Underdog released Woof, a 31.7 MB text-to-speech model for the browser

Underdog released Woof, a text-to-speech model that fits in 31.7 MB. It runs in a web browser on the user's own GPU. The company also put out a short interactive demo at underdog.ai/escape. The page is a retro room with a typewriter, a radio and a desk lamp. Every voice in the room is made live on the visitor's machine.

The one-voice model has 30.7 million parameters. The two-voice model has 38.2 million parameters in a 40.6 MB download. Both are open weights under CC BY-NC 4.0, a non-commercial licence. The two-voice version is on Hugging Face. The browser model downloads once, then works with the Wi-Fi off. TechCrunch covered the demo on 6 November 2026.

What the vendor measured

Underdog reports a 1.42% word error rate for the two-voice model. The test was 500 sentences packed with phone numbers, dates, prices and codes. The one-voice model gets 2.06% wrong on the same test. Fish Audio Drama 3 gets 1.89% wrong, Underdog says. Drama 3 is a 41.5 GB model with 35.9 billion parameters. By Underdog's figures Woof is 1,308 times smaller than Drama 3. The two-voice model gets 25% fewer words wrong than Drama 3 at that smaller size. None of the comparisons comes from an outside benchmark. The company published them on its own Woof page.

Reading numbers out loud

Underdog says Woof runs every line through its own text normalizer before it speaks. Take a phone number like "020 7946 0018". Woof reads it as "zero two zero, seven nine four six, zero zero one eight". "14/03" is read as "the fourteenth of March". A code like "XK-4471" is read one character at a time. Small speech models often read a phone number as one long figure. Others drop part of a code. This is the case Woof was built for, the vendor says.

How it was put together

Woof's audio decoder follows the design of Mimi, by Kyutai. Its training data includes LibriTTS-R, by Yuma Koizumi and colleagues. Underdog's founder Sigil Wen told TechCrunch he is building open models that are "American-made" and "capable". He said the goal is that people have private AI. Underdog also sells a Mac app with a 163 MB speech-to-text model called Bark.

The question for a team building a voice feature is familiar. Can the speech part run inside the app on the user's own device, instead of calling a hosted API? A 31.7 MB model is small enough to travel with the download. Underdog says the one-voice model runs on a single CPU thread faster than real time. If a phone number reads out right, a lot of voice assistant work moves off the server. The catch is the CC BY-NC 4.0 licence. The browser model is free to try. The weights cannot be used in a commercial product without a separate agreement.

Source

Woof, Underdog's text-to-speech models. TechCrunch: Underdog shows off a 32MB speech model in new demo.