Underdog Woof: a 40 MB text-to-speech model for CPU without a GPU

The model ConwayResearch/Underdog-Woof has been published on HuggingFace — a compact text-to-speech generator with two voices: Juno and Mira. The author claims 1.42% word error rate on the Gradium TTS Hard Cases set, a speed of 13.8x real time on a single CPU thread, and operation without a GPU, server, or API key. The model reads phone numbers, dates, prices, and codes the way people pronounce them.

What it means

In our audio generators section, Underdog Woof stands out primarily for its hardware requirements. The model takes up 40.6 MB in the two-voice version, the weights are stored in npz format, and it runs through numpy — without PyTorch or a graphics card. Peak memory consumption on a single CPU thread is stated at 402 MB. For comparison: XTTS from Coqui requires 1.7 GB for its largest weights file and runs through the coqui library, while Fish Speech needs 1.2 GB and the CUDA platform. The Underdog Non-Commercial License 1.0 permits free use in personal projects, research, teaching, and non-profit organizations. Commercial use of the model, derivatives, or outputs requires a separate written license. Training other speech or audio generation models on this model or its outputs is prohibited, as is imitating a real person's voice without consent. When publishing audio, you must credit "Underdog Woof" or "underdog.ai". Similar restrictions apply to XTTS: its weights are also permitted for non-commercial use only, including evaluation by commercial organizations, but without direct or indirect payment. YuE stands apart in this lineup — both its code and weights are under Apache-2.0, but it is a music generation model rather than a speech one, and it needs 24 GB of VRAM. Underdog Woof is a rare case where a text-to-speech model fits entirely into a laptop's system RAM and runs on CPU in real time with headroom to spare. For local voiceover without cloud services, this is a notable shift in hardware requirements, although commercial use remains subject to a separate license.