Pitch detection in Web Audio using autocorrelation
Switch branches/tags
Nothing to show
Clone or download
Pull request Compare This branch is even with markmarijnissen:master.
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
example
LICENSE.txt
README.md
package.json
pitchdetector.js
pitchdetectorcanvas.js

README.md

Pitch Detector

Based on Chris Wilson's work, an improved pitch detector. The Pitch Detector calculates auto-correlation score for a range of frequencies.

See demo at http://lab.madebymark.nl/pitch-detector/example/.

Examples

  • The Y-Axis is the auto-correlation score.
  • The X-Axis is the frequency range, from high (22 Khz) to low (83 Hz).
  • The Red bar is the signal strength (RMS). The little dark red line is the minimal RMS.

Detect best auto-correlation of all frequencies:

Auto-Correlation scores

Detect the first peak auto-correlation, which is the highest frequency. Auto-correlation often detects lower octaves (and harmonies), so we can just stop after the first peak.

Auto-Correlation scores, detect the first peak correlation

Detect a sudden increase in correlation:

Auto-Correlation scores, detect the first increase in correlation

Usage

  • pitchdetector.js contains the PitchDetector (logic only)
  • pitchdetectorcanvas.js allows you to visualize pitch detection on a canvas.
  • example/gui.js is a playground to test and tweak the pitch detector.

Drop pitchdetector.js in your page, or require the CommonJS module using Webpack or Browserify.

First, create a PitchDetector:

var detector = new PitchDetector({
	// Audio Context (Required)
	context: new AudioContext(),

	// Input AudioNode (Required)
	input: audioBufferNode, // default: Microphone input

	// Output AudioNode (Optional)
	output: AudioNode, // default: no output

	// interpolate frequency (Optional)
	//
	// Auto-correlation is calculated for different (discrete) signal periods
	// The true frequency is often in-beween two periods.
	//
	// We can interpolate (very hacky) by looking at neighbours of the best 
	// auto-correlation period and shifting the frequency a bit towards the
	// highest neighbour.
	interpolateFrequency: true // default: true

	// Callback on pitch detection (Optional)
	onDetect: function(stats, pitchDetector) { 
		stats.frequency // 440
		stats.detected // --> true
		stats.worst_correlation // 0.03 - local minimum, not global minimum!
		stats.best_correlation // 0.98
		stats.worst_period // 80
		stats.best_period // 100
		stats.time // 2.2332 - audioContext.currentTime
		stats.rms // 0.02 
	},

	// Debug Callback for visualisation (Optional)
	onDebug: function(stats, pitchDetector) { },

	// Minimal signal strength (RMS, Optional)
	minRms: 0.01,

	// Detect pitch only with minimal correlation of: (Optional)
	minCorrelation: 0.9,

	// Detect pitch only if correlation increases with at least: (Optional)
	minCorreationIncrease: 0.5,

	// Note: you cannot use minCorrelation and minCorreationIncrease
	// at the same time!
 
	// Signal Normalization (Optional)
	normalize: "rms" // or "peak". default: undefined

	// Only detect pitch once: (Optional)
	stopAfterDetection: false

	// Buffer length (Optional)
	length: 1024, // default 1024

	// Limit range (Optional):
	minNote: 69, // by MIDI note number
	maxNote: 80, 

	minFrequency: 440,    // by Frequency in Hz
	maxFrequency: 20000,

	minPeriod: 2,  // by period (i.e. actual distance of calculation in audio buffer)
	maxPeriod: 512 // --> convert to frequency: frequency = sampleRate / period

	// Start right away
	start: true // default: false
})

Then, start the pitch detection. It is tied to RequestAnimationFrame

detector.start()

If you're done, you can stop or destroy the detector:

detector.stop()
detector.destroy()

You can also query the latest detected pitch:

detector.getFrequency() // --> 440hz
detector.getNoteNumber() // --> 69
detector.getNoteString() // --> "A4"
detector.getPeriod() // --> 100
detector.getDetune() // --> 0
detector.getCorrelation() // --> 0.95
detector.getCorrelationIncrease() // --> 0.95

// or raw data
detector.stats = {
	stats.frequency
	stats.detected
	stats.worst_correlation
	stats.best_correlation 
	stats.worst_period
	stats.best_period
	stats.rms
}

Tips & Tricks

Always use an optimization

  • minCorrelation is the most reliable
  • minCorreationIncrease can sometimes give better results.

Use RMS or Peak normalization with minCorrelationIncrease

The increase in correlation strongly depends on signal volume. Therefore, normalizing using RMS or Peak can make minCorrelationIncrease work much better.

Set a frequency range

If you know what you're looking or, set a frequency range.

Warning: minCorrelationIncrease needs a large frequency range to detect a difference. The frequency range must be large enough to include both a low and high auto-correlation.

Ideas to improve algorithm:

  • Draw an "optimal" auto-correlation shape, and calculate mean squared error (MSE) from measured auto-correlation score. When MSE is low enough, a pitch is detected. (or a combination of pitches!)
  • Implement DTMF demodulation as example
  • Learn a shape by recording auto-correlation scores of the perfect example. The resulting shape is the average of all recordes samples. Calculate standard deviation to see if the signal can be detected reliably.

Changelog

0.2.0 (26/02/2015)

  • Used ScriptProcessingNode for faster analysis. Callbacks are still tied to the requestAnimationFrame.
  • Extracted the Canvas draw function into a seperate file.

0.1.0 (25/02/2015)

  • Extract core logic (pitchdetector.js) from the GUI code (example/gui.js)
  • Add a new heuristic: detect a sudden increase in auto-correlation (when approaching the target frequency).
  • Added signal normalization (peak or rms)
  • Updated canvas visualization to draw correlation scores for every frequency.

Contribute

I first want to check if the original author, Chris Wilson, is willing to pull my fork. So please check out the original version at https://github.com/cwilso/PitchDetect.

Credits

Original code from Chris Wilson, improvements (see changelog) by Mark Marijnissen

Contact

© 2015 - Mark Marijnissen & Chris Wilson