Voices that
sound human.
Type anything and hear it in a natural voice, made right on your Mac. Direct how a line should feel, add a laugh where you want one, and clone your own voice. Now in Singularity Desktop.
Narration· Serena, Qwen3-TTS 1.7B
Made on a MacBook Air in 12 seconds.
Every voice here was made on a MacBook Air
Direction
Direct the performance, not just the words.
One line, five deliveries. Describe the feeling in plain words, drop in a sound, or hand the line to a second speaker.
Plain text. The voice reads it the way the words suggest.
Everything in the studio
Don't read about it. Press play.
Every card below was made in Voice Studio on the same MacBook Air, with the network switched off.
Design a voice
Describe it. Hear it.
Ah, a traveller. Sit by the fire, and I shall tell you how the stars were made.
Made on a MacBook Air in 13 seconds.
Clone a voice
Six seconds in. Any line out.
The short clip
A new line, same voice
For this demo the short clip is a synthetic voice. Yours never leaves your Mac.
Multiple speakers
A scene, not just a sentence.
Up to ten speakers, each with their own voice, model and style, joined into one take.
Audio tags
Laugh on cue.
[laugh] You really thought that would work? [sigh] Fine. Let's try it your way.
Chatterbox Turbo
Read aloud anywhere
Select text. Press three keys.
Hey all,
Quick update for the team. The launch is now on Thursday, and the slides are in the shared folder.
Made on a MacBook Air in 0.6 seconds.
Local speech API
Your voices, in your other apps.
Speech API at 127.0.0.1:8096
Turn on the Speech API and any OpenAI-compatible app on your Mac can speak with your voices. Off by default.
Models
One line, every engine.
The same sentence through each voice engine, with how long it took to make. Pick by ear, by speed or by size. Every model downloads once and runs on your Mac from then on.
Qwen
Resemble AI
Canopy Labs
Nari Labs
hexgrad
Gradium
ekwekApple
Fish Audio
KittenML
Mistral
Rumik
Sarvam
OpenBMB
Meituan LongCat
OpenMOSS
StepFun
Maya Research
NVIDIA
Boson AIQwen3-TTS 1.7B
Qwen
1.1× slower than real time
Chatterbox
Resemble AI
2.4× slower than real time
Orpheus 3B
Canopy Labs
2.9× slower than real time
Chatterbox Turbo
Resemble AI
2.1× faster than real time
Qwen3-TTS 0.6B
Qwen
1.3× faster than real time
Kokoro 82M
hexgrad
10× faster than real time
Phonon 31M
Gradium
22× faster than real time
Soprano 80M
ekwek
13× faster than real time
Timed on a MacBook Air with an M4 and 16 GB, model already loaded. Also in the app: Qwen3-TTS Clone and Design, Dia 1.6B for two-speaker dialogue (best with 32 GB or more), and the voices built into macOS. More models, on your Mac and on the Grid, are on the way.
Private by default
Your words, and your voice, stay on your Mac.
0 bytes
of your text or voice leave your Mac
$0
per word, per voice, per month
0
accounts or API keys needed
Once a model is downloaded, Voice Studio works with no internet connection. Your text, recordings and cloned voices stay in your Mac's Application Support folder.
Singularity Grid
Your Mac can speak for others, and get paid.
Switch the app to Grid, pick Kokoro under TTS, and your Mac speaks for developers and agents across the network. Here is roughly what that is worth.
You keep per day
$1.92
Per 30 days
$57.60
An estimate at the live rate: $8 per million bytes of text (minimum $0.0001 a request), 80% to you. Real earnings depend on how many requests the Grid sends your Mac.
One request, start to finish
An app or agent asks for speech
Over the OpenAI-style speech API, paid with credits or per call.
“Read this to me, please.”The text is sealed for your Mac
End-to-end encrypted. Only the node that speaks it can read it.
9f2c·a71e·04bd·e6Your Mac speaks it with Kokoro
Locally, in well under a second for a typical line.
Audio goes back. You get paid.
80% of the price lands in your earnings.
+$0.0019 to you
Your Mac has a voice now.
Voice Studio ships today in Singularity Desktop for Macs with Apple silicon. Windows is on the way.
