hk-cantonese-cc0 emotion steering makes this voice speak angry, fearful or sad. This model was fine-tuned from Qwen3-TTS Base, which has no emotion control; the pack transfers some of Qwen3-TTS CustomVoice's emotion control to it, with no emotion-labelled speech and the same commercial-friendly licence. It works with this model in PyTorch (torch/steer.py) and with the GGUF in ggtts. Listen to the samples on its page.
A LoRA adapter for Qwen3-TTS 12Hz 1.7B that provides a Hong Kong Cantonese female voice. Fine-tuned only on CC-0 Common Voice data and distributed under Apache-2.0. The Qwen3-TTS base underneath 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.
Status: v0.0.5, 2026-10-25. Single donor voice. Built from scratch by the published
recipe, reproducibly.
v0.0.1 stays available at the v0.0.1 tag. Measured bounds, the blind listen and the limitations are in
docs/MODEL_CARD.md in this package.
zh-HK 27.0 to yue 22.0, and every clip is restored with Sidon,
its dynamics put back and its edges trimmed.hk_voice_343 (was hk_cantonese_speaker); the same speaker vector.adapter_model.safetensors 139.5 MB Rank-64 LoRA over 196 talker projections (fp16)
extra_tensors.safetensors 16 KB Baked speaker embedding, codec index 3000
adapter_config.json LoRA config
SHA256SUMS Digests of the two .safetensors files
LICENSE, NOTICE Apache-2.0 and required attribution
scripts/
apply_adapter.py Merge script; builds the runnable model
docs/
MODEL_CARD.md Evaluation, limitations, training
recovery_report.json Per-tensor SVD reconstruction error report
samples/
v0.0.5-e11.flac, v0.0.1-e11.flac The same paragraph, v0.0.5 then v0.0.1
v0.0.5-e21.flac, v0.0.1-e21.flac
v0.0.5-c11.flac, v0.0.1-c11.flac
v0.0.5-c21.flac, v0.0.1-c21.flac
sentences.txt The paragraphs' text and provenance
The three adapter files stay at the repository root, where a loader expects them.
Each row is one paragraph of held-out prose, first draws, read by v0.0.5 and by v0.0.1. The files are in samples/
(FLAC); samples/sentences.txt has the text.
| Paragraph | v0.0.5 | v0.0.1 |
|---|---|---|
| e11 | ||
| e21 | ||
| c11 | ||
| c21 |
Download Qwen/Qwen3-TTS-12Hz-1.7B-Base (the -Base variant, not -CustomVoice or -VoiceDesign).
# Standard inference build
python scripts/apply_adapter.py --base ./Qwen3-TTS-12Hz-1.7B-Base --adapter . --out ./hk-cantonese
# Add --trainable ONLY if you intend to run further fine-tuning
python scripts/apply_adapter.py --base ./Qwen3-TTS-12Hz-1.7B-Base --adapter . --out ./hk-cantonese-trainable --trainable
The two builds are not interchangeable. The inference build sets tts_model_type to custom_voice and drops the
speaker-encoder tensors, so training from it fails at the first batch.
Note: standard
peftloaders cannot load this directly, because the voice relies on a non-LoRA baked embedding row (extra_tensors.safetensors) and the build drops the upstream speaker encoder.scripts/apply_adapter.pyneeds onlytorchandsafetensors. Run it from the repository root.
import torch, soundfile as sf
from qwen_tts import Qwen3TTSModel
m = Qwen3TTSModel.from_pretrained("./hk-cantonese", dtype=torch.bfloat16, device_map="cuda",
attn_implementation="sdpa")
text = "今日天氣好好,我哋出去行下街啦。" # split long text at punctuation, up to 45 characters a chunk
wavs, sr = m.generate_custom_voice(text=[text], speaker=["hk_voice_343"], language=["chinese"])
sf.write("out.flac", wavs[0], sr)
Keep language="chinese"; the adapter carries the Cantonese. Keep the default sampling: greedy decoding and low
temperatures loop.
dandelion-love/qwen3-tts-cantonese-recipe (tag v0.0.5) builds this model from the data up: Common Voice selection,
Sidon cleanup, both LoRA phases, the voice bake, the GGUF conversion and the release gate, stage by stage
(python recipe.py <stage>). Its README has the full steps; in short:
pip install -r requirements.txt and bash scripts/build_gguf_runtime.sh; fetch the base, codec, Sidon and the
recognisers it lists.yue and zh-HK 27.0 from the Mozilla Data Collective yourself, under its terms.VOICE_SOURCE at this model built with scripts/apply_adapter.py: its voice rows are bit-identical to the
ones the release baked.convert, compare the talker md5 with 672040405ec852babf6accc15d82f81a.What reproduces exactly: training is bit-exact for the same inputs and seed (seed = 2); the lab's from-scratch
run of the recipe reproduced this release's md5. The yue 22.0 and donor clip selections are the lab's own and
cannot be re-hosted, so a fresh selection from Common Voice gives a close build rather than the identical one.
The values the recipe uses, for reference:
Qwen/Qwen3-TTS-12Hz-1.7B-Baseyue 22.0 (33,096 clips, 39.6 h) and zh-HK 27.0 (29,511 clips, 31.5 h), each
restored with Sidon, dynamics put back, trimmed. LoRA r 32, alpha 64, LR 1e-4, 1 epoch, --keep_speaker_encoder.sub_talker_weight 0, seed 2, exact seeded
kernels. About 7 h on an RTX 3080 Ti.The two chained r-32 phases are why the recovered adapter is rank 64. It rebuilds the release within 1.76 bf16 steps on the worst
tensor (docs/recovery_report.json); on 20 held-out lines the rebuild reads at the release's CER (0.151 against 0.150).
| Component | Source | Licence |
|---|---|---|
| Base model | Qwen/Qwen3-TTS-12Hz-1.7B-Base |
Apache-2.0 |
| Language phase | Common Voice yue 22.0 and zh-HK 27.0 |
CC-0 |
| Speaker phase | One Common Voice yue HK contributor, 343 clips |
CC-0 |
Adapter and scripts: Apache-2.0.
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. Common Voice 27.0 comes from the Mozilla Data Collective, whose terms forbid re-hosting the dataset or identifying its speakers; none of it is here.
Voice donor: dedicated under CC-0 via Mozilla Common Voice by a pseudonymous contributor.
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.
adapter_model.safetensors 140 MBextra_tensors.safetensors 16.6 kBQwen/Qwen3-TTS-12Hz-1.7B-Base