Underdog AI has unveiled Woof, a remarkably compact text-to-speech model with a reported 31.7MB download and approximately 30.7 million parameters. Demonstrated through an interactive browser game, Woof promises locally generated speech without relying on Cloud inference. The company has also published model weights, although its quality and speed claims still require independent testing.
For creators who simply need a finished voiceover, the DIY AI voice generator provides a managed alternative. Woof addresses a different problem: making speech synthesis small enough to run directly on consumer hardware, including potentially modest devices where larger models are impractical.
TechCrunch reported on the demonstration on 6 November 2026, highlighting both speech and AI-generated music. However, Underdog has released something different. Woof is a speech synthesis model; the music showcased in the experience comes from other local models.
What has Underdog actually released?
Woof converts written text into spoken audio. Its headline achievement is fitting neural speech generation into a download considerably smaller than many existing AI voice systems.
| Specification | Underdog Woof |
|---|---|
| Primary function | Text-to-speech generation |
| Advertised browser model size | 31.7MB |
| Reported parameters | Approximately 30.7 million |
| Browser processing | Local inference using WebGPU |
| Claimed CPU performance | Up to 18 times real-time on one laptop CPU thread |
| Weights | Published by Conway Research |
| Commercial licensing | Restricted under the company’s stated non-commercial terms |
| Independent quality verification | Not established |
Conway Research’s activity on Hugging Face corroborates the model’s release. Developers can inspect the official Underdog Woof model repository for the available files and documentation.
The company’s browser experience, called Escape, offers a retro-style interactive environment where generated voices play a key role. It is an unusually accessible demonstration for a model release, allowing visitors to experience local speech generation without first assembling a Python environment.
Underdog also develops Bark, a separate speech-to-text system. The names are easy to confuse, especially because its earlier language models have also used the Woof name. Developers should identify the specific speech model repository rather than assuming all Underdog models called Woof share the same architecture or purpose.
Does Underdog’s 32MB model generate music?
The demonstration contains generated music, but there is no verified evidence that the 31.7MB Woof text-to-speech model itself composes those tracks.
Underdog describes its wider experience as featuring audio generated offline by local models. Speech synthesis and music generation involve different requirements. A text-to-speech model must turn language into intelligible speech, while a music model must produce rhythm, instrumentation, and musical structure over time.
Reporting that Woof generates both speech and music would therefore overstate its demonstrated capabilities.
This distinction also affects practical adoption. A developer building spoken notifications, reading assistance or an offline voice interface can evaluate Woof directly. Someone building a music creation application would need to investigate the separate model responsible for the musical examples, including its size, runtime and licence.
A 32MB download does not mean 32MB of memory usage
The most interesting engineering question is whether Woof’s compressed model size translates into genuinely lightweight deployment.
A 31.7MB download is attractive for browser distribution and devices with limited storage. However, model weights are only part of the runtime footprint. Applications also need memory for decoded weights, intermediate activations, audio buffers and the inference engine itself.
WebGPU introduces another variable. Browser and device support differ, while GPU memory allocation and shader initialisation can affect performance before the first spoken word is produced.
A model might generate ten seconds of speech in considerably less than ten seconds once warm, yet still feel slow if its first launch requires substantial compilation or loading.
That is why Underdog’s claim of up to 18 times real-time CPU generation should be treated as a vendor benchmark, not a guaranteed user experience. CPU throughput, WebGPU responsiveness and time to first audio are three separate measurements.
Developers working with local speech tools repeatedly encounter a similar practical problem: achieving attractive inference speed is easier than making installation, browser compatibility and voice quality consistent across different machines.
Woof’s small download size may reduce one major source of friction, but it doesn’t automatically eliminate the others.
Underdog’s accuracy claim needs more context
Underdog has promoted a 1.42% word error rate in its launch messaging. That is a potentially interesting result, but it should not be treated as proof that Woof offers the best overall speech quality.
Word error rate measures how accurately transcription recovers spoken words. For synthetic speech, this can help reveal whether numbers, abbreviations and difficult phrases remain intelligible.
It does not adequately measure warmth, emotional expression, accent authenticity, pacing or speaker consistency.
There is also a package-level verification issue. The published benchmark figure should not automatically be attributed to the exact 31.7MB browser configuration without confirming which voice, model variant and evaluation set produced it.
DIY AI’s AI tool evaluation data and audio generation dataset use broader criteria, including voice realism, latency, language coverage, licensing and integration. Woof has not been independently evaluated in that dataset, so assigning it a comparative score now would be premature.
The non-commercial licence limits where Woof can be used
Underdog’s announcement describes Woof as available with open weights under CC BY-NC 4.0. That enables experimentation, but it does not provide the unrestricted commercial deployment rights associated with permissive licences such as MIT or Apache 2.0.
This distinction matters most for developers planning to embed generated speech in a paid application, client product, or commercial service.
Before deployment, inspect the actual repository licence and associated notices. Confirm that the published terms cover the exact weights and runtime being used, and obtain separate commercial permission where required. The company’s announcement alone is not a substitute for checking the distributed files.
Local inference could reduce recurring cloud synthesis charges, but licensing and maintenance may still create costs. A tiny model is not automatically the least expensive solution once you include development time, browser support, updates, and quality assurance.
For a comparison against managed infrastructure, our breakdown of Google Cloud text-to-speech pricing explains how hosted generation is billed and why cost per accepted recording can be more useful than the underlying API rate.
How developers should test Woof before adopting it
A useful evaluation should test more than an impressive sentence in the launch demonstration. We recommend using a repeatable script across Woof and at least one established alternative.
| Test | What to check |
|---|---|
| Pronunciation | Dates, prices, telephone numbers, names and technical abbreviations |
| Naturalness | Whether pacing and emphasis remain convincing across ordinary paragraphs |
| Cold start | Time from opening the application to hearing the first generated audio |
| Warm generation | Audio duration produced per second after the model has loaded |
| Memory | Peak system and GPU memory, not merely the compressed download size |
| Offline operation | Whether speech generation continues after initial assets are cached and connectivity is disabled |
| Compatibility | Behaviour across desktop browsers, mobile devices and CPU-only environments |
| Licensing | Whether the intended personal or commercial use is actually permitted |
For pronunciation testing, use a sentence containing several difficult elements, such as: “Your £1,249.50 invoice is due on 17 December. Please quote reference ZX-407 when calling the Chelmsford office.”
Listen for altered numbers, awkward pauses and incorrectly spoken abbreviations. Then try longer passages, as a voice that sounds excellent in one sentence may struggle to stay consistent over several minutes.
For the offline test, load the required files first, disconnect the network and generate a new sentence rather than replaying previously generated audio. Record any network requests while the model runs. This tests the local-processing claim more meaningfully than simply observing that the application has an offline label.
Finally, compare the same script with larger speech models under equivalent listening conditions. A small model that is slightly less expressive but faster to load may be the better engineering choice for accessibility or device notifications. A narration platform may need the opposite balance.
Who should pay attention to Woof?
Woof is particularly interesting for developers building private reading assistants, local voice interfaces, educational applications and offline accessibility features. Those workloads can benefit from modest model downloads, predictable local processing and reduced dependence on external inference services.
It is less obviously suitable for production voiceovers that require extensive emotional control, multilingual output, reusable voice identities, or established commercial permissions. The browser demonstration does not establish all those capabilities.
For creators choosing a production service rather than integrating model weights, our comparison of the best AI voice and audio tools covers a wider range of narration, cloning, editing and music workflows.
DIY AI verdict: a meaningful local speech release, with testing still to do
Underdog Woof deserves attention because a 31.7MB speech model running locally in a browser offers a practical way to reduce the hardware and deployment burden of AI audio.
The strongest part of the announcement is the combination of a public model release and an interactive demonstration. Developers have something concrete to inspect rather than relying solely on promotional video.
However, independent voice-quality comparisons, realistic memory measurements, device compatibility testing and licensing checks are still necessary. The company’s claimed superiority over much larger models remains unproven across the full range of speech-generation tasks.
We recommend evaluating Woof first for short English speech, offline interfaces, and low-overhead applications. Treat music generation as a separate capability, and do not commit to commercial deployment until you confirm the relevant licensing terms.
