Question/Issue: How to deploy/export a trained wake-word model as generic C/C++ for an unsupported proprietary DSP architecture (Jieli q32DSP).
Project ID: 345812
Context/Use case: I am developing a smart speaker based on the Jieli AC7916 SoC. The device manages local audio capture (PCM ring buffers) and connects to cloud AI backends (LLM, STT, TTS) via WebSockets. I have successfully trained a wake-word model to trigger the voice interaction and need to deploy it to this device. I am looking for guidance on how to export a generic, unoptimized C/C++ inference library that does not depend on ARM CMSIS-NN or specific hardware vector instructions, allowing me to compile it using Jieli’s proprietary compiler.
Steps Taken:
- Successfully trained a custom wake-word/KWS model in the Edge Impulse studio.
- Navigated to the Dashboard/Deployment section to configure the target device and application budget.
- Expanded the “Processor architecture” dropdown but found only ARM Cortex-M and similar standard options available.
Expected Outcome: Clear instructions on which generic architecture to select, or how to directly export a pure standard C/C++ source code package that can be ported to any custom 32-bit MCU/DSP without hardware-specific dependencies.
Actual Outcome: I am unable to find a suitable architecture option in the UI to estimate budget or export the model safely for my proprietary q32DSP core without causing compilation errors related to ARM intrinsics.
Reproducibility:
- [x] Always
- [ ] Sometimes
- [ ] Rarely
Environment:
- Platform: Jieli AC7916 SoC (q32DSP architecture)
- Build Environment Details: Jieli Studio / Code::Blocks with proprietary Jieli compiler
- OS Version: Windows 11
- Edge Impulse Version (Firmware): N/A (Custom deployment)
- Edge Impulse CLI Version: 1.35.0
- Project Version: 1.0.0
- Custom Blocks / Impulse Configuration: Audio Time Series → Audio (MFCC) → Classification (Keras)
Logs/Attachments: Attached screenshot of the “Target device” configuration modal showing the lack of generic/custom DSP options. (Please refer to the uploaded image_a45606.png).
Additional Information: My goal is similar to deploying to an Espressif ESP32, but without the luxury of hardware-specific AI optimizations. I strictly need the most vanilla C++ implementation of the neural network to integrate into a closed-source SDK. Any advice on the optimal deployment method (e.g., C++ Library export with specific flags) would be highly appreciated.