Hello,
I have fine tuned a whisper-base.en with LoRA adaptions using a friends 230 Dysarthric speech WAVs.
I then merged them before exporting to ONNX before finally compiling into an .hef file for loading on my pi5/Hailo-10h HAT.
The problem appears to be at ONNX stage which when compiling runs into the following error:
[ERROR] Compilation failed: Invalid kernel shape for base conv layer base_conv1 (translated from /enc/conv1/Conv). Either the input shape doesn’t match the kernel shape, or the calculated groups number doesn’t match the expected ratio between kernel shape and input shape. Kernel features: 80 Input features: 3000 Groups: 37
I’ve seen mention of it having a 30 second default input frame which understandably will create a massive Mel Spectrogram and wondered if this the problem?
Does anyone know how to restrict this to a smaller input frame with a smaller Mel spectrogram ‘expectation’ ? And also any views on whether this is the real reason why I get the compilation error?
Thanks In advance
Pete