NeuTTS Air

Most high-quality text-to-speech with voice cloning lives behind a cloud API, which means every line of text and every reference voice leaves the device and travels to someone else's servers. That is a problem for embedded products, privacy-sensitive applications, and anywhere connectivity is unreliable. NeuTTS Air, released by the UK startup Neuphonic, is an attempt to close that gap. It is an on-device TTS model built on a small LLM backbone with roughly 360 million active parameters, paired with a neural audio codec. It speaks English and can clone a voice from a few seconds of reference audio without any training step, and it is shipped in formats meant to run on phones, laptops, and single-board computers rather than data-center GPUs.

NeuTTS Air treats speech generation as a language-modeling problem. A compact LLM backbone, in the 0.5B-parameter class, predicts tokens that a 50Hz neural codec called NeuCodec turns back into audio at low bitrate. The context window of around 2048 tokens covers roughly 30 seconds of speech. What makes it practical at the edge is the packaging: the weights ship in GGUF format with Q4 and Q8 quantizations, so they run through llama.cpp on commodity hardware, with an optional ONNX codec decoder to trim latency. Voice cloning is reference-based rather than trained: you supply 3 to 15 seconds of clean mono audio plus its transcript, and the model conditions on that style. By default, outputs carry a Perth perceptual watermark to support detection of synthetic speech.

The natural audience is anyone building voice into a product where shipping audio to an API is awkward or unacceptable. That includes embedded voice agents, talking toys, accessibility tools that read text aloud, kiosks, and assistants on Android or Raspberry Pi class hardware. The project reports real-time-capable throughput on mid-range devices, including roughly 20 tokens per second on a Galaxy A25 with Q4 and much higher figures on desktop CPUs and GPUs, so interactive use on a phone is plausible rather than aspirational. It also fits compliance-sensitive settings, healthcare, legal, internal tooling, where keeping voice data on-device simplifies the data-protection story. For these cases the combination of local inference and instant cloning at this weight class is genuinely uncommon.

The honest caveats start with language and licensing. The Apache-2.0 Air weights are English-only; if you need Spanish, German, or French you move to the NeuTTS-Nano variants, which carry the NeuTTS Open License v1.0, free for research and limited commercial use but requiring a paid arrangement above a revenue threshold. So multilingual and fully open do not currently overlap here. Quality is bounded by the small backbone: Neuphonic itself notes that prosody and naturalness do not match larger server-based models, and there are no independent, peer-reviewed benchmarks to lean on yet. The cloning ability also carries obvious misuse risk, which the watermark mitigates but does not eliminate. Finally, this is a young startup, and impersonator websites with similar names have appeared, so source verification matters.

Measured against ElevenLabs, the obvious commercial alternative, NeuTTS Air is not trying to win on raw fidelity or breadth of voices and languages. ElevenLabs will give you a higher quality ceiling, more languages, and a managed service with no hardware to think about. NeuTTS Air competes on a different axis: it runs offline, keeps audio on the device, and the Air weights are truly open for commercial use. Adopt it when edge deployment, privacy, or per-request cost rule out a cloud API and English is enough. Hold off if you need multilingual output under a fully permissive license, studio-grade prosody, or verified quality numbers before committing. For the right embedded or compliance-driven use case, though, it is a credible open foundation worth prototyping with.