The loudest number in the benchmark that pitted Pocket TTS against Kokoro, Supertonic, and Inflect-Nano is not about speed—in fact, it’s the slowest of the six configurations tested on a 4-core Xeon. It’s that no other CPU-runnable TTS model can clone a voice from five seconds of audio. Pocket TTS does it, with no fine-tuning, no GPU, no commercial restrictions. And that changes who can afford a text-to-speech that sounds like a real person.

The model, developed by Kyutai, is a roughly 100M-parameter network that treats speech as an autoregressive LLM: it generates audio tokens directly, decoded to 24 kHz via the Mimi neural codec. There’s no classic acoustic-model-plus-vocoder split; every token is produced sequentially, which explains two exotic behaviors. Latency is flat: RTF hovers around 0.69–0.76 whether you feed it 12 characters or 1,712. Kokoro, by comparison, swings from 0.49 to 0.83 depending on text length. And it’s a streaming model, suitable for interactive interfaces where the first phoneme fires immediately.

Zero-shot cloning is the key feature. A 5-second reference clip is enough for the system to reproduce accent, timbre, pacing, and even the microphone character of the original. This is not a refinement of a pre-trained vocoder: it’s an architectural leap that none of the other tested models—Kokoro, Supertonic, Inflect-Nano—can make, because they operate with fixed voice sets. In practice, it means an organization can give a voice to its assistants without sending data to a cloud service and without waiting weeks for fine-tuning. Data sovereignty is immediate here: the voice reference never leaves on-premise infrastructure.

For a fair quality comparison, the benchmark pinned Pocket TTS to a preset voice (“alba”). Objective numbers (MOS via UTMOS) place it at 4.10, slightly below Kokoro (4.44–4.46) but above Inflect-Nano and Supertonic’s speed-oriented configurations. Note the Inflect-Nano case: UTMOS gives it 3.48, but to the ear it sounds buzzy and robotic—a known failure mode for quality predictors when rating very small HiFi-GAN vocoders. Moreover, it caps output at about 15 seconds, inflating RTF on long inputs.

On practicality, installation is surprisingly painless: a plain pip install pocket-tts and the system downloads weights on first run. No CUDA builds, no manual path tweaks. For anyone evaluating on-premise deployment, this is an often underestimated differentiator: ease of setup reduces operational cost as much as a good TCO does.

Then there’s the license. Pocket TTS is MIT, essentially permitting any commercial use. Kokoro is Apache 2.0 (also permissive), but Supertonic uses OpenRAIL-M with commercial restrictions. In enterprise contexts where legal compliance is a prerequisite, the MIT license removes an entire layer of negotiation.

The structural reading is non-trivial. A ~100M-parameter model running on CPU that can clone vocal identities suggests that paralinguistic information can be compressed and reproduced without a GPU and with a lightweight neural codec. This lowers the investment floor for personalized synthetic voice scenarios: hospitality, healthcare, remote assistance, IoT devices handling natural voice commands with no cloud roundtrip. Even in journalism or media, where multilingual voice-overs could keep a reporter’s vocal footprint.

At the same time, the technology raises questions about protection from misuse: cloning without consent becomes cheaper and more portable. This is not a new problem, but until now models with such capabilities required computational heft that acted as an implicit barrier. Pocket TTS, running on a 4-core Xeon, shifts the boundary. The governance burden falls entirely on the integrator, not on a technological gatekeeper.

Who loses? Commercial TTS vendors selling predefined voices or cloning as a paid service may see the value of basic personalization eroded. Who wins? Every organization with data-control needs and voice personalization requirements that previously excluded voice cloning because infrastructure costs were too high.

Pocket TTS isn’t the fastest TTS, nor the one with the highest MOS. But the combination of streaming architecture, zero-shot cloning on CPU, and MIT license makes it the only one that can enter environments where the generated voice must sound like a real person, without vocal samples ever leaving the server room.