Underdog Woof puts a small speech model on the CPU—with commercial limits

Underdog’s Woof is a small English text-to-speech model aimed at running on a CPU or in a browser. Its model card distinguishes a 31.7 MB one-voice download from a 40.6 MB two-voice version. Those are download sizes, not claims that the complete application uses only that much memory.

Small files, different measurements

The publisher reports its two-voice model and decoder total about 38.24 million parameters, and reports 402 MB peak memory for one tested voice. It also reports faster-than-real-time generation on one Apple M5 Max CPU thread. Those figures are device-specific publisher measurements; we have not repeated them or tested the model on a Windows PC.

Woof’s model card compares transcription errors against other speech systems. Those results depend on the chosen benchmark, voices and automated evaluation method. They do not establish a universal ranking for naturalness, pronunciation, accents or every use case. TechCrunch reported on the project’s browser demonstration on November 6.

Important terms for creators

The card labels its terms Underdog Non-Commercial License 1.0. It says noncommercial use requires attribution and that commercial uses, including monetized videos and company work, require a separate license. The model and its generated audio also have training restrictions. Check the current license directly before using it in a product, channel or article.

For our part, we like the idea of fast speech generation on hardware people already own, but a compact model is not the same thing as unrestricted open source. Voice quality, accessibility and permission to use the output are all part of the decision.