Miso One

Miso One sits around Miso Labs' Miso TTS 8B release, an open-weights text-to-speech model of roughly 8 billion parameters aimed at expressive, conversational English speech rather than broad multilingual narration. The architecture follows the conversational speech model direction associated with Sesame's CSM and uses discrete audio-code modelling rather than a simple waveform export, with the project citing a published low-latency figure of around 110 ms for voice-agent exploration.

Because it is an open model, the site's documentation layer matters as much as the demo. The public weights and inference code are available for developers who want to inspect the repository, download the checkpoint and benchmark memory and latency in their own CUDA environment; the site is candid that an 8B checkpoint is not a lightweight browser toy and needs real GPU resources. It also states plainly that the current public release is English-only, that prompt-audio generation should be evaluated with consented audio, and that safety notes and watermarking guidance should be reviewed before publishing generated speech.

For people who would rather not install anything, the site offers a hosted voice studio with a set of built-in voices, a script-to-audio workflow, previews at 48 kHz, live translation and caption tooling. Hosted usage runs on a credit system shared across text-to-speech, Voice Design and Voice Clone, with a free allowance and credit packs or plans purchasable. Note that this is an independent third-party site built around the model, not the official Miso Labs domain — always confirm licence, weights and safety terms against the official repository.