Fastest and only macOS Dictation app with on-device STT and custom trained AI enhancement model. Windows, iOS testflight! Android, Linux waitlist open! A local Wispr Flow alternative. Follow https://x.com/fluidvoiceapp for more
test(meeting): Phase 0 clock-domain probe — SCStream PTS is mach-anchored
Shreeraman A Kfd083bd1033c
diff --git a/Fluid.xcodeproj/project.pbxproj b/Fluid.xcodeproj/project.pbxprojindex 616ebbc0..8544115b 100644--- a/Fluid.xcodeproj/project.pbxproj+++ b/Fluid.xcodeproj/project.pbxproj@@ -37,6 +37,7 @@E5EC00000000000000000042 /* VoiceProcessingDuckingProbeTests.swift in Sources */ = {isa = PBXBuildFile; fileRef = E5EC00000000000000000041 /* VoiceProcessingDuckingProbeTests.swift */; };E5EC00000000000000000052 /* VoiceProcessingSpeechDuckingTests.swift in Sources */ = {isa = PBXBuildFile; fileRef = E5EC00000000000000000051 /* VoiceProcessingSpeechDuckingTests.swift */; };E5EC00000000000000000062 /* VoiceProcessingAsrQualityTests.swift in Sources */ = {isa = PBXBuildFile; fileRef = E5EC00000000000000000061 /* VoiceProcessingAsrQualityTests.swift */; };+ E5EC00000000000000000072 /* ClockDomainProbeTests.swift in Sources */ = {isa = PBXBuildFile; fileRef = E5EC00000000000000000071 /* ClockDomainProbeTests.swift */; };A6E2246FA3744C7B3E2DB91C /* MeetingLiveTranscriptionTests.swift in Sources */ = {isa = PBXBuildFile; fileRef = 355C1248D5386D1F815EB7D7 /* MeetingLiveTranscriptionTests.swift */; };2952403C87FFDE41C1BC85D5 /* MeetingLiveReplayTests.swift in Sources */ = {isa = PBXBuildFile; fileRef = 7BE63EB94BB280F38F5CB028 /* MeetingLiveReplayTests.swift */; };C0DE63600000000000000002 /* AudioEngineRetirementDrainTests.swift in Sources */ = {isa = PBXBuildFile; fileRef = C0DE63600000000000000001 /* AudioEngineRetirementDrainTests.swift */; };@@ -100,6 +101,7 @@E5EC00000000000000000041 /* VoiceProcessingDuckingProbeTests.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = VoiceProcessingDuckingProbeTests.swift; sourceTree = "<group>"; };E5EC00000000000000000051 /* VoiceProcessingSpeechDuckingTests.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = VoiceProcessingSpeechDuckingTests.swift; sourceTree = "<group>"; };E5EC00000000000000000061 /* VoiceProcessingAsrQualityTests.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = VoiceProcessingAsrQualityTests.swift; sourceTree = "<group>"; };+ E5EC00000000000000000071 /* ClockDomainProbeTests.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = ClockDomainProbeTests.swift; sourceTree = "<group>"; };355C1248D5386D1F815EB7D7 /* MeetingLiveTranscriptionTests.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = MeetingLiveTranscriptionTests.swift; sourceTree = "<group>"; };7BE63EB94BB280F38F5CB028 /* MeetingLiveReplayTests.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = MeetingLiveReplayTests.swift; sourceTree = "<group>"; };C0DE63600000000000000001 /* AudioEngineRetirementDrainTests.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = AudioEngineRetirementDrainTests.swift; sourceTree = "<group>"; };@@ -211,6 +213,7 @@E5EC00000000000000000041 /* VoiceProcessingDuckingProbeTests.swift */,E5EC00000000000000000051 /* VoiceProcessingSpeechDuckingTests.swift */,E5EC00000000000000000061 /* VoiceProcessingAsrQualityTests.swift */,+ E5EC00000000000000000071 /* ClockDomainProbeTests.swift */,355C1248D5386D1F815EB7D7 /* MeetingLiveTranscriptionTests.swift */,7BE63EB94BB280F38F5CB028 /* MeetingLiveReplayTests.swift */,C0DE63600000000000000001 /* AudioEngineRetirementDrainTests.swift */,@@ -407,6 +410,7 @@E5EC00000000000000000042 /* VoiceProcessingDuckingProbeTests.swift in Sources */,E5EC00000000000000000052 /* VoiceProcessingSpeechDuckingTests.swift in Sources */,E5EC00000000000000000062 /* VoiceProcessingAsrQualityTests.swift in Sources */,+ E5EC00000000000000000072 /* ClockDomainProbeTests.swift in Sources */,A6E2246FA3744C7B3E2DB91C /* MeetingLiveTranscriptionTests.swift in Sources */,2952403C87FFDE41C1BC85D5 /* MeetingLiveReplayTests.swift in Sources */,C0DE63600000000000000002 /* AudioEngineRetirementDrainTests.swift in Sources */,diff --git a/MIC_CAPTURE_MIGRATION_PLAN.md b/MIC_CAPTURE_MIGRATION_PLAN.mdindex e6a434ca..143c8961 100644--- a/MIC_CAPTURE_MIGRATION_PLAN.md+++ b/MIC_CAPTURE_MIGRATION_PLAN.md@@ -118,21 +118,24 @@ establishment. Echo reference alignment is gone.does (`ASRService.swift:4617-4620`). Never `mach_absolute_time()` at callback time.- Invalid `hostTime`: extrapolate from the last valid stamp plus output duration, **capped at 500 ms**;beyond that, record a discontinuity and resync.-- **The constant-offset model is contingent.** This repo measured ~6.4 ppm drift, ~23 ms/hour- (`MeetingProcessingPipeline.swift:722-724`); a fixed offset is ~46 ms wrong at two hours, consuming- the whole budget. Phase 0 decides: same clock domain ⇒ offset is identity and no model is needed;- different domains ⇒ a linear model with periodic re-measurement, not a constant.+- **Resolved by Phase 0 (see §9): same clock domain.** SCStream PTS regressed against mach at+ −0.133 ppm ± 0.628 over 24k samples — the offset is effectively identity and no drift model is+ needed. The ~6.4 ppm drift this repo previously measured (`MeetingProcessingPipeline.swift:722-724`)+ was between the two *audio device* clocks through their samples, not between SCStream's stamps and+ mach; it remains real for echo-reference alignment, which VPIO sessions no longer perform.**Budget: ±50 ms**, justified by turn ordering alone. Stated as a claim to be tested, not derived —v3 wrongly cited the 0.15 s `overlapsRemote` threshold, which is a minimum-overlap predicate.## 9. Phases and gates-**Phase 0 — clock domain (unattended).** Compare *rates*, not arrival deltas: regress SCStream PTS-against wall time and microphone PTS against wall time over 30 minutes. *Gate:* report both slopes in-ppm with confidence intervals. |slope difference| < 20 ppm ⇒ constant offset acceptable; ≥ 20 ppm ⇒-§8's linear model is mandatory. Subtracting `mach_absolute_time()` at arrival is **not** the test — it-conflates queueing jitter with clock domain.+**Phase 0 — clock domain. DONE, 2026-08-15.** Regressed SCStream audio PTS against mach host time+over 8 minutes, 24,027 samples (`ClockDomainProbeTests`, `FLUIDVOICE_CLOCK_PROBE`): slope+**−0.133 ppm, 95% CI ±0.628** — statistically zero, 150× inside the 20 ppm gate. SCStream PTS is+mach-anchored, so both capture paths already share a clock domain: **constant offset only; the+linear model in §8 is not needed.** Residual std 6.9 ms is arrival jitter, confirming that per-sample+arrival deltas would have been the wrong instrument. Caveats: one machine, and the probe needs an+unlocked display in a foreground session (`SCShareableContent` reports zero displays otherwise).**Phase 1 — `MeetingMicrophoneCapture`, standalone.** *Gate*, over 10 minutes on the built-in mic**and** one Bluetooth device: valid `hostTime` on ≥99% of buffers; output Float32 LPCM (else thediff --git a/Tests/FluidDictationIntegrationTests/ClockDomainProbeTests.swift b/Tests/FluidDictationIntegrationTests/ClockDomainProbeTests.swiftnew file mode 100644index 00000000..f1cd5241--- /dev/null+++ b/Tests/FluidDictationIntegrationTests/ClockDomainProbeTests.swift@@ -0,0 +1,151 @@+@testable import FluidVoice_Debug+import AVFoundation+import CoreMedia+import Foundation+import ScreenCaptureKit+import XCTest++/// Phase 0 of the mic-capture migration: is SCStream's audio PTS in the mach host clock domain?+///+/// The migration puts the microphone on AVAudioEngine, whose tap stamps are mach host time by+/// definition. Whether cross-track alignment needs a constant offset or a drift model depends+/// entirely on SCStream's clock: regress its audio PTS against mach time over a long capture and+/// read the slope. Arrival jitter perturbs individual samples, not a long-run slope.+///+/// Gate (MIC_CAPTURE_MIGRATION_PLAN.md §9): |slope − 1| < 20 ppm ⇒ constant offset is acceptable;+/// ≥ 20 ppm ⇒ the linear model in §8 is mandatory.+///+/// Run with FLUIDVOICE_CLOCK_PROBE=<minutes>.+final class ClockDomainProbeTests: XCTestCase {+ func testScreenCaptureAudioClockAgainstMachHostTime() async throws {+ guard let minutesString = ProcessInfo.processInfo.environment["FLUIDVOICE_CLOCK_PROBE"],+ let minutes = Double(minutesString), minutes > 0+ else {+ throw XCTSkip("FLUIDVOICE_CLOCK_PROBE not set")+ }++ let content = try await SCShareableContent.excludingDesktopWindows(false, onScreenWindowsOnly: true)+ guard let display = content.displays.first else { throw XCTSkip("no display") }+ let filter = SCContentFilter(display: display, excludingApplications: [], exceptingWindows: [])++ let configuration = SCStreamConfiguration()+ configuration.width = 2+ configuration.height = 2+ configuration.minimumFrameInterval = CMTime(value: 1, timescale: 1)+ configuration.queueDepth = 3+ configuration.showsCursor = false+ configuration.capturesAudio = true+ configuration.sampleRate = 48_000+ configuration.channelCount = 2+ configuration.excludesCurrentProcessAudio = true++ let recorder = ClockPairRecorder()+ let output = ClockProbeOutput(recorder: recorder)+ let stream = SCStream(filter: filter, configuration: configuration, delegate: nil)+ try stream.addStreamOutput(output, type: .audio, sampleHandlerQueue: DispatchQueue(label: "clock.probe"))+ try await stream.startCapture()+ defer { Task { try? await stream.stopCapture() } }++ let seconds = minutes * 60+ print(String(format: "[clock] capturing for %.0f minutes…", minutes))+ // Progress report every minute so a truncated run still yields a usable slope series.+ var elapsed = 0.0+ while elapsed < seconds {+ let step = min(60.0, seconds - elapsed)+ try await Task.sleep(nanoseconds: UInt64(step * 1_000_000_000))+ elapsed += step+ if let fit = await recorder.regression() {+ print(String(+ format: "[clock] t=%4.0fs n=%6d slope=%+.3f ppm residualStd=%.3f ms ptsSpan=%.1fs",+ elapsed, fit.count, fit.slopePPM, fit.residualStdMs, fit.ptsSpanSeconds+ ))+ }+ }++ guard let fit = await recorder.regression(), fit.count > 1000 else {+ XCTFail("insufficient samples for a slope estimate")+ return+ }+ print(String(+ format: "\n[clock] FINAL n=%d slope=%+.3f ppm (95%% CI ±%.3f) residualStd=%.3f ms gate=%@",+ fit.count, fit.slopePPM, fit.slopeCI95PPM, fit.residualStdMs,+ abs(fit.slopePPM) < 20 ? "CONSTANT OFFSET OK" : "LINEAR MODEL REQUIRED"+ ))+ }+}++/// Records (mach seconds, SCStream PTS seconds) pairs and fits PTS = a + b·mach by least squares.+private actor ClockPairRecorder {+ struct Fit {+ let count: Int+ let slopePPM: Double+ let slopeCI95PPM: Double+ let residualStdMs: Double+ let ptsSpanSeconds: Double+ }++ private var machTimes: [Double] = []+ private var ptsTimes: [Double] = []++ func record(mach: Double, pts: Double) {+ self.machTimes.append(mach)+ self.ptsTimes.append(pts)+ }++ func regression() -> Fit? {+ let n = self.machTimes.count+ guard n > 10 else { return nil }+ let x0 = self.machTimes[0]+ let y0 = self.ptsTimes[0]+ let xs = self.machTimes.map { $0 - x0 }+ let ys = self.ptsTimes.map { $0 - y0 }+ let sumX = xs.reduce(0, +)+ let sumY = ys.reduce(0, +)+ let meanX = sumX / Double(n)+ let meanY = sumY / Double(n)+ var sxx = 0.0+ var sxy = 0.0+ for i in 0..<n {+ let dx = xs[i] - meanX+ sxx += dx * dx+ sxy += dx * (ys[i] - meanY)+ }+ guard sxx > 0 else { return nil }+ let slope = sxy / sxx+ var sse = 0.0+ let intercept = meanY - slope * meanX+ for i in 0..<n {+ let r = ys[i] - (intercept + slope * xs[i])+ sse += r * r+ }+ let residualVariance = sse / Double(max(1, n - 2))+ let slopeStdErr = (residualVariance / sxx).squareRoot()+ return Fit(+ count: n,+ slopePPM: (slope - 1) * 1_000_000,+ slopeCI95PPM: 1.96 * slopeStdErr * 1_000_000,+ residualStdMs: residualVariance.squareRoot() * 1_000,+ ptsSpanSeconds: ys.last ?? 0+ )+ }+}++private final class ClockProbeOutput: NSObject, SCStreamOutput {+ private let recorder: ClockPairRecorder+ private static let timebase: Double = {+ var info = mach_timebase_info_data_t()+ mach_timebase_info(&info)+ return Double(info.numer) / Double(info.denom) / 1_000_000_000+ }()++ init(recorder: ClockPairRecorder) { self.recorder = recorder }++ func stream(_ stream: SCStream, didOutputSampleBuffer sampleBuffer: CMSampleBuffer, of type: SCStreamOutputType) {+ guard type == .audio else { return }+ let pts = CMSampleBufferGetPresentationTimeStamp(sampleBuffer)+ guard pts.isValid else { return }+ let mach = Double(mach_absolute_time()) * Self.timebase+ let seconds = pts.seconds+ Task { [recorder] in await recorder.record(mach: mach, pts: seconds) }+ }+}