Question/Issue:
I am trying to deploy a Keyword Spotting model created using the Edge Impulse tutorial on a Himax WE2 (WE-I Plus) device. The original tutorial is intended for the Arduino Nano 33 BLE Sense board, but I adapted the workflow for Himax WE2.
The model builds successfully and firmware flashes correctly onto the device. However, when running inference using the command: ‘AT+RUNIMPULSE’ the inference fails repeatedly with the following error:
AllocateTensors() failed
Failed to run impulse (-3)
AllocateTensors() failed
Failed to run impulse (-3)
AllocateTensors() failed
Failed to run impulse (-3)
This appears to be related to TensorFlow Lite Micro tensor allocation / memory allocation failure during model initialization.
I would like to understand:
- Whether additional configuration is required specifically for Himax WE2
- Whether the tensor arena size needs to be increased manually
- Whether the generated model exceeds available RAM on the device
- Whether there are known limitations for Keyword Spotting models on Himax WE2
Project ID:
1004839
Context/Use case:
I am experimenting with on-device audio inference / keyword spotting on the Himax WE2 platform using Edge Impulse.
The use case is:
- Real-time keyword spotting
- TinyML audio inference on MCU
- Deployment on Himax WE2 instead of Arduino Nano BLE Sense
I followed the official Edge Impulse Keyword Spotting tutorial: Keyword spotting - Edge Impulse Documentation
The tutorial works with Arduino Nano BLE Sense, but I am attempting to use the Himax WE2 board.
Steps Taken:
- Created a Keyword Spotting project in Edge Impulse Studio
- Followed the official Keyword Spotting tutorial workflow
- Collected/trained audio samples
- Configured MFCC processing block
- Trained neural network classifier
- Enabled quantized model option
- Deployed firmware for Himax WE-I Plus / WE2 target
- Flashed firmware successfully onto Himax WE2 board
- Opened serial terminal
- Executed: AT+RUNIMPULSE
- Observed repeated inference failure with tensor allocation errors
Expected Outcome:
The model should initialize successfully and run real-time keyword spotting inference on the Himax WE2 board.
Expected behavior:
- Successful TensorFlow Lite tensor allocation
- Model inference starts correctly
- Audio predictions displayed on serial console
Actual Outcome:
Inference initialization fails immediately with:
AllocateTensors() failed
Failed to run impulse (-3)
The error repeats continuously whenever AT+RUNIMPULSE is executed.
No inference results are produced.
Reproducibility:
- [x] Always
- [ ] Sometimes
- [ ] Rarely
Environment:
- Platform: Himax WE2
- Build Environment Details:
- OS Version: Ubuntu 22.04
- Edge Impulse Version (Firmware):
AT+INFO
- Edge Impulse firmware *
Firmware build date : May 22 2026
Firmware build time : 09:44:27
ML model author : devjonathan
ML model name : Tutorial: Responding to your voice
ML model ID : 1004839
Model deploy version : 116
Edge Impulse version : v1.93.3
Used sensor : microphone
-
Custom Blocks / Impulse Configuration: MFCC
Logs/Attachments:
AT+RUNIMPULSE
AllocateTensors() failed
Failed to run impulse (-3)
AllocateTensors() failed
Failed to run impulse (-3)
AllocateTensors() failed
Failed to run impulse (-3)
Additional Information:
Things I have not yet tried:
- Increasing
EI_CLASSIFIER_TFLITE_ARENA_SIZE - Reducing MFCC coefficients
- Reducing NN layer size
- Enabling EON Compiler
- Re-training smaller quantized model
Would appreciate guidance on:
- Recommended memory settings for Keyword Spotting on Himax WE2
- Known working configurations
- Recommended tensor arena size
- Whether this model type is officially supported on Himax WE2