What is NeuTTS (Air, Nano, 2E)?
NeuTTS is a family of small text-to-speech models from Neuphonic, built to run on the device rather than in the cloud. NeuTTS Air, released in October 2025, has about 360 million active parameters (552 million with embeddings) and speaks English. NeuTTS Nano, at about 120 million active, comes in English, French, German and Spanish versions. NeuTTS-2E, posted in July 2026, adds six emotions with four fixed English voices and no cloning.
Each model pairs a small language model with Neuphonic's NeuCodec audio codec, and ships as Q8 and Q4 GGUF files for llama.cpp. Air and Nano clone a voice from about 3 seconds of clean audio. Every output carries the Perth watermark by default. Air is Apache 2.0; Nano and 2E use the NeuTTS Open License, which is free for commercial use only below $5 million in annual revenue.
Is NeuTTS (Air, Nano, 2E) worth it?
NeuTTS is for the case where speech must stay on the device: a toy, a kiosk, a privacy-sensitive app. Neuphonic's figures put the Q4 Nano model at 45 tokens a second on a Galaxy A25 phone and 221 on a Ryzen 9 laptop CPU, with the caveat that these count only the language model, not the codec that makes the audio. Cloning from about 3 seconds of audio puts it ahead of Kokoro and Phonon, which ship fixed voices only.
The quality cost is visible in public comparisons. Gradium's blind test of on-device models scored NeuTTS 2,034 Elo against 2,043 for Kokoro and 2,098 for Phonon 31M, with an 8.7% word error rate on hard sentences against 3.5% and 1.5%. Neuphonic's hosted API, which is a separate product, sits 91st of 104 on Artificial Analysis's arena at 942 Elo.
On safety it does better than most open models. The Perth watermark is applied to every clip by default, as in Chatterbox. One trap: installing from source with uv sync turns the watermark off because its dependencies fail to install, while pip install keeps it on. Consent is not covered beyond a one-line request not to misuse the model.
The catch: licences differ inside the family. Only Air is Apache 2.0; the multilingual Nano models and the emotional 2E fall under the NeuTTS Open License, which needs a paid licence at $5 million in revenue, counted across affiliates. The 2,048-token context covers about 30 seconds of audio including the reference prompt, so long text must be chunked.
Our take: NeuTTS earns its place in hardware and offline apps where a cloned voice must run without a network; for anything that can call a server, larger models sound better for the money.
Where it earns its place
A toy maker building a talking plush with a parent's recorded voice
Speak bedtime stories in that voice without sending audio to a server
Record 3 to 15 seconds of the parent, pre-encode it once, and run Air's Q4 GGUF on the toy's processor. Air's Apache 2.0 licence avoids the revenue threshold; the watermark marks every clip.
A healthcare app team with strict data rules
Read patient instructions aloud in French, German or Spanish on the device
Run the matching Nano model with a same-language reference clip, so no text leaves the phone. A company under $5 million revenue can ship it free; above that, budget a Neuphonic licence.
A kiosk developer on a Raspberry Pi
Give an information screen a natural English voice with some emotion
NeuTTS-2E offers four fixed voices and seven emotion settings including neutral, selected with one argument; use the ONNX decoder and GGUF backbone to cut latency.
Before you commit
- Setup
- pip install neutts (add the llama extra for GGUF, the onnx extra for the ONNX decoder); reference clips should be 3 to 15 seconds, mono, clean .wav. Streaming works only with GGUF models.
- Integrations
- Python package, llama.cpp via llama-cpp-python, ONNX decoders, Hugging Face Spaces demos; Neuphonic also sells a separate hosted TTS API.
- Data and privacy
- Runs locally, so text and voice clips stay on the device. Perth watermark on by default with pip installs. No consent check beyond 'Don't use this model to do bad things'.
- Cost at scale
- Air is free at any scale. Nano and 2E are free below $5 million annual revenue across affiliates; the paid licence price is not published.
- Limits
- 2,048-token context, about 30 seconds of audio including the prompt; English only for Air and 2E; benchmark speeds exclude the codec step; output is random unless a seed is fixed.
What’s good
- Runs on phones, laptops and Raspberry Pi via GGUF
- Clones a voice from about 3 seconds of audio
- Perth watermark on every clip by default
- Air under Apache 2.0
What’s not
- 8.7% hard-sentence error rate in Gradium's test, against 1.5% for Phonon
- Nano and 2E need a paid licence at $5 million revenue
- Context of 2,048 tokens, about 30 seconds of audio
- Installing with uv sync silently drops the watermark
Best for
Hardware and app makers who need voice cloning that runs offline on phones or embedded boards, with provenance marking built in.
NeuTTS (Air, Nano, 2E) alternatives
NeuTTS (Air, Nano, 2E): common questions
- Is NeuTTS (Air, Nano, 2E) worth it?
- NeuTTS gets voice cloning onto a phone or a Raspberry Pi with a watermark attached, at the price of sounding like a small model. We rate it 3.6 out of 5.
- How much does NeuTTS (Air, Nano, 2E) cost?
- Air is free under Apache 2.0. Nano and 2E are free for research and for companies under $5 million annual revenue; above that a paid licence is needed.
- What are the downsides of NeuTTS (Air, Nano, 2E)?
- 8.7% hard-sentence error rate in Gradium's test, against 1.5% for Phonon; nano and 2E need a paid licence at $5 million revenue; context of 2,048 tokens, about 30 seconds of audio; installing with uv sync silently drops the watermark.
- Who is NeuTTS (Air, Nano, 2E) best for?
- Hardware and app makers who need voice cloning that runs offline on phones or embedded boards, with provenance marking built in.
- How much does NeuTTS (Air, Nano, 2E) cost for a team or in production?
- Air is free at any scale. Nano and 2E are free below $5 million annual revenue across affiliates; the paid licence price is not published.
- Is NeuTTS (Air, Nano, 2E) safe for company data?
- Runs locally, so text and voice clips stay on the device. Perth watermark on by default with pip installs. No consent check beyond 'Don't use this model to do bad things'.
- What are the best NeuTTS (Air, Nano, 2E) alternatives?
- The closest alternatives we have reviewed are Gradium Phonon, Kokoro-82M, Chatterbox and KittenTTS 2.
Answers checked Nov 7, 2026.