TeALO/qwen3-tts-gguf

Qwen3-TTS 12Hz 0.6B — GGUF for qwen3-tts.cpp

GGUF conversion of Qwen/Qwen3-TTS-12Hz-0.6B-Base for predict-woo/qwen3-tts.cpp.

Why this repo exists

The other Qwen3-TTS GGUF repos on the Hub target CrispASR, a different runtime. Their tensor layout does not match what qwen3-tts.cpp loads:

qwen3-tts.cpp expects	CrispASR conversions ship
talker.blk.N.attn_q.weight	same
code_pred.lm_head.*	code_pred.output.N.weight
pre_tfm.blk.*	absent

These files were produced with qwen3-tts.cpp's own scripts/convert_tts_to_gguf.py and scripts/convert_tokenizer_to_gguf.py, so the layout matches by construction.

Files

File	Size	Role
qwen3-tts-0.6b-q8_0.gguf	1.3 GB	Talker + code predictor + speaker encoder, Q8_0
qwen3-tts-tokenizer-f16.gguf	261 MB	Speech tokenizer / codec, Q8_0

Both filenames are the ones Qwen3TTS::load_models looks for; keep them as-is and place both in one directory.

Usage

./build/qwen3-tts-cli -m models -t "Bonjour" -o out.wav
./build/qwen3-tts-cli -m models -t "Bonjour" -r reference.wav -o cloned.wav

Voice cloning needs only a reference clip — no transcript.

Verified

Loaded and run before publishing: synthesis and zero-shot cloning both produce valid 24 kHz audio at RTF ~3.7 with ~1.6 GB peak RSS on a desktop CPU.

Weights are Apache-2.0, as is the original model.