ConwayResearch publishes Underdog Woof, a 40.6 MB English TTS model
Original title: ConwayResearch/Underdog-Woof — new model trending #30 on Hugging Face
The model card claims 1.42% word error rate for Mira and CPU-only synthesis with numpy.
ConwayResearch released Underdog Woof, a small English text-to-speech model with Juno and Mira voices, trending on Hugging Face. The card says the two-voice package is a 40.6 MB file and runs on a laptop CPU with numpy, without a GPU or API key. Accuracy and speed figures are the publisher’s own measurements against other TTS systems.
Unconfirmed: Speed and error-rate comparisons are the publisher’s own tests, not an independent evaluation.
- The card lists Mira at 1.42% word error rate on Gradium TTS Hard Cases.
- Juno is listed at 13.8 times real time on one Apple M5 Max CPU thread.
- Peak memory is listed as 402 MB for Juno and 601 MB for Mira.
- The one-voice build is listed at 31.7 MB and 2.06% word error rate.
- Organizations
- ConwayResearch
- AI models
- Underdog Woof
From the original article
7 min read at huggingface.co[Underdog Woof beats models 1,000x its size: download size against word error rate on Gradium TTS Hard Cases, with both Woof models on the Pareto frontier ahead of Fish Audio Drama 3, Phonon, Kokoro, Pocket TTS, Supertonic 3, Kitten Nano, Qwen3-TTS and Liquid AI LFM3-Audio-2B](assets/woof-frontier. png) - **1,022x smaller than Fish Audio Drama 3** (41. 5 GB), with **25% fewer word errors** - **1.
By ConwayResearch · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.