Pocket TTS: Kyutai's 100M-Parameter Voice Cloning Model That Runs on CPU Alone

A MODEL THAT CLONES A VOICE FROM 20 SECONDS no gpu. no api. no cloud. one pip install ★ 6.8k pocket-tts by kyutai, mit license, pip install pocket-tts - the lab behind moshi: → 100M params, runs on plain cpu - they tried a gpu, it wasn't even faster → ~6x faster than real-time on a macbook air m4, using 2 cores → ~200ms to the first audio chunk, streams the rest → clone a voice from a short sample, or use 26 built-in ones → 6 languages, handles infinitely long text → small enough to run in the browser via webassembly everyone's renting gpu time to make audio. this fits in your pocket https://lnkd.in/dpXEvrZX

I checked this against Kyutai's official GitHub repo, Hugging Face model card, and their blog.

Confirmed accurate:

• Pocket TTS is Kyutai's lightweight TTS model, designed to run efficiently on CPUs, installable via pip and a function call, requiring no GPU version of PyTorch — matches "no gpu. no api. no cloud. one pip install" precisely. Kyutai is indeed the lab behind Moshi. GitHub
• The "GPU wasn't even faster" claim checks out: the developers explicitly tested GPU acceleration and found no performance improvement over CPU execution. Zen MCP Server
• The model has approximately 100 million parameters — matches "100M params" exactly. Grokipedia
• With 100M-parameter models running on 2 cores, Pocket TTS delivers approximately 200ms latency and 6x real-time speed on modern CPUs — matches the post's "~6x faster than real-time... using 2 cores" and "~200ms to the first audio chunk" almost word for word. Aibit
• The language claim is accurate and specifically confirmed: Kyutai's own blog post title states "Pocket TTS now supports six languages," dated May 2026 — matches "6 languages" exactly. Kyutai
• Pocket-TTS is released under the MIT license — confirmed. AIBit
• Community WebAssembly ports exist, including sherpa-onnx bindings that run PocketTTS with WebAssembly support across many platforms — supports the "small enough to run in the browser via webassembly" claim, though it's worth noting the primary browser demo and some WASM implementations are community contributions rather than solely an official first-party feature, which the post doesn't distinguish. GitHub
• The "infinitely long text" and "clone a voice from a short sample" claims align with Pocket TTS processing continuously without memory explosion on arbitrarily long text, and cloning voice characteristics from a provided audio sample. Brightcoding

Not independently verified:

• I couldn't confirm the exact "26 built-in voices" figure — Kyutai's own docs describe only "a small catalog of voices" without a hardcoded count, and a third-party guide listed 8 named voices as examples, not a definitive total. The number may well be accurate (voice catalogs do get updated), but I found no source stating "26" specifically. GitHub
• I couldn't verify the specific "20 seconds" minimum sample length for voice cloning or the exact "6.8k" star count, though both are plausible and consistent with a genuinely popular, actively-developed open-source repo with dozens of community ports and integrations.

Overall, this is an unusually accurate and well-researched post — nearly every specific technical figure matches Kyutai's own documentation precisely.