Most expressive TTS today assumes a GPU sitting in the cloud. NeuTTS flips that assumption, treating the phone in your pocket as the deployment target rather than an afterthought. The bet underneath it: a speech model small enough to run locally — down to ~120M active params — can still clone a voice from a few seconds of audio, and that locality becomes the feature. Latency, privacy, and offline use stop being trade-offs you negotiate with an API.
What Sets It Apart
- A codec does the heavy lifting. NeuCodec, a proprietary 50Hz neural audio codec, packs enough acoustic detail that a tiny LLM backbone stays realistic — so the Air (~360M) and Nano (~120M) variants reach quality usually reserved for billion-parameter models, but in a mobile footprint.
- One family, phone to datacenter. It ships as GGUF quantizations and runs from a Galaxy A25 (~20-45 tok/s) to an RTX 4090 (~16k-19k tok/s). The same model spans embedded hardware and servers with no rewrite.
- Cloning without a pipeline. Roughly 3 seconds of reference audio is enough to clone a speaker, so there is no fine-tuning step or speaker-enrollment infrastructure to maintain.
- Multilingual, with a short memory. English, Spanish, German, and French are supported (model-dependent) inside a 2048-token context that holds about 30 seconds of audio.
Who It's For
Great fit if you are building voice features for mobile, embedded, or privacy-sensitive apps where shipping audio to a cloud TTS is a dealbreaker, or where predictable offline latency matters more than a leaderboard score. Look elsewhere if you need language coverage beyond these four, studio-grade prosody control, or the absolute top of the realism rankings — at this size there is a quality ceiling, and the ~30-second context makes long-form single-pass narration awkward.
