hk-cantonese-cc0 — GGUF (v0.0.5)
New: emotion presets
hk-cantonese-cc0 emotion steering makes this voice speak angry, fearful or sad, on this GGUF through ggtts:
curl -LO https://huggingface.co/jmtl/hk-cantonese-cc0-qwen3-tts-emotion-steering/resolve/main/presets/angry.safetensors
ggtts-render --steer angry.safetensors --steer-scale 0.1875 --steer-layers 21,22,23,24,25 --text "夠喇!呢個爛攤子你自己即刻執返佢!" --out angry.flac
The presets need a runtime with steering (ggtts); other GGUF runtimes cannot apply them yet. Listen to the samples on its page.
Related models
q8_0 GGUF of hk-cantonese-cc0, a Hong Kong Cantonese female voice for Qwen3-TTS 12Hz 1.7B. Runs on qwen3-tts.cpp: no Python, no PyTorch. The adapter form, the evaluation, samples and the training recipe are in the main repository.
Status: v0.0.5, 2026-10-25. Single donor voice. v0.0.1 stays available at the v0.0.1 tag. Built by the
public recipe (tag v0.0.5), which reproduces this file's md5 from the same input clips.
Changed since v0.0.1: new training data and cleanup, reproducible training, chunks of up to 45 characters, and
the speaker is now named hk_voice_343 (was hk_cantonese_speaker). Measured against v0.0.1 it passes every
release test and a native listener preferred it blind, 15 paragraphs to 8; its loudness range is 1.4 dB narrower.
Details in the main repository's card.
Samples
Each row is one paragraph of held-out prose, first draws, read by v0.0.5 and by v0.0.1 (the GGUF builds, through
qwen3-tts.cpp). The FLAC files and their text are in the
main repository's samples/.
| Paragraph | v0.0.5 | v0.0.1 |
|---|---|---|
| e11 | ||
| e21 | ||
| c11 | ||
| c21 |
Files
| File | Size | md5 |
|---|---|---|
qwen-talker-hk-cantonese-cc0-v0.0.5-q8_0.gguf |
2.0 GB | 672040405ec852babf6accc15d82f81a |
qwen-tokenizer-12hz-q8_0.gguf |
278 MB | 39819bc6f94b35837b94b22cdd0ce95e |
SHA256SUMS |
SHA-256 of both files |
The codec is an unmodified conversion of Qwen/Qwen3-TTS-Tokenizer-12Hz (Apache-2.0), the same file as v0.0.1's,
shipped so this repository is self-contained.
Which model this is a conversion of
The merged checkpoint the recipe builds, not a rebuild of the sibling adapter. The recipe trains, merges,
bakes and converts in one pass; this file is that conversion, and the recipe reproduces its md5 from scratch given
the same input clips (its README says which reproduce and how to rebuild). Rebuilding
from the published adapter with apply_adapter.py gives a slightly different file (md5 dee1d97c7eb8b3605bfec0f88dd24b9f), because the
adapter is an SVD recovery of the merged weights.
Build the runtime
git clone https://github.com/Danmoreng/qwen3-tts.cpp.git
cd qwen3-tts.cpp
git checkout 16bb5afcd06311031c72a8488f8d59660dc2fb46
cmake -B build -DGGML_CUDA=ON # omit -DGGML_CUDA=ON for a CPU-only build
cmake --build build -j
This is the commit these weights were rendered and evaluated on. Build on a native filesystem; under WSL a build
over /mnt is far slower.
Run
-m takes the directory holding the talker; --codec-model takes the codec file.
mkdir -p talker && mv qwen-talker-hk-cantonese-cc0-v0.0.5-q8_0.gguf talker/
./build/qwen3-tts-cli \
-m ./talker \
--codec-model ./qwen-tokenizer-12hz-q8_0.gguf \
--speaker hk_voice_343 \
--seed 20260905 \
-t "今日天氣好好,我哋出去行下街啦。" \
-o out.wav
--speaker hk_voice_343selects the baked voice. This build conditions on a baked speaker, not reference audio, so it cannot voice-clone.--seedmakes a render repeatable. Without it the same text gives a different draw.- The runtime finds the talker by matching
qwen-talkerin the filename, so any name keeping that prefix works. - The CLI writes WAV; convert to FLAC if the output will be kept or compared.
Operational constraints
- Chunk at punctuation, up to 45 characters (,。:;!?). At 70 the model skips whole clauses.
- Keep default sampling. Greedy and low temperatures produce runaway loops.
- Trim the edges before using clip duration as a measure.
- Colloquial Cantonese reads best. Standard written Chinese (書面語) and literary readings mispronounce more.
Licence
Apache-2.0, matching Qwen3-TTS 12Hz 1.7B, which these weights derive from. The voice is one pseudonymous Common Voice contributor's, dedicated under CC-0. The Qwen3-TTS base was trained by its authors on 5M+ hours of data whose sources are not disclosed (technical report), so no purity claim is made for the model as a whole.
Data policy: use openly available data unless its owner restricts this use; honour attribution where required; record provenance; keep identifiable real people out of the voice phase.
Responsible use. The voice is a real person's. Do not present it as a fictional or anonymous voice, do not attempt real-world re-identification, and do not use it for impersonation. Label audio made with it as synthetic where people could take it for a recording. No watermark is embedded; the EU AI Act (Article 50) expects generated audio to be marked, so deployers should apply their own marking.