If your app needs to say a price or an order number, a small speech model is useful only if listeners can understand it.
Underdog released Woof, a compact English text-to-speech model, in early November. The browser demo uses a 31.7 MB download for one voice. Underdog says the demo generates audio on the device and can work offline after the model loads.
The browser demo and the two-voice model are separate
The published model card describes a 40.6 MB edition with two voices and a NumPy runtime for a laptop CPU. Those download figures do not tell you how much memory your browser will use, how long it will take to start on a phone, or what happens after the browser clears the model from storage.
The model card permits noncommercial use with credit. It directs people to obtain a separate license for commercial use. If you want to ship Woof inside an app you sell, check the terms before you build around the weights.
Listen to numbers before judging the voice
Use a short test script with a dollar amount, a date, a phone number, and an order code. Give the recording to someone who has not seen the script and ask them to write down what they hear. A natural voice is pleasant; a dropped digit is a bigger problem for a receipt or an access code.
On a phone, measure the time from opening the demo to the first playback. After the download finishes, disconnect and try a second sentence. Record those as separate results. This is a suggested test; we have not measured Woof on a phone.
Underdog publishes its own word-error comparisons on the model card. Those results describe the developer’s test setup. For another local speech option, see our guide to Phonon in Ollama.
Abdessalam Alaoui
I'm Abdessalam Alaoui, founder of Musthave.AI. I write practical guides to AI models, tools, generative media.
