Note

Contains changes from PR #38 , so it should be merged first.

Summary

Remove unnecessary Numba usage from helpers that are either already vectorized, trivially small, or safely replaceable with NumPy.

File-by-file Changes

synctoolbox/feature/utils.py

• Replaced the column-by-column normalization loop with vectorized NumPy logic:
  • computes all column norms at once with np.linalg.norm(..., axis=0)
  • identifies below-threshold columns with a boolean mask
  • normalizes regular columns in one array operation
  • replaces below-threshold columns with the normalized all-ones unit vector

synctoolbox/feature/pitch.py

• Replaced repeated slice summation in a loop with a cumulative-sum implementation:
  • builds a prefix sum of f_square
  • computes each variable-length window sum as prefix_sum[stop] - prefix_sum[start]
  • writes all segment sums into f_pitch[midi_pitch, :] at once

synctoolbox/dtw/anchor.py

• Removed jit decorator, function is trivially small

synctoolbox/dtw/mrmsdtw.py

• Removed jit decorator, function is trivially small

synctoolbox/dtw/cost.py

• Removed unused from numba import jit.
• Removed stale commented-out #@jit(nopython=True) lines above cosine_distance and euclidean_distance.

Verification & Testing

• Full test suite passes
• Performed a custom regression test by comparing before/after function outputs with tight tolerance