A local voice engine with a sensible target
Pocket TTS has a narrower and more useful pitch than many speech projects: run text-to-speech on a normal CPU, start returning audio quickly, and avoid a remote API. Kyutai packages the model as a Python library, a command-line tool, and a local web server. The model has 100 million parameters, supports streaming, and covers English, French, German, Portuguese, Italian, and Spanish. Voice cloning accepts an ordinary audio file, while prepared voice states can be saved as safetensors for faster reuse.
That combination makes the project easy to place in a real product. It can speak notifications in a private desktop assistant, power an offline screen reader, add voices to a game, or sit behind an internal application that must not send text to a third party. The MIT license also removes much of the licensing friction around the code, although developers still need to check the separate licenses attached to individual voice samples.
The CPU focus is the real advantage
The headline feature is not simply that Pocket TTS can run without a GPU. Many speech systems can do that slowly. Kyutai designed this model around CPU use and publishes a concrete reference result: roughly 200 milliseconds to the first audio chunk and around six times real-time generation on a MacBook Air M4 using two CPU cores. Those are the project's measurements, not a promise for every machine, but they show what the architecture is trying to achieve.
The streaming interface matters just as much as total generation speed. A reader or conversational agent can begin playback before a long response is complete. The README also says the model can handle text of unlimited length, which fits audiobook and document-reading workflows better than systems built around short utterances. In practice, long-form users should still split and inspect output, especially while pronunciation and repeated-word reports remain open.
The API is pleasantly direct. Load a model, turn a supplied voice or audio prompt into a voice state, generate a tensor, and write the samples to a WAV file. The CLI mirrors that simplicity. For repeated requests, the serve command keeps the model in memory and includes a basic browser interface. This is the right operational advice because model loading and conversion of a raw voice prompt are relatively slow.
Voice cloning is useful, not effortless
Passing a WAV file through --voice is a low barrier to experimentation, and exporting its processed state avoids repeating that work later. Pocket TTS also supplies a sizable list of ready-to-try voices across its supported languages. This is enough to compare styles before collecting your own permitted sample.
There are two catches. First, the model reproduces defects in the input recording, so Kyutai recommends cleaning the sample. That turns a one-argument demo into an audio preparation task when quality matters. Second, cloned speech creates legal and ethical obligations. The project explicitly prohibits unauthorized impersonation and deceptive uses. A production team needs documented consent and controls around uploaded samples, not merely a checkbox added after the feature ships.
Language support is broader than the original English-only shape, and larger 24-layer variants are available for non-English languages when quality matters more than speed. Still, language coverage is not the same as linguistic polish. An open report says long numbers, particularly in Italian and other non-English cases, can be read poorly because number-to-word normalization is absent. Another report concerns decimal splitting. Test dates, prices, identifiers, abbreviations, and domain-specific names in every target language.
Rough edges that affect production
The most important limitations are concrete. The README says text-controlled silence for pauses is unsupported. That rules out an easy way to direct pacing for narration, announcements, and character dialogue. Applications can split audio and insert silence themselves, but then they own timing and stitching.
Recent issue activity also shows format and compatibility problems. Users have reported downloaded WAV files with placeholder size fields, an exported-voice regression in the server command, and intermittent inserted or repeated words with exported profiles on Apple Silicon. Some related fixes are already under discussion as pull requests, which is a positive sign, but the reports are reasons to add automated duration checks and human listening tests. Do not assume a valid-looking file is correct because synthesis completed without an exception.
GPU users should reset expectations too. The README says the maintainers did not see a speed benefit from GPU execution because the model is small and uses a batch size of one. This project is optimized for responsive single-stream CPU work, not bulk datacenter synthesis. Teams generating thousands of clips may prefer a system designed for batching or run their own throughput comparison.
Health, documentation, and the buying decision
Pocket TTS is young, but it does not look neglected. The repository was pushed in July 2026, its latest release arrived in May, and issues and pull requests were being updated in August. The combined open count includes both kinds of work, so it should be read as an active queue rather than 75 confirmed bugs. The latest release fixed non-CPU device support and voice cloning with quantization, while adding default voices for each language.
Documentation is a standout. The README covers supported Python versions, CLI commands, server use, voice export, the Python API, unsupported features, community browser ports, and downstream integrations. Separate generated documentation, a model card, a technical report, and a paper give developers somewhere to go beyond the quick start.
Pocket TTS is the right first trial when local CPU speech, streaming, and cloneable voices matter more than studio-grade controls. It is not yet a fire-and-forget narration backend. Build a vocabulary test set, validate generated files, pin a known-good version, and keep a human ear in the release process. With those guardrails, the project offers an unusually capable amount of speech synthesis for its installation and hardware cost.