audio-numerics-review

Reviews audio DSP code for numerical correctness. Use when the user asks to review, audit, or check DSP code for correctness issues — filters, feedback loops, fixed-point arithmetic, accumulation loop

SKILL.md

GitHub

Audio Numerics Review

Floating-point is not real math. IEEE 754 edge cases — denormals, NaN propagation, catastrophic cancellation, precision loss — silently produce wrong audio output without crashing. You won't find them with a debugger; you find them by knowing where to look.

Step 1 — Find numeric-critical paths

Locate every function doing floating-point or integer DSP: filter state loops, feedback delay lines, envelope followers, RMS/energy accumulators, pitch detectors, fixed-point paths (CMSIS-DSP, SIMD int16/int32), any division whose denominator may be zero.

Step 2 — Scan for violations

Violation Where to look Symptom
Denormal floats IIR state vars, reverb tanks, delay feedback Sudden 100× CPU spike when signal decays
NaN / Inf Divisions, sqrt, acos/asin, peak normalizers Silence or noise that can't be traced
Catastrophic cancellation DC-blocking filters, near-Nyquist biquads, first-order diff Noise floor jumps, loss of precision
Fixed-point overflow Q15/Q31 multiply, CMSIS paths, SIMD int16 Audible distortion / wrapping
Precision loss RMS/energy sums, mixing buses, long loops Accumulated DC offset, incorrect levels
FTZ/DAZ absent prepareToPlay/stream-open — no mode setup Denormals hit on some hosts but not others

For code examples and platform-specific notes see references/numerics-pitfalls.md.

Step 3 — Write the review

## Audio Numerics Review: `[file / function]`

### Verdict
[Safe | Has critical violations | Warnings only] — one sentence summary

### Critical Violations
**[Category]: [description]**
`file:line` — `offending code`
Why: one sentence on the numeric risk
Fix: concrete suggestion

### Warnings
[same format]

### What's Done Well
[guarded divisions, double accumulators, FTZ setup, etc.]

### Recommended Fixes (priority order)
1. ...

Quick fix table

Violation Safe alternative
Denormal in feedback ScopedNoDenormals; FTZ+DAZ in prepareToPlay; DC-offset clamp
Division by zero (denom > 1e-6f) ? num/denom : fallback
sqrt/acos/asin unclamped std::sqrt(std::max(0.f, x)), std::clamp inputs
NaN propagating silently assert(!std::isnan(y)) in debug; validate at entry points
Cancellation in coefficients Compute coefficients in double; use numerically stable forms
Fixed-point overflow Promote to wider type before multiply; __SSAT, CMSIS saturating ops
Precision loss in long sum double accumulator or Kahan compensated summation
FTZ/DAZ absent Set explicitly in prepareToPlay; use ScopedNoDenormals per-block