasing Pierre-Gilles about a new service based on Gradium. It’s a French startup specializing in voice AI that’s been making waves. The free subscription allows for 1 hour of text-to-speech per month.
The idea behind the service is to let those who want to tinker a bit choose a custom voice (they even offer the voice of General de Gaulle), or to test TTS for free for new Gladys users.
Once again, Gladys Plus already gives you a ready-to-use voice and supports the project, so don’t hesitate to subscribe to Gladys Plus.
On the developer side, this is a first step that allowed us to add a TTS abstraction, and my next goal is to run local models. @McFlyPartages mentioned supertonic, version 3 was recently released and is even more performant. My initial tests are quite conclusive, but I’m trying to compare it with other models available on HuggingFace. I hope to offer you something soon.
In the meantime, for the Gradium service, I tested it with Airplay. I’ll need testers for Google Cast and Sonos. A Docker image is available bertrandda/gladys:gradium
Hi, thanks for your work.
I know there’s also Voxtral that runs locally with very good performance but hasn’t been tested yet.
If I have time to test your development, I’ll give you feedback. I have Sonos (Ikea) speakers.
Yes I saw, but Voxtral seems to work from a good small config. My supertonic tests on Raspberry PI 4 8GB allowed me to generate audio clips of about ten seconds in ~8 seconds. Here are a few examples generated in Nodejs on RPI 4 (the links are valid for 7 days, I could re-upload them if necessary) :
Other voices are available including male voices
If it runs on PI4 it will run on better machines (with even better results, by increasing the total_steps parameter for example). The RPI 4 CPU is from 2021 (see benchmark below) with newer devices (such as the Beelink offered via the Gladys all-in-one kit) the latency time will certainly be lower.
Note that in what I imagine, the goal of Gradium and later local models is to have alternatives for the audio notification block in scenes only (not the voice assistant widget of the dashboard which requires an LLM and therefore Gladys Plus), so having a few seconds of delay doesn’t seem very serious.
And once a first local model is implemented, nothing prevents adding others that are larger for more powerful machines.
Meanwhile, for those who want to test Gradium, an image with the latest fixes is available bertrandda/gladys:gradium
