| 1 | import Foundation |
| 2 | |
| 3 | /// Monophonic fundamental-frequency estimator using the YIN algorithm |
| 4 | /// (de Cheveigné & Kawahara, 2002) with parabolic interpolation. |
| 5 | final class YINDetector { |
| 6 | var sampleRate: Double |
| 7 | let windowSize: Int |
| 8 | private let threshold: Float |
| 9 | private let rmsGate: Float |
| 10 | private let minClarity: Double |
| 11 | |
| 12 | // Scratch buffers reused across calls to avoid per-frame allocation. |
| 13 | private var diff: [Float] |
| 14 | private var cmnd: [Float] |
| 15 | |
| 16 | init(sampleRate: Double, windowSize: Int = 2048, threshold: Float = 0.15, |
| 17 | rmsGate: Float = 0.006, minClarity: Double = 0.5) { |
| 18 | self.sampleRate = sampleRate |
| 19 | self.windowSize = windowSize |
| 20 | self.threshold = threshold |
| 21 | self.rmsGate = rmsGate |
| 22 | self.minClarity = minClarity |
| 23 | self.diff = [Float](repeating: 0, count: windowSize / 2) |
| 24 | self.cmnd = [Float](repeating: 1, count: windowSize / 2) |
| 25 | } |
| 26 | |
| 27 | struct Result { |
| 28 | let frequency: Double |
| 29 | let clarity: Double // 0…1, higher == more periodic |
| 30 | let level: Double // RMS, 0…1 |
| 31 | } |
| 32 | |
| 33 | /// Returns a pitch estimate for `samples` (must be `windowSize` long) or nil |
| 34 | /// if the signal is too quiet or not periodic enough. |
| 35 | func detect(_ samples: UnsafePointer<Float>, count: Int) -> Result? { |
| 36 | guard count >= windowSize else { return nil } |
| 37 | let half = windowSize / 2 |
| 38 | |
| 39 | // --- RMS gate: ignore near-silence --- |
| 40 | var sumSq: Float = 0 |
| 41 | for i in 0..<windowSize { let v = samples[i]; sumSq += v * v } |
| 42 | let rms = (sumSq / Float(windowSize)).squareRoot() |
| 43 | if rms < rmsGate { return nil } |
| 44 | let level = Double(min(1, rms * 6)) |
| 45 | |
| 46 | // --- Difference function --- |
| 47 | diff[0] = 0 |
| 48 | for tau in 1..<half { |
| 49 | var sum: Float = 0 |
| 50 | var j = 0 |
| 51 | while j < half { |
| 52 | let d = samples[j] - samples[j + tau] |
| 53 | sum += d * d |
| 54 | j += 1 |
| 55 | } |
| 56 | diff[tau] = sum |
| 57 | } |
| 58 | |
| 59 | // --- Cumulative mean normalized difference --- |
| 60 | cmnd[0] = 1 |
| 61 | var running: Float = 0 |
| 62 | for tau in 1..<half { |
| 63 | running += diff[tau] |
| 64 | cmnd[tau] = running > 0 ? diff[tau] * Float(tau) / running : 1 |
| 65 | } |
| 66 | |
| 67 | // --- Absolute threshold --- |
| 68 | var tauEst = -1 |
| 69 | var tau = 2 |
| 70 | while tau < half - 1 { |
| 71 | if cmnd[tau] < threshold { |
| 72 | while tau + 1 < half && cmnd[tau + 1] < cmnd[tau] { tau += 1 } |
| 73 | tauEst = tau |
| 74 | break |
| 75 | } |
| 76 | tau += 1 |
| 77 | } |
| 78 | guard tauEst > 0 else { return nil } |
| 79 | |
| 80 | // --- Parabolic interpolation around the dip --- |
| 81 | let betterTau = parabolicRefine(tauEst) |
| 82 | guard betterTau > 0 else { return nil } |
| 83 | let freq = sampleRate / betterTau |
| 84 | guard freq >= 40, freq <= 2000 else { return nil } |
| 85 | |
| 86 | let clarity = Double(max(0, min(1, 1 - cmnd[tauEst]))) |
| 87 | guard clarity >= minClarity else { return nil } |
| 88 | return Result(frequency: freq, clarity: clarity, level: level) |
| 89 | } |
| 90 | |
| 91 | private func parabolicRefine(_ tau: Int) -> Double { |
| 92 | let half = windowSize / 2 |
| 93 | let x0 = tau > 1 ? tau - 1 : tau |
| 94 | let x2 = tau + 1 < half ? tau + 1 : tau |
| 95 | if x0 == tau { return Double(cmnd[tau] <= cmnd[x2] ? tau : x2) } |
| 96 | if x2 == tau { return Double(cmnd[tau] <= cmnd[x0] ? tau : x0) } |
| 97 | let s0 = cmnd[x0], s1 = cmnd[tau], s2 = cmnd[x2] |
| 98 | let denom = 2 * (2 * s1 - s2 - s0) |
| 99 | if denom == 0 { return Double(tau) } |
| 100 | return Double(tau) + Double(s2 - s0) / Double(denom) |
| 101 | } |
| 102 | } |