Is it possible to use the Hailo compiler to convert an audio classification model to hef? Everything I see online is about object detection? Is there anywhere I can find more information about this?
Hey @gmbluth ,
Yes, it is possible to use the DFC to convert an audio classification model to a .hef
file — as long as the model meets Hailo’s compilation requirements. The reason most examples online focus on object detection is simply because that’s a common use case in vision — but Hailo supports a wide range of architectures, including those for audio.
Requirements for Audio Models
- Must be in supported format: ONNX, TensorFlow, TFLite, or Keras
- Must use operations supported by the compiler backend
- Requires static input shape (no dynamic shapes)
- Must be quantizable for Hailo’s compilation process
Many common audio classification models work well with Hailo, including YAMNet, MobileNet + Spectrogram, Keyword Spotting models, and CRNN architectures.
We’re currently working on Whisper implementation, and there’s a community project from the Hailo hackathon that involves audio processing: hailo-rpi5-examples/community_projects/TEMPO at main · OmriAx/hailo-rpi5-examples · GitHub