This 32MB AI Voice Model Runs in Your Browser. Could It Replace Cloud-Based Text-to-Speech?

Updated November 6, 2026. Woof is only days old, so some of what follows may shift as more people get their hands on it.

On November 5, a startup called Underdog posted a small browser game: a retro room with a typewriter, a glowing CRT monitor and a desk lamp you can click. Type something on the keyboard and the room says it back to you. It's charming, a little silly, and according to TechCrunch it exists to show off a new text-to-speech model called Woof, which the company says can run on most devices with no internet connection at all.

The number that got everyone's attention is the size. Underdog says the version running in your browser is 31.7 MB. For anyone who has watched a cloud speech bill creep up, or felt uneasy about shipping user text off to someone else's servers, that's a tempting thing to read.

It's also a number worth being a bit suspicious of. Woof is days old. Most of the detailed figures come from the company itself, and I haven't benchmarked it on my own hardware. So this isn't a victory lap. It's an attempt to sort what the documentation actually backs up from what's still marketing, and then to answer the question underneath it all: could something this small really stand in for a cloud speech service?

What Woof actually is

Strip away the game and Woof is a small English text-to-speech model with two voices, Juno and Mira. It comes in two sizes. The one-voice model is the one that runs in the browser, at 31.7 MB. The two-voice model published on Hugging Face weighs 40.6 MB. Both generate 24 kHz mono audio, and the decoder that turns the model's output into sound is adapted from Kyutai's open Mimi codec.

Two things in the public material don't quite add up, and I'd rather point them out than smooth them over.

The first is the parameter count. TechCrunch reports 31 million parameters. The model card breaks the two-voice model into about 23 million parameters for the speech model and 15.3 million for the decoder, roughly 38 million in total. The 31 million figure might describe the one-voice build, but nothing I read confirms that.

The second is music. TechCrunch describes a tuner in the game that generates different kinds of music. The model card, though, talks about text-to-speech and nothing else. Until Underdog explains it, I'd treat the music as a feature of the demo, not a documented ability of the model.

As for architecture, the public description is thin. There's no paper-style breakdown. What the card does reveal is the machinery around the model, and that's more telling than you'd expect. A rule-based normalizer rewrites numbers, dates, prices and codes the way a person would say them. Long text gets chopped into sentences. And if a piece comes out far too long, or contains a second or more of dead air, Woof simply tries again, up to three more times, and keeps the best take.

I find that detail oddly reassuring and a little sobering at once. It tells you this is a generative model that occasionally trips over itself, and that some of its polish comes from careful engineering wrapped around the weights, not from the weights alone. The card also mentions a default of 16 steps, which suggests an iterative generation process. That's my guess, though, not something Underdog says.

The benchmark, read slowly

Here's the claim that's driving the buzz. On the Gradium "TTS Hard Cases" set, 500 English sentences, the two-voice model with the Mira voice gets 1.42% of words wrong. The one-voice browser model gets 2.06%. In Underdog's own table, that puts Woof ahead of a long list of cloud services: Microsoft's MAI-Voice-2.1 at 2.64%, Mistral's Voxtral 3 TTS at 2.91%, ElevenLabs Flash v2.5 at 6.63%, and the open Kokoro model at 5.17%. Fish Audio Drama 3, a 41.5 GB open-weights model, lands at 1.89%. That last comparison is where the "1,000 times smaller" line comes from.

It's an impressive table. It also deserves four honest caveats.

First, Underdog says it ran and scored every model itself. Nobody outside the company has replicated the results yet.

Second, the test set is packed with phone numbers, dates and codes. Those are exactly the sentences that reward a model with a purpose-built text normalizer, so Woof was always going to look good here.

Third, the metric itself. Word errors are counted by having other AI models transcribe the audio and compare it to the script. That tells you the speech is accurate and understandable. It tells you nothing about whether the voice sounds warm, whether the pacing feels natural, or whether you could listen to it for twenty minutes without wincing. I found no listening-test score anywhere in the published material.

Fourth, one voice per model in one test isn't the same as every real-world use case.

So the fair reading is this: Woof looks unusually good at saying difficult text correctly, for its size. Whether people enjoy hearing it is a separate question, and a word-error table can't answer it. Other coverage makes the same point; FourWeekMBA notes that the benchmarks are Underdog's own measurements.

Hardware, speed and the browser question

On speed, the numbers are strong. The one-voice model runs at 18 times real time on a single CPU thread of an Apple M5 Max, using plain PyTorch with int8 weights. The two-voice model manages 13.8 times real time with the Juno voice in a numpy runtime.

But an M5 Max is a top-tier laptop chip. It's not what most people carry around. Nothing I found reports results from a mid-range phone or a five-year-old laptop, and that's precisely where a claim like "runs on most devices" gets tested.

Memory is the less-advertised cost. The one-voice model peaks at 852 MB of memory, compared with 402 MB for the two-voice Juno model. So the smaller download is the hungrier one, although the two numbers come from different runtimes, and neither is a measurement of a browser tab.

Browser compatibility is the part I couldn't verify at all. Underdog says "most devices," and I couldn't open the demo's documentation page to see which browsers work or whether WebGPU is required. Similar projects vary a lot here. Liquid AI's in-browser audio demo requires WebGPU in a recent version of Chrome or Edge, while Kokoro.js started out running on WebAssembly. I wouldn't assume Woof is either until Underdog says so.

Offline use looks more solid. The model card says Woof runs on a laptop CPU with nothing but numpy, from a single file, with no GPU, server or API key. That supports the idea of running locally once you've downloaded it. What I can't tell you is whether the browser demo caches its 31.7 MB so it works offline on your next visit.

Browser speech isn't new. The size is.

It's worth remembering that Underdog didn't invent speech in a tab. Kokoro.js already runs the 82-million-parameter Kokoro model entirely in the browser, and tools like TTSLab let you try and compare speech models locally with WebGPU and WebAssembly.

What Woof is trying to change is the price of admission. The model card lists Kokoro at 327 MB against 40.6 MB for Woof's two voices, a gap that FourWeekMBA puts at roughly eight times. A download that much smaller could mean fewer people giving up on a slow connection. That's my reasoning, though, not something anyone has tested.

Cloud versus local: where each one wins

Privacy is the strongest card local inference holds. If the text never leaves the device, there's no third-party server to breach and no retention policy to read. But that's a property of the design, not a guarantee. The rest of the app can still send logs or analytics somewhere, so anything regulated needs a review of the whole system, not just the model.

Cost is trickier than "free versus paid." Cloud speech is generally billed by usage, and I didn't compare current price lists, so check the vendors directly. Local inference shifts the cost to the user's device and to your bandwidth. Sending roughly 32 MB to every new visitor shows up on a CDN bill and in a first-load wait. And for businesses, Woof isn't free at all, which brings us to the license.

Quality and range still favor the cloud. Hosted services typically offer many voices, many languages and expressive controls. Woof offers English and two voices. If your product needs Spanish or Japanese, the conversation ends right there.

Reliability cuts both ways. A local model works on a plane, with no rate limits and no outages. But you're now shipping to hardware you can't see, so a slow phone gives some people a noticeably worse experience than others. A cloud service gives you consistent output, monitoring, and the ability to fix things centrally.

Scalability is a similar trade. Local inference scales with your users' devices at no per-request cost to you, but updating a model means getting everyone to download it again, and your support surface grows with every device type.

The license is the real story for founders

TechCrunch calls the model open for anyone to download. The weights are downloadable, yes. But what governs your use is the Underdog Non-Commercial License 1.0. It's free for personal projects, research, teaching and non-profits, as long as you give credit. Using Woof or its audio in a paid product, an ad, a monetized video, a company's work, or anything else that earns money needs a commercial license from Underdog.

There are two more clauses founders should notice. You can't use Woof or its audio to train, fine-tune or distill another speech or audio model. And anything built with it has to show "Voice by Underdog Woof."

In plain terms, that's open weights, not open source in the usual sense, since a non-commercial restriction doesn't meet the Open Source Definition. For a startup, embedding Woof means starting a conversation with Underdog. The card points to a licensing email address, and I found no public pricing. The "no distillation" rule matters too, because it closes off the tempting shortcut of using Woof to generate training data for your own voice model.

Non-commercial terms also have a reputation for fuzziness. Creative Commons' own NC licenses have long been criticized for being vague about what counts as commercial. Woof's wording is more explicit than most, but if you run ads on a site, voice-over on those pages very plausibly counts. Read the license, and ask Underdog before assuming otherwise.

Where it fits, and where it doesn't

The comfortable fits are personal and educational projects, research prototypes, accessibility experiments, offline kiosks and games, and privacy-first apps. That matches how Underdog describes its own goals: the founder says he wanted models that run on users' devices so people own their data, and the company also released Bark, a 167 MB speech-to-text model. Put the two together and you can imagine a fully local voice loop. TechCrunch's own take is modest: Woof probably won't replace bigger models, but it could power local assistants or personalized podcasts.

The poor fits are easy to list. Multilingual products. Anything that needs many voices or voice cloning, which the card doesn't document. High-volume phone systems. Expressive long-form narration where flatness would stand out. And anything that needs a service-level agreement.

As for what developers are saying, I found no substantial Woof-specific discussion yet; it's simply too new. Older Hacker News threads do show the long-running complaint that open-source speech lags commercial neural voices, alongside the counterpoint that lightweight local engines are good enough for plenty of jobs. Both predate Woof, so read them as background, not evidence about it.

The verdict

Woof won't replace cloud text-to-speech across the board. As a complement, though, it looks credible for a narrow but real slice of work: English, short utterances, accuracy-sensitive text like order numbers and dates, modest volume, and situations where privacy or offline operation matters. That holds only if the speed and memory figures survive contact with ordinary hardware.

The bigger story is the direction of travel. A 31.7 MB download scoring competitively on an accuracy test suggests the floor for on-device speech is rising quickly. Underdog's founder predicts that most of today's capable models will run on devices people already own within months. That's a prediction, not a result, and it's worth keeping that distinction in mind.

If you want to judge it for yourself, listen to the official samples, test it on the weakest device you care about, read the license line by line, and ask for commercial terms before you build anything on top of it.

So here's the question I'd love to hear your take on: if a model this small can read tricky sentences accurately on your own device, what would still make you pay for cloud voice? Is it the quality, the language coverage, or just not wanting to maintain it yourself?