Swift AVFoundation timing info for audio measurements - ios

Question

I am creating an application that will take an audio measurement by playing some stimulus data and recording the microphone input, and then analysing the data.
I am having trouble accounting for the time taken to initialise and start the audio engine, as this varies each time and is also dependant on the hardware used, etc.
So, I have an audio engine and have installed a Tap the hardware input, with input 1 being the microphone recording, and input 2 being a reference input (also from the hardware). The output is physically Y-Split and fed back into input 2.
The app initialises the engine, plays the stimulus audio plus 1 second of silence (to allow propagation time for the microphone to record the whole signal back), and then stop and close the engine.
I write the two input buffers as a WAV file so that I can import this into an an existing DAW. to visually examine the signals. I can see that each time I take a measurement, the time difference between the two signals is different (despite the fact the microphone is not moved and the hardware has stayed the same). I am assuming this is to do with the latency of the hardware, the time taken to initialise the engine and the way the divice distributes tasks.
I have tried to capture the absolute time using mach_absolute_time of the first buffer callback on each installTap function and subtracting the two, and I can see that this does vary quite a lot with each call:
class newAVAudioEngine{
var engine = AVAudioEngine()
var audioBuffer = AVAudioPCMBuffer()
var running = true
var in1Buf:[Float]=Array(repeating:0, count:totalRecordSize)
var in2Buf:[Float]=Array(repeating:0, count:totalRecordSize)
var buf1current:Int = 0
var buf2current:Int = 0
var in1firstRun:Bool = false
var in2firstRun:Bool = false
var in1StartTime = 0
var in2startTime = 0
func measure(inputSweep:SweepFilter) -> measurement {
initializeEngine(inputSweep: inputSweep)
while running == true {
}
let measureResult = measurement.init(meas: meas,ref: ref)
return measureResult
}
func initializeEngine(inputSweep:SweepFilter) {
buf1current = 0
buf2current = 0
in1StartTime = 0
in2startTime = 0
in1firstRun = true
in2firstRun = true
in1Buf = Array(repeating:0, count:totalRecordSize)
in2Buf = Array(repeating:0, count:totalRecordSize)
engine.stop()
engine.reset()
engine = AVAudioEngine()
let srcNode = AVAudioSourceNode { _, _, frameCount, AudioBufferList -> OSStatus in
let ablPointer = UnsafeMutableAudioBufferListPointer(AudioBufferList)
if (Int(frameCount) + time) <= inputSweep.stimulus.count {
for frame in 0..<Int(frameCount) {
let value = inputSweep.stimulus[frame + time]
for buffer in ablPointer {
let buf: UnsafeMutableBufferPointer<Float> = UnsafeMutableBufferPointer(buffer)
buf[frame] = value
}
}
time += Int(frameCount)
return noErr
} else {
for frame in 0..<Int(frameCount) {
let value = 0
for buffer in ablPointer {
let buf: UnsafeMutableBufferPointer<Float> = UnsafeMutableBufferPointer(buffer)
buf[frame] = Float(value)
}
}
}
return noErr
}
let format = engine.outputNode.inputFormat(forBus: 0)
let stimulusFormat = AVAudioFormat(commonFormat: format.commonFormat,
sampleRate: Double(sampleRate),
channels: 1,
interleaved: format.isInterleaved)
do {
try AVAudioSession.sharedInstance().setCategory(.playAndRecord)
let ioBufferDuration = 128.0 / 44100.0
try AVAudioSession.sharedInstance().setPreferredIOBufferDuration(ioBufferDuration)
} catch {
assertionFailure("AVAudioSession setup failed")
}
let input = engine.inputNode
let inputFormat = input.inputFormat(forBus: 0)
print("InputNode Format is \(inputFormat)")
engine.attach(srcNode)
engine.connect(srcNode, to: engine.mainMixerNode, format: stimulusFormat)
if internalRefLoop == true {
srcNode.installTap(onBus: 0, bufferSize: 1024, format: stimulusFormat, block: {(buffer: AVAudioPCMBuffer!, time: AVAudioTime!) -> Void in
if self.in2firstRun == true {
var info = mach_timebase_info()
mach_timebase_info(&info)
let currentTime = mach_absolute_time()
let nanos = currentTime * UInt64(info.numer) / UInt64(info.denom)
self.in2startTime = Int(nanos)
self.in2firstRun = false
}
do {
let floatData = buffer.floatChannelData?.pointee
for frame in 0..<buffer.frameLength{
if (self.buf2current + Int(frame)) < totalRecordSize{
self.in2Buf[self.buf2current + Int(frame)] = floatData![Int(frame)]
}
}
self.buf2current += Int(buffer.frameLength)
if (self.numberOfSamples + Int(buffer.frameLength)) <= totalRecordSize{
try self.stimulusFile.write(from: buffer)
self.numberOfSamples += Int(buffer.frameLength) } else {
self.engine.stop()
self.running = false
}
} catch {
print(NSString(string: "write failed"))
}
})
}
let micAudioConverter = AVAudioConverter(from: inputFormat, to: stimulusFormat!)
var micChannelMap:[NSNumber] = [0,-1]
micAudioConverter?.channelMap = micChannelMap
let refAudioConverter = AVAudioConverter(from: inputFormat, to: stimulusFormat!)
var refChannelMap:[NSNumber] = [1,-1]
refAudioConverter?.channelMap = refChannelMap
//Measurement Tap
engine.inputNode.installTap(onBus: 0, bufferSize: 1024, format: inputFormat, block: {(buffer2: AVAudioPCMBuffer!, time: AVAudioTime!) -> Void in
//print(NSString(string:"writing"))
if self.in1firstRun == true {
var info = mach_timebase_info()
mach_timebase_info(&info)
let currentTime = mach_absolute_time()
let nanos = currentTime * UInt64(info.numer) / UInt64(info.denom)
self.in1StartTime = Int(nanos)
self.in1firstRun = false
}
do {
let micConvertedBuffer = AVAudioPCMBuffer(pcmFormat: stimulusFormat!, frameCapacity: buffer2.frameCapacity)
let micInputBlock: AVAudioConverterInputBlock = { inNumPackets, outStatus in
outStatus.pointee = AVAudioConverterInputStatus.haveData
return buffer2
}
var error: NSError? = nil
//let status = audioConverter.convert(to: convertedBuffer!, error: &error, withInputFrom: inputBlock)
let status = micAudioConverter?.convert(to: micConvertedBuffer!, error: &error, withInputFrom: micInputBlock)
//print(status)
let floatData = micConvertedBuffer?.floatChannelData?.pointee
for frame in 0..<micConvertedBuffer!.frameLength{
if (self.buf1current + Int(frame)) < totalRecordSize{
self.in1Buf[self.buf1current + Int(frame)] = floatData![Int(frame)]
}
if (self.buf1current + Int(frame)) >= totalRecordSize {
self.engine.stop()
self.running = false
}
}
self.buf1current += Int(micConvertedBuffer!.frameLength)
try self.measurementFile.write(from: micConvertedBuffer!)
} catch {
print(NSString(string: "write failed"))
}
if internalRefLoop == false {
if self.in2firstRun == true{
var info = mach_timebase_info()
mach_timebase_info(&info)
let currentTime = mach_absolute_time()
let nanos = currentTime * UInt64(info.numer) / UInt64(info.denom)
self.in2startTime = Int(nanos)
self.in2firstRun = false
}
do {
let refConvertedBuffer = AVAudioPCMBuffer(pcmFormat: stimulusFormat!, frameCapacity: buffer2.frameCapacity)
let refInputBlock: AVAudioConverterInputBlock = { inNumPackets, outStatus in
outStatus.pointee = AVAudioConverterInputStatus.haveData
return buffer2
}
var error: NSError? = nil
let status = refAudioConverter?.convert(to: refConvertedBuffer!, error: &error, withInputFrom: refInputBlock)
//print(status)
let floatData = refConvertedBuffer?.floatChannelData?.pointee
for frame in 0..<refConvertedBuffer!.frameLength{
if (self.buf2current + Int(frame)) < totalRecordSize{
self.in2Buf[self.buf2current + Int(frame)] = floatData![Int(frame)]
}
}
if (self.numberOfSamples + Int(buffer2.frameLength)) <= totalRecordSize{
self.buf2current += Int(refConvertedBuffer!.frameLength)
try self.stimulusFile.write(from: refConvertedBuffer!) } else {
self.engine.stop()
self.running = false
}
} catch {
print(NSString(string: "write failed"))
}
}
}
)
assert(engine.inputNode != nil)
running = true
try! engine.start()
So The above method is my entire class. Currently each buffer call on installTap writes the input directly to a WAV file. This is where I can see the two end results differing each time. I have tried adding the startTime variable and subtracting the two, but the results still vary.
Do I need to take into account my output will have latency too that may vary with each call? If so, how do I add this time into the equation? What I am looking for is for the two inputs and outputs to all have relative time, so that I can compare them. The different hardware latency will not matter too much, as long as I can identify the end call times.

Answer 1

If you are doing real-time measurements, you might want to use AVAudioSinkNode instead of a Tap. The Sink Node is new and introduced along with AVAudioSourceNode you are using. With installing a Tap you won't be able to get precise timing.