What Is Miso One? Inside Miso Labs' 8B Open-Weights Emotive TTS Model
A practical guide to Miso One and Miso TTS 8B: how the 8B open-weights voice model works, its 110ms latency claim, one-shot voice cloning, local setup, real use cases, and honest limitations.
Miso One has turned into one of the most searched names in open voice AI, and the reason is straightforward. It puts an expressive, 8-billion-parameter text-to-speech model into the open-weights world, where developers can download the checkpoint, read the code, and run it on their own hardware instead of calling a closed API. For anyone watching speech AI move from hosted-only services toward self-hostable models, Miso One is a release worth understanding in detail.
The model behind the name is Miso TTS 8B, built by Miso Labs and published with open weights in early June 2026. It targets emotionally expressive English speech, voice continuation from a short audio prompt, and low-latency generation aimed at voice agents and research. If you want the official overview, the demo, and the model facts in one place, the Miso One page is the most direct starting point.
This guide explains what Miso One actually is, how Miso TTS 8B works underneath, what it can and cannot do today, and how to evaluate it for your own project — separating the confirmed facts from the launch-day marketing.
Miso One is the product-facing name for Miso TTS 8B, the open-weights text-to-speech model from Miso Labs. It helps to keep the two terms apart. "Miso TTS 8B" is the model itself: the 8-billion-parameter checkpoint, the inference code, and the published weights. "Miso One" is the public identity around that model — the overview page, the hosted demo, and the model facts that most people reach first when they search for it.
That framing matters because nearly every search for Miso One is model-intent. People are not shopping for a generic narrator app. They want the official files, the Hugging Face weights, the GitHub repository, a demo they can listen to, and a clear answer to one question: is this model good enough, and small enough to serve, for what they are building? The Miso One overview is organized around exactly those questions.
Three facts define the release. It is open weights, so the checkpoint and inference code are public rather than locked behind a paid endpoint. It is English-first, focused on quality and expressiveness in English rather than broad multilingual coverage. And it is large for a speech model at 8 billion parameters, which means it is built for real GPU inference, not a lightweight in-browser toy.
Miso TTS 8B follows the architecture direction popularized by Sesame's Conversational Speech Model (CSM). In practice, that means it is not one monolithic network but a pairing: a large backbone based on a Llama 3.2-style design, plus a smaller autoregressive audio decoder that turns the backbone's predictions into sound.
Rather than generating a raw waveform directly, the model produces discrete audio codes — Mimi audio codes, from the same residual-codec family used across CSM-style systems. To capture fine detail, it relies on residual vector quantization (RVQ): each audio token is represented as 32 codebook indices drawn from 2048-way codebooks. Each codebook adds another layer of refinement, and the decoder sums those layers back into audio. This layered approach is a big part of why the model can carry emotional texture and prosody that simpler vocoders tend to flatten.
The other defining trait is context conditioning. Miso TTS 8B can take prior audio as input, not just text. That audio prompt shapes speaker tone, pacing, and identity in the generated speech, which is the mechanism behind its one-shot voice cloning and voice-continuation behavior. For the canonical technical details, the model card on Hugging Face (MisoLabs/MisoTTS) and the MisoLabsAI/MisoTTS GitHub repository are the authoritative sources, and the Miso One model page pulls the most important facts together for a faster read.
The word that follows Miso One everywhere is "emotive." Miso Labs positions the model around expressive, human-sounding delivery — warmth, rhythm, natural pauses, and emotional shading instead of flat narration. Some early coverage framed it more dramatically, describing an 8B open model built to "out-emote" humans. That is headline language, and the honest approach is to treat the strongest emotion claims as something to test with your own scripts rather than accept on faith.
The underlying ambition is real, though. Expressive delivery is the hardest part of text-to-speech. Most models can produce intelligible audio; far fewer can vary emotion convincingly across a long passage, land a pause, or keep a consistent character through a paragraph. Miso TTS 8B is explicitly tuned for that conversational, emotionally varied register, which is why so many evaluators compare it against other recent speech models rather than against basic narrators.
The first capability is expressive English speech. The current public model concentrates on emotion, pacing, and conversational delivery in English. If your project needs that lifelike, human-reading quality in English, this is the model's core strength.
The second is one-shot voice cloning. Because Miso TTS 8B conditions on prompt audio, it can continue or imitate a speaker's tone from a short reference sample. Evaluators usually test this carefully: short prompts, noisy prompts, longer continuations, and emotional shifts, all with consented audio only.
The third is low latency. Miso Labs publishes a roughly 110ms latency figure, which is the number that draws voice-agent builders to the model. Treat it as an orientation point rather than a guarantee. Real latency depends on your GPU, batching, serving code, and prompt length, so the published claim is a reason to benchmark, not a substitute for benchmarking.
The fourth is open access and self-hosting. The weights and inference code are public under a modified MIT license, so teams that need control over inference, data handling, or on-premises deployment can run the model in their own environment instead of routing audio through a third-party service.
There are three practical paths, and which one you start with depends on how deep you need to go.
If you just want to listen, start with the hosted demo and the overview on miso-one.com. A quick listen tells you whether the voice quality, emotional range, and English pronunciation are in the right zone before you invest engineering time.
If you want to inspect the model, go to the Hugging Face page (MisoLabs/MisoTTS) for the model card, license, and weights, and to the MisoLabsAI/MisoTTS GitHub repository for the inference code and setup instructions. This is the right step for confirming checkpoint status, safety notes, and hardware requirements before downloading anything.
If you want to run it locally, plan for real infrastructure. At 8 billion parameters, Miso TTS 8B expects a capable GPU and a working CUDA environment. Install the repository, download the weights, and benchmark latency and memory use against your own target hardware and response lengths. Miso Labs has also signaled that hosted API access is coming, so if self-hosting is not practical for you yet, that is a path to watch.
Voice agents are the headline use case. The combination of expressive delivery and a low-latency design is aimed squarely at interactive, conversational systems where a flat or laggy voice breaks the experience. The honest caveat is that "agent-ready" depends on your serving stack, so the model is a strong candidate to benchmark rather than a finished agent.
Expressive narration and creator audio are a natural fit too. Where a project needs warm, character-rich English reading — audio stories, explainer videos, character voices, longer-form narration — the model's emotional range is the point. Open weights also mean a studio or tool can build the voice layer directly into its own pipeline.
Research and evaluation are a major reason the model exists in open form. Speech researchers can probe prosody, stability, and how an 8B CSM-style model compares with other recent systems. Because the weights are inspectable, the model is useful for studying the architecture itself, not just its outputs.
Finally, on-premises and privacy-sensitive workflows benefit from self-hosting. Teams that cannot send audio to an external API — for compliance, data-residency, or confidentiality reasons — can keep generation inside their own environment.
Being specific about limits is the fastest way to avoid disappointment, so it is worth stating them plainly.
Miso One is English-only today. It should not be described as a broad multilingual product. If your roadmap depends on many languages, the current public release is not the right fit, and that is exactly the kind of trade-off the Miso One model facts are meant to make clear.
It is also heavy. An 8B checkpoint is not a casual download, and running it well takes real GPU resources. For quick, browser-based, or multilingual experiments, a lighter hosted demo will often be more practical than standing up local inference for an 8B model.
There are safety considerations as well. Generated audio is watermarked by default (Miso Labs cites SilentCipher), and prompt-audio voice cloning carries clear consent obligations. Clone only voices you have permission to use, review the official safety and watermarking guidance, and test edge cases before publishing any generated speech publicly. Because this is a newly released model, treat the model card, license, and safety notes as the source of truth and re-check them as the project evolves.
It helps to place Miso One in context. The closest reference point is Sesame's CSM, which popularized the conversational-speech architecture Miso TTS 8B builds on. Sesame open-sourced a 1B CSM variant under Apache 2.0; Miso One is a larger 8B model published under a modified MIT license, with emotion and one-shot cloning as its headline selling points. Bigger is not automatically better for every workload, but the extra scale is part of why Miso Labs leans on expressiveness as the differentiator.
Against closed services, the trade is the familiar one: hosted voice APIs are convenient and require no infrastructure, while an open-weights model like Miso One trades convenience for control, inspectability, and self-hosting. And against lightweight, multilingual, browser-first tools, the difference is intent. A multilingual in-browser demo is built for fast, low-friction generation across languages; Miso One is built for expressive English quality and local deployment that you evaluate seriously before adopting. Most teams end up choosing based on which trade-off matches their product, not on a single benchmark.
A practical path for anyone who found Miso One through search and needs to decide whether it is worth installing.
Start with the model card. Open the Hugging Face page and GitHub repo to confirm license, checkpoint status, safety notes, and hardware requirements before downloading anything.
Listen to the demo. Use the hosted demo on Miso One to judge voice quality, emotional range, and English pronunciation with your own text before committing engineering time.
Run local inference. Install the repository, download the 8B weights, and benchmark real latency and memory use in your own CUDA environment rather than relying on the published numbers.
Test prompt audio. Evaluate voice continuation with consented audio prompts, including short prompts, noisy prompts, and longer continuations, and check whether speaker identity stays stable.
Decide where it fits. Use your results to choose between self-hosting now, waiting for hosted API access, or using a different speech model for production.
Miso One is the public name for Miso TTS 8B, an 8-billion-parameter open-weights text-to-speech model from Miso Labs. It is built for expressive English speech, voice continuation from prompt audio, and low-latency generation. The overview, demo, and model facts live at miso-one.com.
The model ships with open weights and public inference code under a modified MIT license, so you can download and run it yourself. Always confirm the exact license terms on the official Hugging Face model card before redistributing or using outputs commercially.
The current public release is focused on English. It should not be treated as a broad multilingual model today, which is one of the most important facts to verify before adopting it.
Yes. The weights and inference code are public, and the model is designed for local inference. Because it is an 8B model, plan for a capable GPU and a working CUDA setup, and benchmark latency and memory on your own hardware.
It supports one-shot voice cloning and voice continuation by conditioning on prompt audio. Use consented audio only, and review the watermarking and safety guidance before publishing generated speech.
It is a powerful, newly released model rather than a turnkey product. The realistic approach is to evaluate it — voice quality, latency on your hardware, prompt-audio behavior, and safety — and start from the official files on the Miso One page, Hugging Face, and GitHub.
Miso One matters because it brings emotionally expressive English speech into the open-weights world at a serious 8B scale. The strengths are clear: expressive delivery, one-shot voice cloning, a low-latency design, and full self-hosting. So are the trade-offs: English-only for now, real GPU requirements, and safety responsibilities around watermarking and voice consent.
The right way to approach it is as an evaluation, not a leap of faith. Read the model card, listen to the demo, benchmark on your own hardware, and test prompt audio with consented samples before deciding where Miso TTS 8B belongs in your stack. If that sounds like your project, the cleanest place to begin is the official Miso One overview, which links the demo, model facts, and source files in one place.
