This voice is my phone model, and it's only 25% trained.
A 126 MB on-device TTS model learning to talk (and express) from 400 h of expressive speech, trained on 2x NVIDIA V100S GPUs.
Full model tonight.
Progress on tiny expressive TTS: QC kept 99.7 % of 240K clips (386 h). The 24-layer teacher finished 20K steps in 13 h on my two V100S GPUs.
Now distilling it into a 6-layer, ~126 MB on-device model.
Results tonight.
1
22
Training a tiny on-device voice model that can talk naturally like laugh, sigh and gasp.
I had a big model generate 240 K lines of audio, filtered out the bad takes, and now teaching a 126 MB model to talk like it.
Results in about a day
20
Distilled Dia-1.6B into a 126 MB on-device TTS. Speech came out clean (WER 0.005), but the pauses and chuckles sounded fake, and Dia's own takes did too. The dataset was the problem, not the training. Ran a bake-off of 3 datasets, Orpheus won, regenerated 414 h of data. Retraining now.
1
48
Hey truckers, long haulers, dispatchers, shippers and trucks!
You need to see this: haitruck.us
28
Fateh retweeted
P.S. I'm open to new roles.
If you like what AXe does and want someone who builds tools like it on your team, get in touch.
cameroncooke.com
Fateh retweeted
My AI Fundamentals Course is close to ready to go live.
aifundamentals.dev
I'd love to give a few of you early access, for free, to test it out.
Reply & Repost if you're interested, invite goes out tomorrow!
this is an evidence how you can create great marketing videos without putting your app on dance floor (all the colors and unnecessary animations)
1
48
How I went from Full Time Senior Product Engineer to an unemployed vibe coder, here are the steps:
[Thread]
36
On your mac.
No backend.
No API bills.
Fully offline.
I built Rona for Mac with SwiftUI, a voice companion powered by Apple Intelligence and a 174.5 MB local TTS model.
You talk. It talks back, in different tones, warm & sharp.
Source code: github.com/Fatehkhann/rona-v…
The stack behind on-device voice personas:
Trainer: Sesame CSM-1B by @sesame . It speaks every training line in context: persona voice → user turn → reply, so the delivery stays conversational.
Trainee: Kyutai Pocket TTS, fine-tuning the 24-layer teacher, then distilling to 6 layers for iOS.
Building voice personas for a fun experimental iOS app that run fully on-device. No cloud, no audio leaving the phone.
Blended voices, conversational style distilled from Sesame CSM into Pocket TTS.
Fine-tuning now.
Results soon!
63
Building voice personas for a fun experimental iOS app that run fully on-device. No cloud, no audio leaving the phone.
Blended voices, conversational style distilled from Sesame CSM into Pocket TTS.
Fine-tuning now.
Results soon!







