To use with llama.cpp
It says: “Runs in stock llama.cpp with embeddings=true.”
Could you please provide the command-line syntax for running it? I just can't seem to get it to work.
And how do I then get the WAV file or access it via HTTP?
Oh... The announce is written by Claude Code, the README is not precise enough.
Actually we built a codec.cpp which is served for LLM backbone (llama.cpp) + audio tokenizer (a seperated graph for audio enc-decoder model) structure.
For Chatterbox you can run it in llama.cpp without patch, but you will got uninterpretable outputs. You need codec.cpp to decode it.
Have you tried this version in a language other than English?
For example, ru:
./tts-cli synthesize --backbone chatterbox-mtl-t3-q4_k_s.gguf --model chatterbox-mtl-codec-q4_k_m.gguf --text "[ru] Привет! Это проверка на русском" --output out_ru.wav
It produces something incomprehensible.
Français - better, but "ç" - is like "c".
--text "[fr] Salut ! C'est un test en français."
if trying, as described in the card
text = "[ru]" + unicodedata.normalize("NFKD", "Привет! Это проверка на русском.lower())
text = "[fr]" + unicodedata.normalize("NFKD", "Salut ! C'est un test en français.".lower())
makes no difference
Thanks for the report! codec.cpp's Chatterbox tokenizer turned all non-ASCII letters (Cyrillic, "ç", etc.) into [UNK]. We had only tested English and Chinese, and both happen to avoid the bug. It's fixed in 67a30d3. Just pull and rebuild; you don't need new GGUFs.