# Miso One
Miso One (shipped as `MisoTTS`) is Miso Labs' 8B [[Text-to-Speech (TTS)|text-to-speech]] model, released with open weights on 3 June 2026. Miso Labs is a Y Combinator company building what they call emotive foundation voice models. The pitch is emotional range, not raw intelligibility.
## What's different
**It conditions on audio, not just text.** Most TTS takes a string and a speaker embedding. Miso One takes text *and* an audio context, so the output tracks the tone of what came before rather than reading each sentence in a vacuum. That's the whole bet: emotion is a property of the conversation, not of the sentence.
Architecture, as published:
- Residual vector quantization, 32 codebook indices over 2048-entry codebooks
- 7.7B temporal backbone plus a 300M depth decoder
- One-shot voice cloning from roughly a 10-second sample
- ~110ms latency claim
- English only at launch
Weights are on [[HuggingFace]] as `MisoLabs/MisoTTS`, code on GitHub (~3.2k stars). Licensed under a modified MIT for the weights, with training data and pipeline unreleased. Open-weight, not open-source. See [[AI Open Weight Models]].
## What I'd want before trusting it
The claims are unusually unaccompanied by evidence.
- The 110ms latency figure ships without a methodology, hardware spec, or batch size. Treat it as an orientation point.
- No published word error rate, speaker-similarity score, or MOS naturalness comparison against [[ElevenLabs]], [[Orpheus TTS]], [[Qwen3-TTS]] or anything else. For a model whose entire claim is expressiveness, the absence of a MOS number is conspicuous.
- Full-duplex and turn-taking are listed as future work, so this is a synthesis model, not a conversational stack.
- Hosted API was announced as coming, not shipped.
- It needs a high-VRAM GPU for interactive use. 8B params of audio tokens is not a laptop workload.
As of August 2026 there's no successor in the family. The GitHub org holds `MisoTTS` and a docs repo, nothing else.
## Why it's still interesting
On-premise plus one-shot cloning plus permissive weights is a combination the hosted incumbents don't offer. If you're building voice into a product where the audio can't leave your infrastructure, the shortlist is short, and this joins it alongside [[VibeVoice]] and Orpheus.
The audio-conditioning idea is the part likely to outlive this particular checkpoint. Once a model can hear the previous turn, "read this text with feeling" stops being a prompt-engineering problem.
## References
- Repository: https://github.com/MisoLabsAI/MisoTTS
- Weights: https://huggingface.co/MisoLabs/MisoTTS
- Miso Labs: https://www.misolabs.ai/blog/miso-tts-8b
- Y Combinator profile: https://www.ycombinator.com/companies/miso-labs
- MarkTechPost coverage: https://www.marktechpost.com/2026/06/04/miso-labs-releases-misotts-an-8b-emotive-text-to-speech-model-with-open-weights/
- Launch video: https://www.youtube.com/watch?v=HizlJgDbac8
- Elvis Saravia's post: https://x.com/omarsar0/status/2062227352416067984
## Related
- [[Text-to-Speech (TTS)]]
- [[Voice Cloning]]
- [[VibeVoice]]
- [[Orpheus TTS]]
- [[Qwen3-TTS]]
- [[LuxTTS]]
- [[Gemini 3.1 Flash TTS]]
- [[ElevenLabs]]
- [[AI Open Weight Models]]
- [[HuggingFace]]
- [[Y Combinator]]
- [[Knowii Voice AI]]