qwen3-tts-0.6b-customvoice for the Ryzen AI NPU

Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice converted to run on the Ryzen AI NPU of AMD Strix Halo inside halogen-flash-server, beside the Flash model on the GPU and behind the same port. Questions go to the Discord. Bugs go to the server repository's issues.

These files run only inside the halogen-flash-server container, on the NPU. They will not load in transformers or any other runtime.

What is here

devices/                                 511 MB  the NPU programs for the talker and the code predictor, and its NOTICE.md
qwen3-tts-0.6b-customvoice.hnpw          516 MB  the talker's and the code predictor's weights, converted for the NPU
tokenizer/                                11 MB  the tokenizer
hf/                                      2.4 GB  the source checkpoint's own files, read on the CPU (the speech decoder among them)

Use it

Start halogen-flash-server 0.18.14 or later with HALOGEN_DOWNLOAD on and HALOGEN_NPU_MODELS=qwen3-tts-0.6b-customvoice. The server fetches these files into /models/npu/qwen3-tts-0.6b-customvoice/ and checks each one's size and checksum before it starts.

It speaks a text on POST /v1/audio/speech in OpenAI's request shape, with the model's nine preset voices and its languages, as wav or pcm, or pcm streamed as it is made. The talker and the code predictor run on the NPU and the speech decoder on the CPU, so the GPU stays with the LLM. On one Strix Halo it speaks about 1.5 times faster than real time, with the first audio in about 0.3 seconds. This size takes no instructions; the 1.7B does.

The host needs the NPU driver, XRT with its NPU plugin, and the GPU's fabric clock held while the NPU works beside the GPU. The setup and every request shape are in docs/NPU.md.

Source and license

Converted from Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice at revision 85e237c12c027371202489a0ec509ded67b5e4b5; hf/ holds that revision's files unchanged. Apache-2.0, as the source. devices/NOTICE.md lists the third-party notices the NPU programs carry.