idle-intelligence/pocket-tts-gguf

Pocket TTS, GGUF Q8_0

Q8_0 quantized versions of kyutai/pocket-tts-without-voice-cloning in GGUF format, for text-to-speech in the browser with WebAssembly. The layout mirrors Kyutai's repo: the original English model at the root, one folder per language under languages/.

Try the demo

Files

Path	Source in Kyutai's repo	Kyutai revision	Size	sha256
pocket-tts-q8_0.gguf	tts_b6369a24.safetensors	February 2026	134,356,064 B	6861029e…
tokenizer.model	tokenizer.model	February 2026	59,339 B	d461765a…
languages/english/pocket-tts-q8_0.gguf	languages/english/model.safetensors (= english_2026-09)	4e1e0a3	133,837,984 B	ac2fcfc7…
languages/french/pocket-tts-q8_0.gguf	languages/french/model.safetensors	4e1e0a3	133,837,984 B	74fa7e52…
languages/german/pocket-tts-q8_0.gguf	languages/german/model.safetensors	4e1e0a3	133,837,984 B	8aca1b53…
languages/spanish/pocket-tts-q8_0.gguf	languages/spanish/model.safetensors	4e1e0a3	133,837,984 B	f284ce80…
languages/portuguese/pocket-tts-q8_0.gguf	languages/portuguese/model.safetensors	4e1e0a3	133,837,984 B	3fe9a5ea…
languages/italian/pocket-tts-q8_0.gguf	languages/italian/model.safetensors	4e1e0a3	133,837,984 B	e67642e6…

Each languages/<name>/ folder also holds that language's tokenizer.model, byte-identical to Kyutai's.

Revision 4e1e0a3 is 4e1e0a3e611c51c0b4ed8174fc10f32a54644303 (2026-09-24), the one the pocket-tts 3.3.0 Python package pins.

Voices

Voices are not included. Use Kyutai's, from the same revision as the model:

• root model: embeddings_v2/<voice>.safetensors
• languages/<name>/ models: languages/<name>/embeddings/<voice>.safetensors at revision 4e1e0a3

A voice file is a KV cache computed by one specific model, so a model and its voices must come from the same revision. On 2026-10-01 Kyutai re-uploaded every non-English language, models and voices together. Voices from that upload played through these 2026-09-24 models give broken speech (repeated syllables, wrong lengths), most audibly in Italian.

What's included

Each GGUF contains the TTS decoder pipeline only: the transformer backbone, the flow matching network, the Mimi decoder and decoder transformer, and the output projection. The Mimi encoder (used for voice cloning) is left out, which saves about 28% of the original size.

Quantization

Per-block Q8_0 (block size 32): one f16 scale and 32 int8 values per block, on all linear and projection weights. Norms, biases, embeddings and the SEANet decoder convolutions stay F32. Worst-tensor SQNR is above 39 dB for every language and above 40 dB for the root model.

Runtime

Weights stay quantized at runtime. Matmuls use a tiled WASM SIMD128 quantized matmul kernel (a fork of Candle), about 2x realtime on a desktop browser.

Usage

These files are made for tts-web, a browser text-to-speech engine built with Candle and WebAssembly.

Acknowledgments

Based on Kyutai's Pocket TTS, a 100M-parameter text-to-speech model.

This is an independent port by Idle Intelligence, not affiliated with or endorsed by Kyutai Labs.

License

CC-BY-4.0, same as the original model.