Help needed: RKNN conversion fails on Vosk model — "Expand input shape not correct"

[dulimov]

Hi all,

We’re currently working on porting the Vosk ASR model to RK3588 using rknn-toolkit2.
We’ve made a public repo to track our work and issues:
🔗 https://github.com/dulimov/vosk-rknn

Our conversion script:
📜 conv.py

Conversion log with errors:
📄 conv_output.log

During conversion we get the following error:

W The shape is not correct: [3,2] vs [3,1]
E Convert node[Expand_28] failed, type[Expand]!
E Catch exception when convert op[Expand_28]!
RuntimeError: Expand input shape not correct.
We've tried simplifying and modifying the ONNX model (e.g., editing shapes, simplifying nodes, removing some ops), but the error persists.

❓ Question: How can we correctly convert this model? Any guidance on how to handle this Expand shape mismatch would be very helpful.

Thanks a lot in advance!

Processing top-level node: /encoder_proj/MatMul, OpType: MatMul
Processing top-level node: /encoder_proj/Add, OpType: Add
ensure_correct_conv_attributes: Changes made to some Conv attributes (either specific or general).
--- Sanity check for saved file: vosk_models_onnx/encoder_prepared.onnx ---
ERROR: ONNX Runtime failed to load vosk_models_onnx/encoder_prepared.onnx: [ONNXRuntimeError] : 1 : FAIL : Load model from vosk_models_onnx/encoder_prepared.onnx failed:Node (/encoder_embed/conv/0/Conv) Op (Conv) [ShapeInferenceError] Attribute dilations has incorrect size

--- Processing model: decoder ---
--- Sanity check for saved file: vosk_models_onnx/decoder_prepared.onnx ---
ONNX Runtime loaded vosk_models_onnx/decoder_prepared.onnx successfully.

--- Processing model: joiner ---
--- Sanity check for saved file: vosk_models_onnx/joiner_prepared.onnx ---
ONNX Runtime loaded vosk_models_onnx/joiner_prepared.onnx successfully.

--- Creating Dummy Calibration Data ---

--- Converting to RKNN ---

Processing encoder model: vosk_models_onnx/encoder_prepared.onnx

I Loading : 0%| | 0/717 [00:00<?, ?it/s]
I Loading : 100%|██████████████████████████████████████████████| 717/717 [00:00<00:00, 13636.09it/s]

I FoldConstant : 0%| | 0/5324 [00:00<?, ?it/s]
I FoldConstant : 0%| | 1/5324 [00:00<00:16, 331.15it/s]