sevagh/pitch-detection

YIN vs McLeod pitch detection accuracy

apaatsio opened this issue · 7 comments

apaatsio commented

Are there benchmarks of the pitch detection accuracy of the YIN and McLeod algorithms? Is there any significant difference?

sevagh commented

It's a good question. You could probably modify https://raw.githubusercontent.com/sevagh/pitch-detection/master/test/test_instruments.cpp to compare YIN and MPM side-by-side.

Personally, since figuring out how to do YinFFT (and thus bringing the performance in line with MPM), I prefer YIN due to the low pitch cutoff for MPM:

Simply put MPM doesn't function well for low pitches under 80Hz e.g.: #50

I don't remember why I picked the value 80Hz. Maybe from real testing, or maybe using the guitar low E (~82.4Hz) as a threshold.

Here's some newer methods I've seen (maybe time for a pitch-detection-v2 project?):

https://code.soundsoftware.ac.uk/projects/pyin
https://github.com/marl/crepe

  • 👍1
sevagh commented

I'm considering using https://github.com/EliosMolina/audio_degrader to run some real/degradation tests against YIN and McLeod to make them prove their worth

sevagh commented

Here's a sneak preview of a test using my (hopefully soon-to-be-released) audio-degradation-toolbox.

First I got a file (Flute, Db7) from http://theremin.music.uiowa.edu/MIS.html

I applied some degradations:

sevagh:audio-degradation-toolbox $ cat degradations.json
[
  {
    "name": "noise",
    "color": "white",
    "snr": 6
  },
  {
    "name": "noise",
    "color": "violet",
    "snr": 13
  },
  {
    "name": "mp3",
    "bitrate": 32
  },
  {
    "name": "mp3",
    "bitrate": 32
  },
  {
    "name": "mp3",
    "bitrate": 32
  }
]

Running pitch detection against the pure clip (converted from aiff to wav):

sevagh:pitch-detection $ ./wav_analyzer/wav_analyzer ~/repos/audio-degradation-toolbox/Flute-Db7-pure.wav
sample rate: 44100
len samples: 133120
frame size: 2
seconds: 3.01859
channels: 1
mpm: 2267.48
yin: 2268.88
pyin: 12422.3, 0.305056
pyin: 2268.88, 0.682327
pyin: 1133.89, 0.012614
pmpm: 2267.48, 1

A Db7 should be 2217.46 but it could be out of tune I guess.

Finally, with the degradations applied:

sevagh:pitch-detection $ ./wav_analyzer/wav_analyzer ~/repos/audio-degradation-toolbox/Flute-Db7-degraded.wav
sample rate: 44100
len samples: 135983
frame size: 2
seconds: 3.08351
channels: 1
mpm: 2268.01
yin: 2273.61
pyin: 12790.1, 0.305056
pyin: 2273.61, 0.592254
pyin: 2.20599, 0.102687
pmpm: 2268.01, 1

MPM is closer to what it calculated with the clean clip. YIN is straying further.

I'll be formalizing all of this in a shiny new mega test suite, hopefully with some head-to-head tables to put in the README.

  • 👍2
j-funk commented

Thank you for doing all of this.

apaatsio commented

That's really cool @sevagh. Thank you for taking the time to do this!

adamski commented

Seems there is not much difference in accuracy. However, I am interested in the fact that YIN can detect lower frequencies more easily. I switched from YIN to MPM a long time ago, partly due to the lack of an FFT-based version at the time. Maybe time to revisit..