Miso One — Text to Speech With Actual Emotion
Miso Labs' MisoTTS, embedded here and running on their own Space
Most TTS is intelligible and flat. Miso One is an ~8B-parameter open-weights model that goes after the other half — warmth, pacing, feeling. Type English text into the panel below to hear it, or hand it a short clip to clone.
What you see above is the official MisoTTS Space on Hugging Face, embedded directly. Expect the first generation to lag while the GPU spins up. Miso One and MisoTTS come from Miso Labs; this page is our own write-up wrapped around their demo.
So what is Miso One?
Miso Labs — a Y Combinator-backed voice-AI startup — released this in early June 2026 under the technical name MisoTTS. At roughly 8 billion parameters it is not small, and the size is spent on something specific: delivery. Warmth, pacing, emotional shape. The places where synthetic speech usually gives itself away.
Architecturally it is a Llama 3.2-style backbone feeding a smaller autoregressive audio decoder — an RVQ Transformer taking cues from Sesame's CSM — which emits Mimi audio codes conditioned on your text plus any audio context you supply. Roughly ten seconds of reference is enough for one-shot cloning, and because Miso Labs shipped open weights, none of this has to run on someone else's machine.
The panel at the top of this page is that same Space, so there is nothing to install before forming an opinion.
What it does differently
Expressive delivery, cloning from a short sample, and weights you can actually download.
Emotion that survives synthesis
The target is a reading rather than a recitation: warmth, natural pacing, feeling that tracks the sentence — the qualities most TTS still flattens out.
One-shot voice cloning
About ten seconds of reference audio is enough for the model to take on a voice and read new lines in it. Only ever do this with the speaker's consent.
Fast enough to converse
Miso Labs puts first-token latency near 110 ms, which would make back-and-forth interaction viable. That is their measurement, not an independent one.
How the ~8B model is put together
Llama 3.2-style backbone, a smaller audio decoder behind it — an RVQ Transformer after Sesame's CSM — producing Mimi audio codes from your text.
Open weights, self-hostable
Open weights under a modified MIT licence, which means the option to run it locally and never send text or audio anywhere.
Made to be integrated
Pull MisoLabs/MisoTTS from Hugging Face and wire it into your own stack. A hosted API is listed as coming soon.
English, done properly
English is the sole focus for now, and the payoff is intonation, emphasis and phrasing that follow the punctuation you actually wrote.
Reads the room
Conditioning on audio context as well as text means the delivery can be steered toward the mood you want rather than a default register.
Four steps to your first line of audio
All of it happens in the demo above. No account, no setup.
Put your text in
Drop English text into the demo above. Write the punctuation you mean — it is what shapes the pacing and the emphasis.
Add a voice to copy (optional)
Want a particular voice? Give it a clean clip of around ten seconds and it will read your text in that voice instead.
Set the controls, then run it
Work through whatever expressiveness and length controls the Space exposes and hit generate. The first run is the slow one while the GPU wakes up.
Listen, download, or take it in-house
Play it back and download it. For anything in production, run the open MisoTTS weights on your own infrastructure and the data stops travelling.
Where an emotive voice earns its place
Products where flat narration is the thing holding the experience back.
Narration and voiceover
Explainers, ads and social video that need a voice with some life in it, without a studio booking in the calendar.
Agents people can stand talking to
Assistants, IVR trees and support bots stop sounding like machinery when the response arrives quickly and with intonation.
Long-form: audiobooks and courses
Hours of script rendered at a steady pace and tone — the consistency matters more than any single sentence does.
Characters and game dialogue
Give each character a distinct voice via one-shot cloning, whether you are prototyping a scene or shipping it.
Accessibility
Articles, documents and interfaces read aloud in something a person can listen to for an hour without fatigue.
Podcast and content production
Intros, ad reads and recurring segments — anywhere the same voice needs to turn up week after week.
Getting more out of it
- Punctuate for the ear, not the page: commas, full stops and line breaks are what Miso One reads pacing from.
- Cloning works best from a clean mono clip of roughly ten seconds with as little room noise as you can manage.
- Break a long script at natural sentence boundaries and generate section by section rather than in one pass.
- That first slow generation is the ZeroGPU Space starting up; everything after it is quicker.
- Keep confidential text out of the public demo entirely — if the work is sensitive, run the open weights yourself.
Questions people ask about Miso One
The model, the open-source release, and the company behind both.
So what is Miso One?
An open-weights TTS model from Miso Labs, built around expressive English delivery, able to clone a voice from a short reference clip. The panel above is Miso Labs' own MisoTTS demo, running on Hugging Face.
Is Miso One the same as MisoTTS?
Same model, two names: “Miso One” is what it is called as a product, “MisoTTS” is the open-source release. GitHub, the Hugging Face model page and the Space all use the latter.
Who created Miso One?
Miso Labs, a Y Combinator-backed startup working on emotive voice AI. The open-weights release came in early June 2026.
Is Miso One free and open source?
The demo costs nothing to use, and the weights are published under a modified MIT licence. Read the exact terms in the repository before you build anything commercial on it — “modified” is doing work in that sentence.
What languages does Miso One support?
English, and only English so far. Nothing else is documented as supported at launch.
Can Miso One clone a voice?
Yes — Miso Labs describes one-shot cloning from about ten seconds of reference audio. Restrict that to voices you have permission to use, and check the law and platform rules where you are.
How large is the model and how fast is it?
Around 8 billion parameters — a Llama 3.2-style backbone plus an audio decoder. On speed, Miso Labs cites first-token latency of roughly 110 ms; treat that as the vendor's own figure until someone measures it independently.
Does my data stay private?
It can, if you self-host: open weights mean the audio and text never have to leave your hardware. The demo embedded above is hosted on Hugging Face, so treat it as public and keep confidential material out.
Go and listen to it
One line of text is enough to tell whether the emotion is real. If it convinces you, the MisoTTS weights are there to build on.