Recognizing Kiwi Bird Calls in Audio Recordings

My project for WSC 2017 was identifying kiwi calls in audio recordings. The project can be broken down into 2 main steps:

1. Finding clips that contain noise that could be a kiwi
2. Identifying the clips that actually contain a kiwi call

Finding Clips

Data

The data is an audio recording taken overnight in Northland, New Zealand

Filtering

I started off by filtering the audio to be between 1200Hz and 3600Hz in order to remove the majority of the noise. I then normalized the audio to make the volume consistent.

AudioNormalize[HighpassFilter[LowpassFilter[audio, Quantity[1200, "Hertz"]],  Quantity[3600, "Hertz"]]]

At first I just took any intervals of audio that were above a certain threshold.

AudioIntervals[audioProcessed, #RMSAmplitude > 0.02 &]

This yielded many very short clips so I then extended each clip 1 second in each direction and merged clips that were within 2 seconds of each other.

({#[[1]] - 1, #[[2]] + 1} & /@ AudioIntervals[audioProcessed, #RMSAmplitude > 0.02 &]) 
//. {pre___, {c_, d_}, {x_, y_}, post___} /; d > x - 2 -> {pre, {c, y}, post}

This still had the problem of containing many clips just over 2 seconds long so I removed any clips less than 10 seconds asas kiwi calls are longer than ten seconds.

Cases[int, {x_, y_} /; y - x > 10]

I then limited the length of each of the clips to exactly 10 seconds to make it easier for the machine learning.

AudioTrim[#,10]&/@clips

Identifying calls

Unsupervised Learning

Since my data was unclassified I initially tried to use unsupervised learning for clustering the audio clips, although this didn't yield any particularly meaningful results.

FeatureSpacePlot[clips, LabelingFunction -> (#2[[2]] &), PerformanceGoal -> "Quality"]

The middle section is about 50% kiwi calls and 50% not kiwi calls, as is the circle around the outside so I have no idea what the feature extractor is looking at.

Classifying Data

Since unsupervised learning didn't work particularly well I decided to manually classify every single clip.

Doesn't that sound fun...

clipId = 1;
clipClasses = Range@Length[clips2];
Dynamic@Row[{clipId, "/", Length[clips2], 
   If[clipId <= Length[clips2], clips2[[clipId]], "DONE!"] k, 
   Button["Yes", clipClasses[[clipId]] = "Kiwi"; clipId = clipId + 1],
    Button["No", clipClasses[[clipId]] = "NotKiwi"; 
    clipId = clipId + 1]}]
Dynamic[Row[{clipId - 1, clipClasses[[clipId - 1]]}]]

Neural Network - Take One

I used 200 of the clips as training data for the neural network.

data=Thread[clips,clipClasses];
training=RandomSample[data,200];
Counts[training]
test=Complement[data,training];
Counts[test]

I used a neural net to classify the audio clips as it was the best at classifying them.

cf=Classify[training,Method->"NeuralNetwork",PerformanceGoal->"Quality"]
cm=ClassifierMeasurements[cf,test]

The accuracy of the neural network was extremely poor though

cm["Accuracy"]
0.559565

Neural Network - Take Two

I then tried downsampling the audio from 44.1kHz to 10kHz to reduce the amount of extraneous data the neural network has to work with.

clipsSmall = AudioResample[#, Quantity[10, "Kilohertz"]] & /@ clips;

This reduced the amount of data without significantly changing the audio

cm["Accuracy"]
0.453762

Well back to the drawing board I guess

Neural Network - Take Three

This time I tried a different approach, since the neural network seemed to handle audio extremely poorly I instead input the spectrogram of the audio into the neural net.

data= Thread[Image[Abs[SpectrogramArray[#]]] & /@ clipsSmall -> clipClasses[[All, 1]]]

The image processing side has had a lot more work done so this should work much better.

cm["Accuracy"]
0.820756

At this point I ran out of ways to improve the score and ran up against the limit of the accuracy that I was able to classify the clips so I'm going to call that a success.

Finding calls

Finally finding the calls, which is the simple now that we can find and classify potential calls

FindCalls[clip_] := (Module[{int = {}, audioProcessed, clips, classes},
   audioProcessed = ProcessAudio[clip];
   int = ProcessAudioIntervals[audioProcessed];
   int = Cases[int, {x_?NumberQ, y_?NumberQ} /; y - x >= 10];
   clips = Which[Length[int] == 0, {},
     Length[int] == 1, 
     AudioTrim[#, 10] & /@ {AudioTrim[audioProcessed, int]},
     Length[int] > 1, 
     AudioTrim[#, 10] & /@ AudioTrim[audioProcessed, int]];
   classes = 
    KiwiCallClassifier[
     Image[Abs[
         SpectrogramArray[
          AudioResample[#, Quantity[10, "Kilohertz"]]]]] & /@ clips];
   Thread[{Extract[clips, Position[classes, "Kiwi"]], 
     Quantity[#, "Seconds"] & /@ 
      Extract[int, Position[classes, "Kiwi"]][[All, 1]]}]])

FindCalls[Import@"C:\\Users\\Isaac\\Desktop\\Programming\\data\\Kiwi Audio\\Processing\\20170604k-53.mp3"] // Grid

Reflections

Good

Actually finding clips with noise was quite a simple and easy task, I just messed around with the frequencies on the filters and the threshold for a while.

Bad

In hindsight I realize that this wasn't a problem that was particularly suited to unsupervised learning, as there are other more general features than "kiwi" or "not kiwi" for a feature extractor to identify (although I still have no idea what it was doing).

Worst

Another thing that's important for undertaking a machine learning project like this is having a lot of data. If I do another project like this I'm definitely going to use more, already classified data instead of spending 2 hours listening to birds screaming and loud backgound noise.

This is the training data I used for the neural net https://drive.google.com/file/d/0B4VdlZ57AG6BcXBMMi1NQnQyRms/view?usp=sharing