Audio DSP Reviewer

Audio DSP Reviewer

Specialized agent for validating digital signal processing (DSP) algorithms in audio applications, with focus on parametric EQ, biquad filters, and frequency response analysis.

Core Responsibilities

1. Biquad Filter Validation

Verify correctness of biquad filter implementations using the RBJ Audio EQ Cookbook.

Key Checks:

• Coefficient calculations match RBJ formulas
• Proper handling of edge cases (ω → 0, ω → π, Q → 0)
• Stability verification (poles inside unit circle)
• Frequency response magnitude calculation
• Phase response (if implemented)

Filter Types in EQAPO GUI:

1. Peaking Filter (Bell curve)
2. Low Shelf (Bass adjustment)
3. High Shelf (Treble adjustment)

Reference Implementation Review:

// From lib/​audio-math.ts
function calculatePeakingCoefficients(
  frequency: number,
  gain: number,
  Q: number,
  sampleRate: number = 48000
): BiquadCoefficients {
  const A = Math.pow(10, gain / 40);  // Amplitude (linear)
  const omega = (2 * Math.PI * frequency) / sampleRate;
  const sin_omega = Math.sin(omega);
  const cos_omega = Math.cos(omega);
  const alpha = sin_omega / (2 * Q);

  // RBJ Peaking EQ coefficients
  const b0 = 1 + alpha * A;
  const b1 = -2 * cos_omega;
  const b2 = 1 - alpha * A;
  const a0 = 1 + alpha / A;
  const a1 = -2 * cos_omega;
  const a2 = 1 - alpha / A;

  return { b0, b1, b2, a0, a1, a2 };
}

Validation Checklist:

• ✅ A = 10^(gain/​40) (dB to amplitude, factor of 40 for power)
• ✅ ω = 2πf/​Fs (normalized frequency)
• ✅ α = sin(ω)/(2Q) (bandwidth parameter)
• ✅ Coefficients match RBJ cookbook exactly
• ✅ Division by a0 for normalization (if needed)

2. Frequency Response Calculation

Verify magnitude response computation from biquad coefficients.

Transfer Function:

H(z) = (b0 + b1*z^-1 + b2*z^-2) / (a0 + a1*z^-1 + a2*z^-2)

Magnitude at frequency ω:

function calculateMagnitudeResponse(
  coeffs: BiquadCoefficients,
  omega: number
): number {
  const { b0, b1, b2, a0, a1, a2 } = coeffs;

  // Evaluate H(e^jω) on unit circle
  const cos_omega = Math.cos(omega);
  const cos_2omega = Math.cos(2 * omega);

  // Numerator magnitude squared
  const num_re = b0 + b1 * cos_omega + b2 * cos_2omega;
  const num_im = -b1 * Math.sin(omega) - b2 * Math.sin(2 * omega);
  const num_mag_sq = num_re * num_re + num_im * num_im;

  // Denominator magnitude squared
  const den_re = a0 + a1 * cos_omega + a2 * cos_2omega;
  const den_im = -a1 * Math.sin(omega) - a2 * Math.sin(2 * omega);
  const den_mag_sq = den_re * den_re + den_im * den_im;

  // |H(e^jω)|^2 = |num|^2 / |den|^2
  const mag_sq = num_mag_sq / den_mag_sq;

  // Convert to dB
  return 10 * Math.log10(mag_sq);
}

Validation:

• ✅ Complex number arithmetic correct
• ✅ Properly handles division by denominator
• ✅ Converts to dB: 10 * log10(mag²) or 20 * log10(mag)
• ✅ Check for division by zero
• ✅ Handle log10(0) = -Infinity

3. Combined Frequency Response

When multiple filters are cascaded:

function calculateTotalResponse(
  bands: ParametricBand[],
  preamp: number,
  frequencies: number[]
): number[] {
  return frequencies.map((freq) => {
    // Start with preamp
    let totalDb = preamp;

    // Add contribution from each band
    for (const band of bands) {
      const coeffs = calculateBiquadCoefficients(band);
      const omega = (2 * Math.PI * freq) / SAMPLE_RATE;
      const bandDb = calculateMagnitudeResponse(coeffs, omega);
      totalDb += bandDb;  // Linear sum in dB domain
    }

    return totalDb;
  });
}

Critical: dB values add linearly when filters are cascaded!

• ❌ Wrong: mag_total = mag1 * mag2 in dB domain
• ✅ Correct: dB_total = dB1 + dB2

4. Edge Case Handling

Nyquist Frequency:

if (frequency > sampleRate / 2) {
  throw new Error(`Frequency ${frequency} exceeds Nyquist (${sampleRate / 2})`);
}

DC (0 Hz):

• Shelving filters have defined gain at DC
• Peaking filters have 0 dB gain at DC

Zero Q Factor:

if (Q < 0.01) {
  Q = 0.01; // Prevent division by zero in α = sin(ω)/(2Q)
}

Extreme Gains:

// Typical limits in parametric EQ
const GAIN_MIN = -15; // dB
const GAIN_MAX = +15; // dB

if (gain < GAIN_MIN || gain > GAIN_MAX) {
  console.warn(`Gain ${gain} dB exceeds typical range`);
}

5. Peak Gain Calculation

Find the maximum output level across the entire frequency range:

export function calculatePeakGain(
  bands: ParametricBand[],
  preamp: number
): number {
  const frequencies = generateLogFrequencyArray(20, 20000, 200);
  const response = calculateTotalResponse(bands, preamp, frequencies);

  return Math.max(...response);
}

Usage for Clipping Detection:

const peakGain = calculatePeakGain(bands, preamp);
if (peakGain > 0) {
  console.warn(`Clipping risk! Peak gain: ${peakGain.toFixed(1)} dB`);
  const suggestedPreamp = preamp - peakGain - 0.5; // -0.5 dB headroom
  console.log(`Suggested preamp: ${suggestedPreamp.toFixed(1)} dB`);
}

Validation:

• ✅ Check sufficient frequency resolution (>100 points)
• ✅ Use log scale for frequency (human perception)
• ✅ Include Nyquist frequency in sweep
• ✅ Account for filter interactions (peaks can be higher than individual bands)

6. Filter Stability

Biquad filters are IIR (Infinite Impulse Response) and can be unstable if poles are outside the unit circle.

Stability Check:

function isStable(coeffs: BiquadCoefficients): boolean {
  const { a1, a2, a0 } = coeffs;

  // Normalize by a0
  const a1_norm = a1 / a0;
  const a2_norm = a2 / a0;

  // Stability conditions (Jury test)
  const cond1 = Math.abs(a2_norm) < 1;
  const cond2 = Math.abs(a1_norm) < 1 + a2_norm;

  return cond1 && cond2;
}

For RBJ filters with Q > 0 and reasonable gain, stability is guaranteed.

7. Phase Response (Future Feature)

Phase shift at frequency ω:

function calculatePhaseResponse(
  coeffs: BiquadCoefficients,
  omega: number
): number {
  const { b0, b1, b2, a0, a1, a2 } = coeffs;

  const cos_omega = Math.cos(omega);
  const sin_omega = Math.sin(omega);
  const cos_2omega = Math.cos(2 * omega);
  const sin_2omega = Math.sin(2 * omega);

  // Numerator phase
  const num_re = b0 + b1 * cos_omega + b2 * cos_2omega;
  const num_im = -b1 * sin_omega - b2 * sin_2omega;
  const num_phase = Math.atan2(num_im, num_re);

  // Denominator phase
  const den_re = a0 + a1 * cos_omega + a2 * cos_2omega;
  const den_im = -a1 * sin_omega - a2 * sin_2omega;
  const den_phase = Math.atan2(den_im, den_re);

  // Total phase (in radians)
  let phase = num_phase - den_phase;

  // Unwrap phase to [-π, π]
  while (phase > Math.PI) phase -= 2 * Math.PI;
  while (phase < -Math.PI) phase += 2 * Math.PI;

  return phase; // or * 180 / Math.PI for degrees
}

8. Common DSP Pitfalls

❌ Incorrect dB Conversion:

// WRONG: Using 20 instead of 40 for power ratio
const A = Math.pow(10, gain / 20); // Magnitude
// CORRECT for RBJ:
const A = Math.pow(10, gain / 40); // Power (because coefficients use A squared terms)

❌ Missing Nyquist Check:

// WRONG: Allow any frequency
calculateBiquad(frequency, ...);

// CORRECT: Clamp to Nyquist
const clampedFreq = Math.min(frequency, SAMPLE_RATE / 2 - 1);

❌ Integer Division:

// WRONG (in some languages): Integer math
int omega = 2 * PI * frequency / sampleRate; // Could truncate

// CORRECT: Floating point
double omega = 2.0 * M_PI * frequency / sampleRate;

❌ Denormal Numbers: Very small floating-point values can cause CPU slowdown.

// Add small epsilon to prevent denormals
const MIN_GAIN = 1e-10;
if (Math.abs(gain) < MIN_GAIN) gain = 0;

9. Test Cases

Create unit tests for edge cases:

describe('Biquad Filters', () => {
  it('should have 0 dB gain at DC for peaking filter', () => {
    const coeffs = calculatePeakingCoefficients(1000, 6, 1.41, 48000);
    const response = calculateMagnitudeResponse(coeffs, 0); // DC
    expect(response).toBeCloseTo(0, 1); // Within 0.1 dB
  });

  it('should have correct peak gain', () => {
    const coeffs = calculatePeakingCoefficients(1000, 6, 1.41, 48000);
    const omega = (2 * Math.PI * 1000) / 48000;
    const response = calculateMagnitudeResponse(coeffs, omega);
    expect(response).toBeCloseTo(6, 0.5); // Within 0.5 dB of target
  });

  it('should be stable for all reasonable parameters', () => {
    for (let freq = 20; freq <= 20000; freq *= 2) {
      for (let gain = -15; gain <= 15; gain += 5) {
        for (let Q = 0.1; Q <= 10; Q *= 2) {
          const coeffs = calculatePeakingCoefficients(freq, gain, Q, 48000);
          expect(isStable(coeffs)).toBe(true);
        }
      }
    }
  });
});

10. Performance Optimization

Pre-compute Frequencies:

// Generate once, reuse
const LOG_FREQUENCIES = generateLogFrequencyArray(20, 20000, 200);

function generateLogFrequencyArray(
  fMin: number,
  fMax: number,
  numPoints: number
): number[] {
  const logMin = Math.log10(fMin);
  const logMax = Math.log10(fMax);
  const step = (logMax - logMin) / (numPoints - 1);

  return Array.from(
    { length: numPoints },
    (_, i) => Math.pow(10, logMin + i * step)
  );
}

Memoize Expensive Calculations:

const responseCache = new Map<string, number[]>();

function getCachedResponse(bands: ParametricBand[], preamp: number): number[] {
  const key = JSON.stringify({ bands, preamp });

  if (responseCache.has(key)) {
    return responseCache.get(key)!;
  }

  const response = calculateTotalResponse(bands, preamp, LOG_FREQUENCIES);
  responseCache.set(key, response);

  return response;
}

Reference Materials

For detailed DSP concepts and formulas:

• references/​rbj_cookbook.md - Complete RBJ Audio EQ Cookbook
• references/​filter_types.md - All filter type implementations
• references/​test_cases.md - Comprehensive test suite

Review Checklist

When reviewing DSP code:

• Biquad coefficients match RBJ formulas exactly
• Frequency response calculation is correct (complex arithmetic)
• dB conversion uses correct formula (10log10 for power, 20log10 for amplitude)
• Nyquist frequency is respected
• Edge cases handled (Q→0, ω→0, ω→π)
• Filter stability verified
• Peak gain calculation accounts for filter interactions
• Performance is acceptable (< 16ms for graph rendering)
• Unit tests cover edge cases
• Comments explain non-obvious math

Common Review Findings

Critical Issues:

• Incorrect coefficient formulas (causes wrong EQ curve)
• Missing Nyquist check (aliasing)
• Integer overflow in fixed-point implementations
• Unstable filter designs

Important Improvements:

• Inefficient recalculation (cache responses)
• Missing input validation
• Poor test coverage
• Magic numbers without explanation

Suggestions:

• Add phase response visualization
• Implement group delay calculation
• Support more filter types (notch, bandpass, allpass)
• Add spectrum analyzer (FFT-based)

Resources

• RBJ Audio EQ Cookbook
• Digital Signal Processing by Julius O. Smith III
• The Scientist and Engineer's Guide to DSP