Qwen3-TTS-12Hz-0.6B-CustomVoice GGUF

GGUF packages of Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice for audio.cpp (qwen3_tts family).

Each file is a standalone package: the talker, the speech tokenizer, config.json, the tokenizer files and the audio.cpp model spec are all embedded. Point --model at the .gguf file (or at a directory holding only one of them).

The Q8_0 file uses the converter's default selection, which matches the official Qwen3-TTS-12Hz-0.6B-Base Q8_0 package from audio-cpp/audio.cpp-gguf: talker attention/MLP projections, text projection, codec head and code-predictor heads are Q8_0; norms stay BF16, text and codec embeddings F16, speech-tokenizer convolutions F32.

Usage

Built-in speakers: Vivian, Serena, Uncle_Fu, Dylan, Eric, Ryan, Aiden, Ono_Anna, Sohee.

audiocpp_cli --task tts --family qwen3_tts \
  --model qwen3-tts-12hz-0.6b-customvoice-q8_0.gguf \
  --text "今天天气不错，我们一起去公园散步吧。" \
  --request-option voice_id=Vivian \
  --instruct "Very happy." \
  --out out.wav

--instruct is optional. Streaming (--mode streaming) works the same way as the 1.7B CustomVoice package.

How it was made

Converted with audiocpp_gguf built from audio.cpp commit 52118521b338aaa376ad58350ec772342993485b:

audiocpp_gguf \
  --input model_weights=Qwen3-TTS-12Hz-0.6B-CustomVoice/model.safetensors \
  --input speech_tokenizer_weights=Qwen3-TTS-12Hz-0.6B-CustomVoice/speech_tokenizer/model.safetensors \
  --root Qwen3-TTS-12Hz-0.6B-CustomVoice \
  --family qwen3_tts \
  --output qwen3-tts-12hz-0.6b-customvoice-q8_0.gguf \
  --type q8_0

(--type bf16 for the BF16 file.)

Validation

On CPU with the audio.cpp v0.9.2 library, the Q8_0 file was run with all nine speakers on short Chinese, English, Japanese and Korean sentences, an emotion instruction, and a 35-second Chinese passage; the BF16 file with Vivian on a short sentence and the same passage. Transcribing every output with Qwen3-ASR-0.6B (Q8_0) returned the input text, differing only in punctuation. The 35-second passage showed no long silences (longest pause 1.2 s for Q8_0, 0.7 s for BF16).

License

Apache-2.0, the license of the original Qwen3-TTS weights.