YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
RKNN TTS engines — RK3588
Pre-compiled RKNPU2 engine images (.rknn) for the RKNN inference paths of
audio.cpp, built on x86_64 with rknn-toolkit2 2.3.2
(do_quantization=False, i.e. fp16 weight storage) and validated by running the matching
audio.cpp sessions on an RK3588 board.
An .rknn is compiled for one target SoC and is not portable between chips: everything here is
target_platform="RK3588" only.
| File | Family / exec_plan role |
Source artefact | Engine |
|---|---|---|---|
rk3588/piper-lessac-w93.rknn |
piper_tts → decoder.main |
streaming decoder ONNX of en_US-lessac-medium (Piper), 93-slice window, graph z [1,192,93] + y_mask [1,1,93] |
19.6 MB |
rk3588/generator_w512.rknn |
kokoro_tts → generator.main |
Kokoro-82M (iSTFTNet) generator ONNX, 512-frame window | 93.5 MB |
rk3588/qwen3-tts/vocoder_stateful_16_fp16.rknn |
qwen3_tts → speech_decoder.main |
Qwen3-TTS-12Hz speech decoder, stateful, chunk 16 | 244.9 MB |
rk3588/qwen3-tts/vocoder_stateful_64b_fp16.rknn |
qwen3_tts → speech_decoder.main |
Qwen3-TTS-12Hz speech decoder, stateful, chunk 64 | 259.7 MB |
The file names are kept identical to the on-board production names so an exec_plan can point at
a downloaded copy unchanged.
Build provenance
Every engine was produced by the converters in tools/rknn/ of audio.cpp (the same scripts that
build from source ONNX: prepare_piper_tts_rknn_vocoder.py, the Kokoro and Qwen3-TTS prepare
scripts), from publicly published source models, with the toolkit pin the converters document
(2.3.2). SHA256SUMS pins the exact bytes. A local rebuild from the same ONNX with the same
toolkit reproduces the engine bit-for-bit in the simulator outputs (the embedded build path makes
the container hashes differ; the compiled graph does not).
Residency
The audio.cpp RKNN runtime prices these by what the driver reports
(RKNN_QUERY_MEM_SIZE), not by file size: e.g. the Qwen3-TTS pair is resident at 263 MiB and
331 MiB on RK3588, above their 245/260 MB file sizes.
Licenses
A compiled engine is a derivative of its source model weights and inherits their terms:
- Kokoro-82M: Apache-2.0.
- Qwen3-TTS-12Hz: Apache-2.0.
en_US-lessac-medium(Piper,rhasspy/piper-voices): the voice's model card points to the Lessac Blizzard 2013 dataset license, which allows non-commercial use only.
Not hosted here
The CosyVoice3 flow_estimator_seq* ladder (700–860 MB per rung, eight rungs) is deliberately
not published: build it from the shipped ONNX with tools/rknn/prepare_cosyvoice3_rknn_flow.py
for your own SoC and aperture.