Underdog Woof

Underdog Woof: Small Text-to-Speech That Runs on One CPU Thread

Underdog Woof is an English text-to-speech model from Underdog, sized for a laptop or a browser: one voice in 31.7 MB, or two voices (Juno and Mira) in 40.6 MB. It reads phone numbers, dates, prices and codes the way people say them and runs on a CPU with numpy alone, with no GPU, server or API key.

Key Features

• Two sizes: a one-voice model of 30.7M parameters (31.7 MB) and a two-voice model of 38.2M parameters (40.6 MB), both in int8.

• Runs in the browser: the one-voice model runs on the Woof page and keeps working with Wi-Fi off once loaded; typed text stays on the device.

• Numbers said right: a built-in rule-based normalizer rewrites phone numbers, dates, times, money, codes, units and URLs into spoken form before synthesis.

• Lexicon respelling: hard names are respelled from a lexicon; the normalizer can be switched off to supply your own text processing.

• Self-checking generation: long text is split into sentences, and any piece that runs too long or contains a long silence is generated again, up to three more times.

• CPU-only runtime: the two-voice model ships with a numpy runtime and a Python API and command line; output is 24 kHz mono audio.

• Reproducible output: passing a seed gives the same audio for the same text every time.

• Published scores: Underdog reports 1.42% word error rate on Gradium's 500-sentence hard-cases set (Mira voice) and 13.8x real time on one Apple M5 Max CPU thread (Juno voice).

Use Cases

• Voice assistants and phone agents: reading booking confirmations, order numbers, flight details and other structured data aloud.

• On-device apps: speech output inside an app with no server in the loop.

• Research and teaching: a small open-weights model for non-commercial experiments with local speech synthesis.

Licensing

The weights on Hugging Face ship under the Underdog Non-Commercial License 1.0: free for personal projects, research, teaching and non-profits with credit ("Voice by Underdog Woof"). Commercial use needs a paid licence from Underdog, and the model and its audio may not be used to train other speech or audio models.

Getting Started

• Website: https://underdog.ai/woof

• Model weights: https://huggingface.co/ConwayResearch/Underdog-Woof

Underdog Woof gives developers a compact English voice that handles numbers and codes and runs locally on a single CPU thread, for non-commercial use or under a commercial licence.