This repository documents development of IR-Approx.
For a more detailed introduction, see PROPOSAL.md.
For a basic demo, run python main.py, which will generate the output seen above.
The Trainer class defined in main.py trains an IIR kernel of increasing size to approximate an impulse response (magnitudes only) until an error threshold is met. For the demo response, the algorithm takes 7.48 seconds to reach the error threshold on an M2 Max.
Run python benchmark.py to run the trainer with increasing max_poles, which demonstrates the time complexity of the training algorithm. Current results look to be O(n^2).
notebook.ipynb demonstrates training on a real-world impulse response (from a guitar speaker cabinet), as well as Trainer's function for generating a C++ header that implements the designed filter with second-order sections.
- Partitioning such that training on larger impulse responses won't have reduced frequency resolution
- A time-domain training step that accounts for the delayed onset of a real-world impulse response
- Parallelization and other optimizations to reduce the time complexity of training