Serveurperso/pocket-tts-GGUF

Pocket TTS GGUF

GGUF weights for pockettts.cpp, a C++17/GGML port of Pocket TTS (Kyutai): a flow language model over continuous Mimi latents, zero shot voice cloning, streaming, 24 kHz mono. Every released language pack, 6 and 24 layers. Runs on CPU, CUDA, Metal, Vulkan.

Files

Two GGUFs per language pack:

pocket-tts-{pack}-{variant}.gguf flow LM + flow head + Mimi encoder and decoder + tokenizer + text rules pocket-tts-{pack}-voices.gguf the 27 predefined voices of the pack

Each language is its own model: the flow LM, its tokenizer, its text rules and its voices belong together and do not mix across packs. The Mimi codec is identical in every pack.

language	6 layers	24 layers
English	english_2026-09	english_2026-09_24l
English	english_2026-04	english_2026-04_24l
English	english_2026-01
English	english
French	french	french_24l
German	german	german_24l
Italian	italian	italian_24l
Portuguese	portuguese	portuguese_24l
Spanish	spanish	spanish_24l
Dutch	dutch	dutch_24l

variant	6 layers	24 layers	use case
F32	438 MB	1.34 GB	reference, debug, conversion
BF16	234 MB	687 MB	source faithful, max precision
Q8_0	138 MB	379 MB	recommended default
Q4_K_M	91 MB	229 MB	lowest VRAM

The voices GGUF weighs 170 MB for a 6 layer pack and 681 MB for a 24 layer pack: a voice is the full flow LM KV state after its prompt, in F32. Voices: alba, anna, azelma, bill_boerst, caro_davy, charles, cosette, daan, eponine, estelle, eve, fantine, george, giovanni, jane, javert, jean, juergen, lola, marius, mary, michael, paul, peter_yearsley, rafael, stuart_bell, vera.

Quick start

git clone --recurse-submodules https://github.com/ServeurpersoCom/pockettts.cpp.git
cd pockettts.cpp && ./buildcuda.sh
./models.sh

cd examples
./tts.sh        # predefined voice  -> tts.wav
./clone.sh      # voice cloning     -> clone.wav

Any pack, by hand:

huggingface-cli download Serveurperso/pocket-tts-GGUF \
    pocket-tts-french_24l-Q8_0.gguf pocket-tts-french_24l-voices.gguf \
    --local-dir models
echo "Bonjour tout le monde." | ./build/pocket-tts \
    --model models/pocket-tts-french_24l-Q8_0.gguf \
    --voices models/pocket-tts-french_24l-voices.gguf \
    --voice estelle -o out.wav

Backends

Set GGML_BACKEND to force a device, otherwise the runtime picks the best one available.

value	target
CUDA0	NVIDIA GPU, fastest path on Ada / Blackwell
Vulkan0	Cross vendor GPU (AMD / Intel / NVIDIA)
Metal	Apple Silicon GPU
CPU	CPU fallback, x86 variant auto selected

Quantization policy

Only the 2D matmul weights follow the variant. Everything else keeps the source precision:

tensor	dtype across all variants
Conv and transposed conv kernels (SEANet, upsample)	F32, loaded as F16 / F32
1D tensors (norms, biases, layer scales, freqs)	F32
EOS head	F32
Voices	F32

Every released pack ships BF16 weights, so the BF16 variant is the exact reference; F32 is a lossless upcast.

Prohibited use

Use of the model must comply with all applicable laws and regulations and must not result in, involve, or facilitate any illegal, harmful, deceptive, fraudulent, or unauthorized activity. Prohibited uses include, without limitation, voice impersonation or cloning without explicit and lawful consent; misinformation, disinformation, or deception (including fake news, fraudulent calls, or presenting generated content as genuine recordings of real people or events); and the generation of unlawful, harmful, libelous, abusive, harassing, discriminatory, hateful, or privacy-invasive content.

License

Upstream model: Pocket TTS by Kyutai, CC-BY-4.0 GGUF tooling: pockettts.cpp, MIT