Pocket TTS Japanese โ€” GGUF

takuma104/pocket-tts-ja-cc (the 6-layer japanese model): voice-cloning Japanese TTS, fine-tuned from Kyutai's Pocket TTS on CC BY data. It reads plain Japanese, kanji included.

Two files make one model, as with llama.cpp's multimodal projectors: the model file is the text-to-speech LM backbone, run by llama.cpp; the codec file holds the audio side (decoder / vocoder, speaker encoders and the adaptor between the two), run by codec.cpp. The runtimes below wire the two graphs together, on CPU or GPU (CUDA, Vulkan, Metal), fully offline.

Files

Every run needs one model file + one codec file from this repo.

--backbone --model (codec) Total
Recommended pocket-tts-ja-bf16.gguf codec-bf16.gguf 236 MB
Smaller pocket-tts-ja-q8_0.gguf codec-q8_0.gguf 135 MB
Phones (fastest on ARM CPUs) pocket-tts-ja-q8_0.gguf codec-q8_0-lm-q8_0.gguf 126 MB
Smallest pocket-tts-ja-q4_0.gguf codec-q8_0.gguf 96 MB
Other files
File Size
pocket-tts-ja-q3_k_l.gguf 46 MB
pocket-tts-ja-q3_k_m.gguf 41 MB
pocket-tts-ja-q4_1.gguf 51 MB
pocket-tts-ja-q4_k_m.gguf 49 MB
pocket-tts-ja-q4_k_s.gguf 47 MB
pocket-tts-ja-q5_0.gguf 56 MB
pocket-tts-ja-q5_1.gguf 60 MB
pocket-tts-ja-q5_k_m.gguf 57 MB
pocket-tts-ja-q5_k_s.gguf 56 MB
pocket-tts-ja-q6_k.gguf 66 MB

Command line (codec.cpp)

With codec.cpp's tts-cli (build it; it bundles llama.cpp), then:

hf download BricksDisplay/Pocket-TTS-ja-GGUF pocket-tts-ja-bf16.gguf codec-bf16.gguf --local-dir Pocket-TTS-ja

./build/tts-cli synthesize --backbone Pocket-TTS-ja/pocket-tts-ja-bf16.gguf --model Pocket-TTS-ja/codec-bf16.gguf \
    --ref-audio my_voice.wav --text "ใ“ใ‚“ใซใกใฏ,ไปŠๆ—ฅใฏใ„ใ„ๅคฉๆฐ—ใงใ™ใญ." --output out.wav

All tts-cli options (voices, voice cloning, sampling, GPU) are described in the codec.cpp README.

Python

The codec_cpp package runs the same pipeline (it builds codec.cpp with its bundled llama.cpp, so it needs CMake and a C++ compiler):

pip install "codec-cpp @ git+https://github.com/mybigday/codec.cpp" huggingface_hub soundfile
import codec_cpp, soundfile
from huggingface_hub import hf_hub_download

tts = codec_cpp.TTS(hf_hub_download("BricksDisplay/Pocket-TTS-ja-GGUF", "pocket-tts-ja-bf16.gguf"),
                    hf_hub_download("BricksDisplay/Pocket-TTS-ja-GGUF", "codec-bf16.gguf"))
audio, sr = tts.synthesize("ใ“ใ‚“ใซใกใฏ,ไปŠๆ—ฅใฏใ„ใ„ๅคฉๆฐ—ใงใ™ใญ.", ref_audio="my_voice.wav")
soundfile.write("out.wav", audio, sr)

ref_audio= clones the voice; tts.stream(text) yields audio as it is generated. More in the Python docs.

Sampling

The codec files carry the original's settings (its japanese.yaml): the flow temperature 0.3 (codec.lm.temperature) and 5 frames kept after the end of speech (codec.lm.frames_after_eos). Change them with --override-kv, e.g. --override-kv codec.lm.temperature=float:0.5.

As in the original, write the text the way its config does: ใ€‚ โ†’ ., ใ€ โ†’ ,, quotes and brackets removed, full-width letters and digits as ASCII, and a final ..

Voices

There is no builtin voice: every run needs --ref-audio with a few seconds of clean speech (any sample rate). Save it with --save-speaker to reuse it without the clip. The original repo's voices/ has ten synthetic Japanese voices.

License

cc-by-4.0, the same as the original model. Credit Pocket TTS Japanese (CC BY) by takuma104 and Pocket TTS by Kyutai (paper) when you use these files; here they are converted to GGUF and quantized.

Downloads last month
380
GGUF
Model size
38.3M params
Architecture
pocket_mimi
Hardware compatibility
Log In to add your hardware

3-bit

4-bit

5-bit

6-bit

8-bit

16-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for BricksDisplay/Pocket-TTS-ja-GGUF

Quantized
(2)
this model

Paper for BricksDisplay/Pocket-TTS-ja-GGUF