Pocket TTS use cases: the Kyutai CPU voice, envelope by envelope

Pocket TTS is Kyutai's 100M-parameter speech model, and its use cases follow one rule: it goes where the hardware is thin. On two laptop CPU cores it speaks about six times faster than real time. In a browser tab it runs on WebAssembly. On a phone it reaches about real time through an outside port. This page maps what people build with it, envelope by envelope, from screen readers to game voices, with dated numbers and the cloning rules attached.

One term first. A hardware envelope is the machine that runs the model: a laptop, a server, a browser tab, a phone, a small board, or a hosted speech lane. That envelope is what constrains a build, so the page sorts every use case by it. The model has its own deep dive on the card page: Pocket TTS on this platform. For head-to-head numbers against Kokoro, use the comparison guide: Pocket TTS vs Kokoro on CPU. Every fact below was read from the official sources on 28 October 2026. The lane facts are this platform's own, dated where they appear.

Four facts that set the use cases

Two readings of that table shape the page. First, the cloning floor is what sets the model apart from fixed-voice rivals. Second, the clarity tie is what makes small machines good enough to ship: the model tied a rival more than 7 times its size on words heard wrong. Kokoro is the other CPU model in that table. The split between the two is short. Kokoro is the fixed-voice throughput pick. Pocket TTS is the one that can imitate a voice from a short sample. The full trade lives in the comparison guide above.

The envelope ladder

The ladder is the page's core. Every row names what the model does on that machine, with a dated basis. Speeds are the publisher's own numbers unless the row says otherwise.

Three honesty notes on the ladder. The GPU row is weaker than it looks. On Apple Silicon the publisher saw no GPU gain at a batch size of one, and the README now carries that correction in its own text. A GPU helps on x86 servers: about 2.6 times the speed with a small server card. It never decides whether the model runs.

The small models trade clarity for speed, and the publisher's own table shows the trade. On words heard wrong, its 18.2M tier edged the released model on its own October evaluation: 0.85 against 0.92 percent. The 1.9M model did not: 1.72 percent. One line of one test run adds 0.3 points of that, because the model invented a word and the speech checker scored it in a loop. Use the 1.9M tier for short prompts and tight chips, not narration.

Long text is a local strength. The README says the model streams and can take unbounded input, because it makes audio frame by frame and never waits for a full pass. The hosted lane caps each request at 2,500 characters instead, and the split is the client's job. That contrast is the biggest behavioral gap between the top and bottom rows of the ladder.

What people build, with receipts

The use cases below are not invented. Every named project comes from the README's own projects list, read on 28 October 2026. The families are ours. The receipts are theirs.

Reading it aloud

The oldest use of a speech model is reading the screen. The projects list carries a browser screen reader, a desktop reader that speaks clipboard text from hotkeys and steps by sentence, and a tool that turns PDFs into read-along audiobooks with the words highlighted on the page. A multi-model audiobook tool lists Pocket TTS among its narrators. One caution from the launch discussion: some screen-reader users still want the old robotic voices at very high speeds, because those stay clear where natural voices smear. Test at your users' real speeds.

Games and interactive voices

Games want many voices, fast starts, and no per-line bill. The list carries a Unity 6 build, a tool that speaks game and app text in real time, and a mod that lets players talk to named game characters in those voices. A voice app on Android, iOS, Mac, and Windows uses the model as a system voice, with streaming and low latency. Games are also where cloning consent is easiest to get right. Use the actor's voice with the actor's written permission.

Agents that talk back

A voice agent needs speech that starts fast and costs little per line. That is the use case the ranking tutorial serves: a January 2026 developer post built a voice agent with Pocket TTS as its voice. The list adds more. A local voice harness for Macs handles turn taking, dictation, and talking to coding tools. Several wrappers expose the model behind the OpenAI speech request shape, so agent code calls it unchanged. A chat server clones any voice from 20 seconds of audio, then talks back in it. Note the two numbers. The model's own floor is about 5 seconds of audio, per the technical report. The 20-second figure is that one server's stated capture length, not a model limit.

The smart home and the server shelf

A smart speaker that speaks locally is the classic use for a CPU speech model. The list carries a Home Assistant container that speaks through the Wyoming protocol. It carries a single-file C++ runtime with its own HTTP server, for boxes where Python is too heavy. The publisher's own pages agree on the split: Pocket TTS is the CPU model, and the larger Kyutai TTS 1.6B is the server model behind their Unmute voice product. If your machine is a home server or a small box, this model was built for you.

A voice in your own language

The training code shipped in August 2026, and the community now trains its own voices and languages. The README lists community models for Czech, Hindi, Korean, Farsi, and eight more, plus one multilingual Indian-languages model. Seven publisher languages cover most products today. The community tier is how a language outside that set gets a voice at all, on your own hardware, with your own data.

What cloning asks of you

Cloning is the feature that sells the model. It is also the feature with rules. Three rules, all from the publisher's own text.

The floor is small but typed. About 5 seconds of sample audio is enough, per the technical report. The clone keeps the accent, the cadence, the mood, and even the room and the mic of the sample. That last part is a warning. The README's own advice is to clean the sample first, because a noisy recording comes back as a noisy voice.

Consent is a use condition, not a suggestion. The README's use rules ban cloning without the explicit and lawful consent of the voice's owner. They ban passing made audio off as a real person's words. A no-cloning build of the model ships next to the standard one. That is the line, drawn in shipping code rather than a policy page.

Voices carry licenses, one per voice. The voice repository lists a license for each. Most of the catalog is permissive. At least one source set in it is marked non-commercial. The 228 donated voices are CC0. Read the license of the exact voice before it ships in a product. The hosted lane on this platform settles the question for its own ids: 17 fixed voices, all CC BY 4.0 or CC0, and no cloning on that surface. The vendor-by-vendor cloning rules have their own guide: TTS voice cloning and streaming limits.

When the hosted lane is the right envelope

The local route is free in money and costly in care. On Linux, a plain pip install of the model pulls the CUDA build of PyTorch: about 3 GB of wheels a CPU user never touches. The fix is one published index flag. The model must stay loaded to answer fast. Someone has to mind the box. Break-even math for a local box against a metered lane is worked for a sibling model in Kokoro TTS local vs API cost. The general formula lives in the parent guide: TTS local vs API break even.

The hosted route trades a metering line for all of that. As of 2026-10-28, the pocket-tts lane is one of the speech lanes on this platform, with kokoro-tts beside it for fixed-voice throughput. Both follow the OpenAI speech request shape, so client libraries work as-is. The speech catalog page lists the full set. This lane meters by input character against a prepaid wallet. The current rate lives on the pricing page, never typed into a page. A wallet at zero refuses the next request at the door. A runaway loop's worst case is the balance you loaded.

Five contract facts, measured and dated, shape use on this lane. The request cap is 2,500 characters. Longer input gets a 400 before any synthesis starts, and refused requests are not billed. A full-cap request runs about 48 seconds, from this platform's own 2026-09-09 measurements, so read timeouts belong above a minute. Retries re-bill: audio answers do not fit the replay budget, so a retrying client pays twice. The voice list is a free read, one call to GET /v1/audio/voices?model=pocket-tts, charged nothing. The lane measured 3.0 million characters per day at a concurrency of eight. That is a planning ceiling, not a per-request promise.

A first call, whole and runnable:

curl -sS $IRONSTRATUM_BASE_URL/v1/audio/speech \
  -H "Authorization: Bearer $IRONSTRATUM_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "pocket-tts",
    "input": "The first line, read by a CPU voice.",
    "voice": "alba"
  }' \
  --no-buffer \
  --output first-line.wav

The voice field takes one of the 17 checked ids, with alba as the default. Split longer text at sentence boundaries and join the files at playback. A mid-sentence split sounds like a pause, and nothing worse.

Decision rules

Frequently asked questions

What is Pocket TTS used for?

Five families cover most builds, all with named receipts in the README's projects list. Reading: screen readers, clipboard readers, PDF narration. Games: a Unity build, a real-time speech tool, character-voice mods. Voice agents: the January 2026 voice-agent tutorial, local voice harnesses, OpenAI-shaped wrappers. Small servers: a Home Assistant container and a C++ runtime. Plus community-trained languages. The common thread is thin hardware. Each family exists because a CPU, a browser, or a phone is the machine that has to speak.

Can Pocket TTS run on a phone?

Yes, through an outside port rather than an official app. LiteRT graphs run the model on Android phone GPUs at about real time on a Pixel 8a, read 2026-10-28. A voice app built on the model ships on Android, iOS, Mac, and Windows as a system voice. The publisher agrees on the target: its May 2026 post names laptops and phones.

Does Pocket TTS need a GPU?

No. The pitch is CPU. The 100M model runs about six times real time on two cores of a MacBook Air M4. Small English models, released October 2026, reach 25 times real time on one thread at 1.9M parameters. A GPU helps in one measured case: about 2.6 times faster on an x86 cloud VM with a Tesla T4. On Apple Silicon the publisher measured no GPU gain at a batch size of one, and says so in the README.

Is Pocket TTS free?

The model is. MIT license, open weights, and training code out since August 2026. Local runs cost hardware and power, nothing else. A hosted lane charges per character for the machine and the care. The rate lives on the pricing page. The local-versus-hosted math is a linked guide.

How many seconds of audio does voice cloning need?

About 5 seconds, per Kyutai's technical report of 13 January 2026. This page keeps the number typed and dated. The clone keeps the accent, the cadence, and the recording conditions of the sample. One chat server advertises a 20-second capture. That is its own stated length, not the model's floor. Wherever you clone, the terms require the owner's explicit, lawful consent.

What languages does Pocket TTS support?

Seven, as of the repository read on 28 October 2026: english, french, german, portuguese, italian, spanish, and dutch. The May 2026 release carried six. Dutch joined later. A voice recorded in one language can speak another, sometimes with an accent shift. Community-trained models add roughly a dozen more for local runs. The hosted lane here serves English voices today.

Can I ship a Pocket TTS voice in a commercial product?

Check the voice's own license first. The voice repository carries a per-voice list. Most entries are permissive and the donated voices are CC0, but at least one source set is marked non-commercial. Consent is separate from licensing and always required. The hosted lane here serves 17 checked voices, all CC BY 4.0 or CC0. That settles the license question for those ids. It does not settle consent for anything you clone yourself.

Rates and metering units for every lane sit on the pricing page. To try the hosted envelope: sign up, create a key, load a small wallet, and point the snippet above at your first line. If your build belongs on a laptop or a board, the official links in this page carry the rest.

Last verified: 2026-10-28