Performance: Optimize Mel filterbank application with sparse indices

[ysdede]

Performance: Optimize Mel filterbank application with sparse indices

What changed:

• Precomputed _fbStart and _fbEnd indices for each Mel filter in the MelSpectrogram constructor.
• Updated computeRawMel to iterate only over the non-zero range of the filterbank matrix.

Why it was needed:

• Profiling showed that the Mel filterbank application took ~78% of the total processing time.
• The filterbank matrix is ~98.5% sparse (zeros), meaning most multiplications were wasted.

Impact:

• ~3.7x speedup for process() calls.
• 5s audio processing time dropped from ~69ms to ~18.5ms.
• 10s audio processing time dropped from ~155ms to ~35ms.
• No regression in accuracy (validated against ONNX reference).

How to verify:

• Run npm test to verify functional correctness and see benchmark logs.

---

PR created automatically by Jules for task 1580082174720780620 started by @ysdede

[kiloconnect]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Files Reviewed (2 files)

• .jules/bolt.md - Documentation (new file)
• src/mel.js - Sparse filterbank optimization

Review Notes

The sparse filterbank optimization is well-implemented:

• Precomputes start/end indices in constructor for each mel bin
• Uses these indices in computeRawMel to iterate only over non-zero filterbank values
• This avoids ~98.5% of unnecessary multiplications mentioned in the documentation

No security vulnerabilities, runtime errors, or logic bugs were identified.