Modelpublic

dbrains/Qwen3-TTS-12Hz-Speaker-Encoder-GGUF

sourceHugging Faceapache-2.0updated 5mo agoView on Hugging Face

0likes25downloads

Model Card

Qwen3-TTS-12Hz-Speaker-Encoder-GGUF

A 24 MB GGUF carrying only the speaker-encoder (tok_enc.*) tensors of `Qwen/Qwen3-TTS-12Hz-1.7B-Base`. The full Base GGUF is ~2.4 GB; ~99% of those bytes (talker, code-pred, vocoder) are dead weight if all you want is voice cloning on top of a separate talker (e.g. VoiceDesign).

Extracted via scripts/extract_spk_enc.py from `dbrain/qwen3-tts.cpp`.

File

—Qwen3-TTS-12Hz-Speaker-Encoder-F16.gguf — F16, 24 MB

Compatibility

Requires `dbrain/qwen3-tts.cpp`. The --speaker-encoder / --hf-repo-se sidecar load path is fork-only — upstream `khimaros/qwen3-tts.cpp` expects the encoder tensors to live inside a Base GGUF, not as a separate file.

Mainline llama.cpp does not support qwen3-tts.

Usage

bash

qwen3-tts-server \
    --hf-repo    khimaros/Qwen3-TTS-12Hz-1.7B-VoiceDesign-GGUF:Q8_0 \
    --hf-repo-se dbrains/Qwen3-TTS-12Hz-Speaker-Encoder-GGUF:F16

License

Apache-2.0 — derivative of `Qwen/Qwen3-TTS-12Hz-1.7B-Base` (speaker-encoder subset only).