Pocket TTS
KyutaiKyutai's small text-to-speech model, about 100M parameters, built to run faster than real time on two laptop CPU cores, with streaming output and voice cloning from a short sample. It speaks English and several other European languages, installs as a Python package with a command line and a server, and runs in the browser through community ports. The repository also carries the code to train new voices and languages.
Openness
3 medium confidence- weights
- open (Hub checkpoints; the voice-cloning model sits behind an automatically approved form and the model without voice cloning is ungated)
- data
- closed (the data behind the released model is not stated; the training guide's example recipe uses the public HiFiTTS-2 corpus)
- code
- open (training code with configs for training from scratch and for fine-tuning)
- license
- CC-BY-4.0 (model weights on the Hub; the code is MIT)
The weights are licensed CC-BY-4.0, which asks only for attribution, and Kyutai has published the code to train the model from scratch. What it does not say is which data the released model was trained on.
- https://huggingface.co/api/models/kyutai/pocket-tts recorded 2026-11-12
Hub record: "gated":"auto", "license":"cc-by-4.0", with voice embeddings in safetensors in the file list.
- https://huggingface.co/api/models/kyutai/pocket-tts-without-voice-cloning recorded 2026-11-12
Hub record: "gated":false, "license":"cc-by-4.0".
- https://huggingface.co/kyutai/pocket-tts-without-voice-cloning/raw/main/README.md recorded 2026-11-12
Card front matter "license: cc-by-4.0"; "Small model size, 100M parameters"; "**New (August 2026):** We've released the training code!"
- https://raw.githubusercontent.com/kyutai-labs/pocket-tts/HEAD/LICENSE recorded 2026-11-12
Code LICENSE: "Permission is hereby granted, free of charge, to any person obtaining a copy of this software" (MIT text).
- https://raw.githubusercontent.com/kyutai-labs/pocket-tts/HEAD/training/README.md recorded 2026-11-12
Training guide: "If you have a GPU and some speech data, you can train your own Pocket TTS!"; "Our example uses HiFiTTS-2 ([nvidia/hifitts-2](https://huggingface.co/datasets/nvidia/hifitts-2))"; "You can later re-run with 2000h of data to get results on par with the official model"; configs `finetune.yaml` and `finetune_language.yaml`.
Adoption
2 high confidenceHugging Face downloads of the two checkpoints over the trailing month. The pocket-tts pip package fetches weights from the Hub as well, so its own download count is not added.
- https://huggingface.co/api/models?author=kyutai&search=pocket&limit=100 recorded 2026-11-12
"downloads":1459 for kyutai/pocket-tts and 11920 for kyutai/pocket-tts-without-voice-cloning.
Capability
3 low confidencePocket TTS has no listener rating, and the only shared-test-set results Kyutai publishes are for models trained with its own recipe, which it says match the released model. On that basis it sits level with F5-TTS, which reports on the same LibriSpeech test set.
- https://artificialanalysis.ai/text-to-speech/leaderboard recorded 2026-11-12
The 104-row leaderboard has no Pocket TTS entry.
- https://raw.githubusercontent.com/kyutai-labs/pocket-tts/HEAD/training/README.md recorded 2026-11-12
Training guide, Evaluate: "LSD from scratch, EMA weights, evaluated on the **full** LibriSpeech test-clean cross-sentence list (1127 items...)": "HiFiTTS-2, 8 GPUs | 24 | 31,700 | 400k | 0.82% | **0.929** | 4.33".
Verified 2026-11-12, except some scores, still awaiting re-confirmation