Qwen3-TTS for RK3588 NPU

English | 简体中文

1. Model Introduction

Qwen3-TTS is a text-to-speech model. This repository provides the RK3588 deployment artifacts (RKNN graphs + weights) together with the worker manifest (qwen3-tts.json) that the wdnwebui compute layer needs.

Field	Value
Model ID	wardner/Qwen3-TTS
Display Name	Qwen3-TTS
Model Type	TTS
Chip	RK182x
Host platform	RK3588
Runtime	RKLLM + RKNN
Capability	swcap_text_to_speech
Worker factory	qwen3_tts
Memory estimate	~1900 MB
Init time	~30 s

2. Files

Path	Description
README.md	English documentation
README_CN.md	Chinese documentation
qwen3-tts.json	Worker manifest — plain text on purpose
config.json	Model configuration and file list
SHA256SUMS	SHA-256 checksums
*.rknn	RKNN graphs (talker / code predictor / speech decoder / speaker encoder / text projection)
*.weight	Companion weights for the graphs above
*.fp16.bin	FP16 embedding tables (talker text/input embed, codec embed)
speaker_encoder_mel.npy	Speaker encoder mel filter bank
tokenizer.json	Tokenizer

qwen3-tts.json is deliberately not stored in Git LFS. It is the "manual" for this distribution unit and must stay readable as plain text so it can be diffed, reviewed and previewed on the Hub.

3. Download

hf download wardner/Qwen3-TTS --local-dir ./Qwen3-TTS

Or with git (requires Git LFS):

git clone https://huggingface.co/wardner/Qwen3-TTS

If huggingface.co is unreachable from your network (DNS interception / blocked access), point the client at a mirror instead:

HF_ENDPOINT=https://hf-mirror.com hf download wardner/Qwen3-TTS \
  --local-dir ./Qwen3-TTS

Note: with git clone the weights are Git LFS pointers until git lfs pull completes. hf download fetches the real files directly, and does not create a .git directory — neither is required for the model to work; see §4.

4. Installation (wdnwebui)

Clone the repository into the model root so the folder holds both the weights and the worker manifest:

git clone https://huggingface.co/wardner/Qwen3-TTS /userdata/models/Qwen3-TTS

Then open the web UI → 算力配置 (Compute) → the card appears under 「本地算力」 (the backend rescans the model root every ~2 s, so no restart is needed) → enable it.

The folder name does not have to match the model name: the SDK follows the directory that the manifest was found in, via the WDNAI_QWEN3_TTS_DIR environment variable declared in the manifest.

5. SHA-256 Verification

cd ./Qwen3-TTS
sha256sum -c SHA256SUMS

All entries must report OK before deployment.

6. Notes and Limitations

• The TTS model carries three FP16 embedding tables (text embed, input embed, codec embed, ~750 MB in total). They are loaded directly and are not released through the generic embed-table path; memory is reclaimed by malloc_trim on unload.
• available_targets in the manifest is 182x — the RKNN graphs are compiled for the 1.82 TOPS NPU generation.

7. Copyright and License

The upstream Qwen3-TTS model is distributed under the Apache License 2.0. The files in this repository are RKNN-converted derivatives, not unmodified upstream weights. See LICENSE and the upstream model documentation.