Accurate timer using AudioUnit

Question

I'm trying to make an accurate timer to analyze an input. I'd like to be able to measure 1% deviation in signals of ~200ms. My understanding is that using an AudioUnit will be able to get <1ms. I tried implementing the code from Stefan Popp's example After updating a few things to get it to work on xcode 6.3, I have the example working, however:

1. While I do eventually want to capture audio, I thought there should be some way to get a notification, like NSTimer, so I tried an AudioUnitAddRenderNotify, but it does exactly what it says it should - i.e it's tied to the render, not just an arbitrary timer. Is there some way to get a callback triggered without having to record or play?

2. When I examine mSampleTime, I find that the interval between slices does match the inNumberFrames - 512 - which works out to 11.6ms. I see the same interval for both record and play. I need more resolution than that.

I tried playing with kAudioSessionProperty_PreferredHardwareIOBufferDuration but all the examples I could find use the deprecated AudioSessions, so I tried to convert to AudioUnits:

Float32 preferredBufferSize = .001; // in seconds
status = AudioUnitSetProperty(audioUnit, kAudioSessionProperty_PreferredHardwareIOBufferDuration, kAudioUnitScope_Output, kOutputBus, &preferredBufferSize, sizeof(preferredBufferSize));

But I get OSStatus -10879, kAudioUnitErr_InvalidProperty.

Then I tried kAudioUnitProperty_MaximumFramesPerSlice with values of 128 and 256, but inNumberFrames is always 512.

UInt32 maxFrames = 128;
status = AudioUnitSetProperty(audioUnit, kAudioUnitProperty_MaximumFramesPerSlice, kAudioUnitScope_Global, 0, &maxFrames, sizeof(maxFrames));

[EDIT] I am trying to compare the timing of an input (user's choice of MIDI or microphone) to when it should be. Specifically, is the instrument being played before or after the beat/metronome and by how much? This is for musicians, not a game, so precision is expected.

[EDIT] The answers seem re-active to events. i.e. They let me precisely see when something happened, however I don't see how I do something accurately. My fault for not being clear. My app needs to be the metronome as well - synchronize playing a click on the beat and flash a dot on the beat - then I can analyze the user's action to compare timing. But if I can't play the beat accurately, the rest falls apart. Maybe I'm supposed to record audio - even if I don't want it - just to get inTimeStamp from the callback?

[EDIT] Currently my metronome is:

- (void) setupAudio
{
    AVAudioPlayer *audioPlayer;
    NSString *path = [NSString stringWithFormat:@"%@/click.mp3", [[NSBundle mainBundle] resourcePath]];
    NSURL *soundUrl = [NSURL fileURLWithPath:path];
    audioPlayer = [[AVAudioPlayer alloc] initWithContentsOfURL:soundUrl error:nil];
    [audioPlayer prepareToPlay];

    CADisplayLink *syncTimer;
    syncTimer = [CADisplayLink displayLinkWithTarget:self selector:@selector(syncFired:)];
    syncTimer.frameInterval = 30;
    [syncTimer addToRunLoop:[NSRunLoop mainRunLoop] forMode:NSDefaultRunLoopMode];
}

-(void)syncFired:(CADisplayLink *)displayLink
{
    [audioPlayer play];
}   

Answer 1

You should be using a circular buffer, and performing your analysis on the signal in chunks that match your desired frame count on your own timer. To do this you set up a render callback, then feed your circular buffer the input audio in the callback. Then you set up your own timer which will pull from the tail of the buffer and do your analysis. This way you could be feeding the buffer 1024 frames every 0.23 seconds, and your analysis timer could fire maybe every 0.000725 seconds and analyze 32 samples. Here is a related question about circular buffers.

EDIT

To get precision timing using a ring buffer, you could also store the timestamp corresponding to the audio buffer. I use TPCircularBuffer for doing just that. TPCircularBufferPrepareEmptyAudioBufferList, TPCircularBufferProduceAudioBufferList, and TPCircularBufferNextBufferList will copy and retrieve the audio buffer and timestamp to and from a ring buffer. Then when you are doing your analysis, there will be a timestamp corresponding to each buffer, eliminating the need to do all of your work in the render thread, and allowing you to pick and choose your analysis window.

Answer 2

If you are using something like cross-correlation and/or a peak detector to find a matched sample vector within an audio sample buffer (or a ring buffer containing samples), then you should be able to count samples between sharp events to within one sample (1/44100 or 0.0226757 milliseconds at a 44.1k Hz sample rate), plus or minus some time estimation error. For events more than one Audio Unit buffer apart, you can sum and add the number of samples within the intervening buffers to get a more precise time interval than just using (much coarser) buffer timing.

However, note that there is a latency or delay between every sample buffer and speaker audio going out, as well as between microphone sound reception and buffer callbacks. That has to be measured, as in you can measure the round trip time between sending a sample buffer out, and when the input buffer autocorrelation estimation function gets it back. This is how long it takes the hardware to buffer, convert (analog to digital and vice versa) and pass the data. That latency might be around the area of 2 to 6 times 5.8 milliseconds, using appropriate Audio Session settings, but might be different for different iOS devices.

Yes, the most accurate way to measure audio is to capture the audio and look at the data in the actual sampled audio stream.