neuphonic

Following our #1 trending NeuTTS Air release in October, we’re pushing the limits of "small" even further. We’ve just released NeuTTS Nano, an ultra-lightweight speech foundation model under Apache 2.0.

The main idea: Frontier-quality text-to-speech with a tiny 120M active parameter footprint - 3x smaller than our previous model.

Why we built this: - Extreme Edge Deployment: At this size, real-time TTS is now viable on the most resource-constrained hardware, including older Raspberry Pis and mobile devices, without needing a local GPU; - Llama3 Architecture: Built off a Llama3-based LM + codec architecture, it maintains high-quality prosody and realism despite the 3x size reduction; - Privacy & Control: Full on-device inference means sensitive data never leaves the hardware, with zero API costs or rate limits; - Instant Cloning: Supports zero-shot voice cloning with as little as 3 seconds of audio.

Repo: https://github.com/neuphonic/neutts HF: https://huggingface.co/neuphonic/neutts-nano

We’d love feedback from the HN community on performance benchmarks across different CPU architectures and edge devices.

TTS models can be deployed on most devices (we do a lot of CPU work), but the real bottleneck for these on-device deployments is on the LLM side. Even a "small" 700m parameter LLM was problematic. Also running 3 models in parallel on a single device is not super straightforward.

We’ve been playing with what's truly possible for low-latency, privacy-first voice agents, and just released a demo: Agent Santa.

The entire voice-to-text-to-speech loop runs locally on a sub-$250 Nvidia Jetson Orin Nano.

The ML Stack:

- STT: OpenAI Whisper EN tiny - LLM: LiquidAI’s 700M-parameter LFM2 - TTS: Our NeuTTS (zero-cost cloning, high quality)

The whole thing consumes under 4GB RAM and 2GB VRAM. This showcases that complex, multi-model AI can be fully deployed on edge devices today.

We'd love to hear your feedback on the latency and potential applications for this level of extreme on-device efficiency.

Piper is a VAE model which is quite robotic. This is a speech language model, which sound quite realistic.

You can listen to the model on this video => https://www.youtube.com/watch?v=YAB3hCtu5wE

The codec is open source: https://huggingface.co/neuphonic/neucodec

We’ve just released Neuphonic TTS Air, a lightweight open-source speech foundation model under Apache 2.0.

The main idea: frontier-quality text-to-speech, but small enough to run in realtime on CPU. No GPUs, no cloud APIs, no rate limits.

Why we built this: - Most speech models today live behind paid APIs → privacy tradeoffs, recurring costs, and external dependencies. - With Air, you get full control, privacy, and zero marginal cost. - It enables new use cases where running speech models on-device matters (edge compute, accessibility tools, offline apps).

Repo: https://github.com/neuphonic/neutts-air

Would love feedback from HN on performance, applications, and contributions.