hfink / matchbox Goto Github PK

Hi, thanks for sharing your code. I took a dive into the code to see if the project could be configured to work as a simple humming recognition library (which I intend to share to the community also).

Just wondering if I could ask from the expert whether your code could easily be converted? I'm imagining this scenario:

1. You have pre-recorded a 4 second tune that someone has hummed. So you have that stored locally on the device.
2. You use the record feature on the app to record a user's attempt to mimic the hum.
3. You score the user's closeness to the original.

Based on your experience and approach, is this possible? If the original hum was from a man, would a women's hum be markedly different in terms of the min distance score?

I'm using your MFCC processor in a custom project and some songs return inf and NaN float values for some frames.

More information:
I am using a custom audio file loader and run MFCC processor in a single run - all audio samples are processed in a single buffer, not buffer by buffer when the samples are read. Only a single channel float format is sent to MFCC processor, values between -1.0 and 1.0 of course.

In some cases the songs I use to test first few frames return NaN (the rest of them are fine apparently). That causes the mean operation to place NaN values across whole vector.

Also the AVAsset reader apparently does some preprocessing itself (mixing, normalizing or something), do you have any information regarding that? Because audio loading system I use is built on Extended Audio Services and sample values are different than sample values obtained from AVAssetReader. This leads to MFCC be quite different if they are calculated using AVAsset reader source or Extended Audio Services (I am actually using TheAmazingAudioEngine audio file loading operation).