malaiwah/qwen3-tts-preset-voices

Qwen3-TTS preset voice embeddings

The 9 named speakers from Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice packaged as a small sidecar bundle usable with the -Base checkpoint.

• bundle.safetensors — 9 × 2048-d bfloat16 rows, ~37 KB total
• bundle.json — metadata (speaker name → spk_id, gender, supported languages)

Each row is lifted from talker.model.codec_embedding.weight in the CustomVoice checkpoint at the speaker-ID index from its config.json. With these rows, you can:

1. Deploy only Qwen/Qwen3-TTS-12Hz-1.7B-Base (no need to download the CustomVoice checkpoint separately)
2. Use the named voices via the standard voice-cloning API path

Usage

from safetensors.torch import load_file
from qwen_tts import VoiceClonePromptItem

rows = load_file("bundle.safetensors")  # dict[str, Tensor(2048,)]

item = VoiceClonePromptItem(
    ref_code=None,
    ref_spk_embedding=rows["vivian"].cuda().bfloat16(),
    x_vector_only_mode=True,
    icl_mode=False,
    ref_text=None,
)
wavs, sr = base_model.generate_voice_clone(
    text="Hello, world.", language="English", voice_clone_prompt=[item],
)

Provenance & validation

Audio A/B comparisons — CustomVoice native vs -Base with lifted row — are published as a separate dataset: malaiwah/qwen3-tts-customvoice-ab-clips. Writeup and discussion: Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice community thread #45.

Speakers

name	spk_id	gender	languages
aiden	2861	m	en,zh
dylan	2878	m	en,zh
eric	2875	m	en,zh
ono_anna	2873	f	ja,en,zh
ryan	3061	m	en,zh
serena	3066	f	en,zh
sohee	2864	f	ko,en,zh
uncle_fu	3010	m	zh
vivian	3065	f	en,zh