How to find the highest peaks in an audio signal?

Hi,

I have a set of audio files(.wav). Each audio file is the recording of a user typing a 8 letter word. Each letter when typed produces two sounds 1. when key is pressed, 2. when key is released. Total 16 sounds are produced for one complete typing. My question is, How can I find the highest peak of each press and release and how can I extract the timing information when a peak is occurred?

To simply my question, How to extract the timing information when a sound is produced?

How can i solve this problem. Please anyone guide me on this.

Thanks in advance.

[Star Strider]

It might help if you posted a representative sample of your .wav file. There could be spectral characteristics that would help identify them as well. (That could be typewriter-dependent, so it might not generalise across typewriters.)

Reference (for those born after 1970): Typewriter.

[RamSum]

Please find the image of audio

[Star Strider]

I was hoping for the .wav file, not the image. I would have done a spectrogram of it to see if the key-press and key-release had different (and ideally repeatable) spectra, then if they did, design appropriate digital filters to detect them.

Oh, well...

[RamSum]

Sorry I misunderstood. Here is the audio file of the typing.

Answer 1

[Greg Dionne]

Try something like this to get you started:

y = audioread('U1_S1_T1.wav')
yLeft = y(:,1);
yrmsenvLeft = sqrt(conv(yLeft.^2,ones(200,1)/200,'same'))
plot([yLeft yrmsenvLeft])

Answer 2

[Image Analyst]

I know it seems obvious, but how about just thresholding??

loudElements = signal > someValue;  % Logical vector.
loudIndexes = find(loudElements);  % Linear indexes - actual index numbers.

Anyway, it's a start. If you need something more robust and sophisticated, you'll have to tweak it and develop something better.

[Image Analyst]

After seeing your signal, I'd do this:

loudElements = abs(signal) > 0.1;  % Logical vector.
loudIndexes = find(loudElements);  % Linear indexes - actual index numbers.

[RamSum]

I have tried it, but the problem is there are still many peaks even if i gave threshold. Thanks for the advice.

[Image Analyst]

Yes, there are . What were you expecting? Just one? How many keys were hit? Looking at the plot, exactly what kind of data would you want to be returned to you?

[RamSum]

Actually I was expecting 16 peaks for each key press and release (the highest peaks). Total 8 keys were hit (16 if you consider key press and release). The main aim is to extract the timing information when the sound is produced, so that I can extract features(Duration and latency values). I choose to find the highest peaks because we can identify by seeing the plot that the sound will be at highest peaks

Or is there any other approach i can solve this task? Please help

[Image Analyst]

So for each key press and release there are 16 peaks? And this happened 8 times for the 8 keys typed in your audio recording? So there should be 128 peaks = 16 * 8?

Do you have the Image Processing Toolbox? You might be able to use bwlabel on the thresholded signal to count the separate "runs" above the threshold signal value.

[RamSum]

No No, you understood it wrong. For example, if a user is pressing letter "A" first he will press the key and release it. when the user touches the key (key press) a sound is produced. In the same way during the release (Key release) another sound is produced. So, for one letter (key typing) two sounds are produced. Similarly, there is a 8 letter word was typed during my data collection. Total 16 sounds for whole word.

[Image Analyst]

That's what I first thought, until you said, quote, "I was expecting 16 peaks for each key press and release" - they key word being "each".

[RamSum]

I don't have an image processing tool box. How can i get it? Does it cost, if does how much?

[Image Analyst]

Type "ver" to see if you have it. If you want it, you'll have to contact the Mathworks.