Text-to-speech runs in the browser with Woof

Woof brings text-to-speech to the browser, running directly in the web app for fast, local voice generation in real time without uploads or installs today.

The text typed on a virtual typewriter comes to life. Around it, a cathode-ray screen, a desk lamp, and a retro setting make up the game Escape from Riverbark, designed to showcase Woof, Underdog’s text-to-speech model. The demonstration runs its voices directly in the browser, thanks to WebGPU, which provides access to the device’s graphics capabilities.

What matters is this shift in computation: speech reading no longer requires sending each text to a server responsible for generating the audio. Once the page and the model are loaded, the demo can continue offline, according to its publisher. For a game or a reading application, this makes it possible to generate lines on demand on the device being used, without relying on a remote voice service. Two versions behind the launch numbers

The Woof presentation distinguishes two formats that the launch message lumps together a bit too quickly. The 31.7 MB file corresponds to a single-voice version, with about 30.7 million parameters. Its reported error rate is 2.06%.

The 1.42% result belongs to another version: a 40.6 MB model, equipped with two English voices, Juno and Mira, and about 38.2 million parameters. It is Mira that achieves this score. The distinction matters: the best published result and the smallest download do not describe the same configuration. The documented uses here concern English, with no equivalent demonstration for French. Reading the right words is not enough to rank a voice

To establish its comparisons, the team uses 500 English sentences from the Gradium TTS Hard Cases corpus. Dates, amounts, and codes test the way a written text is turned into speech. The produced audio is automatically transcribed, then compared with the expected text to count incorrect, missing, or added words.

In this evaluation carried out by Underdog, the best voice of each model is retained. The error rate indicates reading fidelity; it does not, by itself, measure naturalness, expressiveness, or listening pleasure. The ranking presented therefore is not enough to establish an overall superiority over all competing voices.

The same clarification applies to the promise of a model “1,000 times smaller”: it refers to a comparison between the 40.6 MB version and the 41.5 GB of files in Fish Audio Drama 3, according to the sizes retained by the publisher. This ratio describes the downloads being compared, not a uniform advantage over all models on the market. Downloading the parameters does not free up all uses

The parameters are available in the public model repository. Their distribution falls under open-weights models, with restrictions that matter for creators. Despite the CC BY-NC 4.0 mention on the site, the published license file bears the name “Underdog Non-Commercial License 1.0.”

This text authorizes personal use, research, and teaching under attribution conditions. Commercial use of the downloaded model and its audio outputs requires separate written authorization, including for monetized videos or advertisements. One exception authorizes commercial uses carried out free of charge within Underdog’s own sites and applications. It does not cover integration of the model into another service. A small download, but hardware requirements to distinguish

The file size does not correspond to the memory needed to produce the voice. For its single-voice version, the publisher reports a peak of 852 MB in its CPU test process, part of which comes from software libraries. This measurement does not describe the consumption of the WebGPU demonstration.

Performance on a computer does not guarantee identical execution on a phone or watch either, two platforms mentioned in the announcement. The browser must support the technology used, and the device must have the necessary resources. Offline operation, for its part, requires having previously downloaded the page and the model files.