VoiceDesign generates speech from a natural-language voice description (instruct), no reference audio required. Great for designing reusable voice personas: design once → save the generated wav → extract a speaker embedding from it → use that embedding with the Base model for fast inference.

Files

File	Type	Size	Notes
`qwen3-tts-1.7b-voicedesign-f16.gguf`	F16	3.84 GB	Recommended. Same precision as upstream safetensors.

You also need the tokenizer/codec gguf — convert it from Qwen/Qwen3-TTS-Tokenizer-12Hz using the same convert_tts_to_gguf.py script, or grab one from the existing cgisky/qwen3-tts-custom-gguf repo (look for qwen3-tts-tokenizer-f16.gguf).

Quickstart with qwen3-tts.cpp CLI

bash

./qwen3-tts-cli \
  --model models \
  --model-name qwen3-tts-1.7b-voicedesign-f16.gguf \
  --text "Hello, I'm a custom-designed voice." \
  --instruction "warm middle-aged male voice, slow pace, friendly tone" \
  --output designed_voice.wav

Persona-design flow

Generate a reference clip with VoiceDesign:
bash
1./qwen3-tts-cli -m models --model-name qwen3-tts-1.7b-voicedesign-f16.gguf \ 2 -t "The quick brown fox jumps over the lazy dog." \ 3 --instruction "deep authoritative male voice" \ 4 -o persona_reference.wav
Extract a reusable speaker embedding from that clip (using the Base model):
bash
1./qwen3-tts-cli -m models --model-name qwen3-tts-1.7b-base-f16.gguf \ 2 -t "test" -r persona_reference.wav \ 3 --dump-speaker-embedding persona.json
Use the saved embedding for fast, consistent inference:
bash
1./qwen3-tts-cli -m models --model-name qwen3-tts-1.7b-base-f16.gguf \ 2 -t "Anything you want to say." \ 3 --speaker-embedding persona.json \ 4 -o output.wav

Conversion details

Converted with scripts/convert_tts_to_gguf.py from the qwen3-tts.cpp repo. No quantization (F16, identical numeric precision to the original safetensors).

bash

python scripts/convert_tts_to_gguf.py \
  --input Qwen3-TTS-12Hz-1.7B-VoiceDesign \
  --output qwen3-tts-1.7b-voicedesign-f16.gguf \
  --type f16

Hardware tested

NVIDIA RTX 3090 (compute capability 8.6) — works
CUDA 12.6, driver-side ggml-cuda (with GGML_CUDA_USE_GRAPHS)

License

Apache 2.0, inheriting from upstream Qwen3-TTS-VoiceDesign.

Citation

bibtex

@misc{qwen3tts,
  title  = {Qwen3-TTS},
  author = {Qwen Team},
  year   = {2026},
  url    = {https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign}
}

justmaier/Qwen3-TTS-12Hz-1.7B-VoiceDesign-GGUF

Qwen3-TTS-12Hz-1.7B-VoiceDesign — GGUF