pocket-tts mit Homebrew, Nix installieren

Projektlinks

• Homepage
• Paketmanager-Seite

Befehle und Aliase

• pocket-tts

Projektgeschichte

Kyutai introduced Pocket TTS from its work on continuous audio language models. The associated research was published in 2025, and the project exposed the model through an open repository, Python API and command-line interface.

During 2026 the project added official multilingual models, quantization, training code, community model documentation and increasingly small CPU-oriented models and executables.

Adoptionsgeschichte

Pocket TTS was designed to make neural speech generation practical on ordinary CPUs rather than requiring a hosted API or accelerator. Its CLI, Python API, streaming output and voice prompting support local applications and experiments.

The project developed an ecosystem of community-trained language models and alternative implementations. Later releases added multilingual official models, training code and small native runtime binaries.

Wie es verwendet wird

Users can generate audio from the pocket-tts CLI or Python API, selecting a built-in voice or supplying an audio prompt to clone a voice.

Model configuration may be supplied as a local or remote YAML argument, but the project does not prescribe a persistent configuration-file location or credentials file.

Warum Paket-Nerds sich dafür interessieren

Pocket TTS packages low-latency neural speech synthesis and voice cloning into a comparatively small local tool. This makes modern TTS accessible through a normal package-manager workflow instead of a remote service.

Its evolution from a Python inference package to documented training code and native runtime binaries is notable for users building reproducible, offline speech pipelines.

Zeitleiste

• 2025: The Continuous Audio Language Models research underlying Pocket TTS was published.
• 2026: Version 2 added official multilingual models.
• 2026: Version 3 published the training code.
• 2026: Version 4 added sub-20-million-parameter models and native runtime binaries.

Related projects

• Kyutai publishes the Pocket TTS model weights and voice assets through its official Hugging Face organization. Community ports and trained models extend the inference code to additional runtimes and languages.

Quellen

• https://arxiv.org/abs/2509.06926
• https://github.com/kyutai-labs/pocket-tts/releases
• https://huggingface.co/kyutai/pocket-tts
• https://kyutai-labs.github.io/pocket-tts/