Neural Guitar Pedal – Optimizing NAM for Daisy Seed Arm Cortex-M7
2 days ago
- #embedded-systems
- #audio-processing
- #neural-networks
- NAM loader developed for Electrosmith Daisy Seed, an ARM Cortex-M7 board popular in DSP-based audio products.
- Challenges included adapting NeuralAmpModelerCore for embedded hardware with tight memory and real-time constraints.
- Initial tests showed processing 2 seconds of audio took over 5 seconds, highlighting inefficiencies.
- Optimizations focused on model size, compute efficiency, and model loading, leading to significant improvements.
- A new compact binary model format (.namb) was created for embedded devices, simplifying model transfer and loading.
- Post-optimization, processing time reduced to 1.5 seconds for the same audio length, with additional headroom for effects.
- Insights from this project are informing the design of Architecture 2 (A2) and the Slimmable NAM approach.
- Source code and tools from the project are being published for community use and further development.