paint-brush
Guitar Tuner: Pitch Detection for Dummiesby@jeremygustine
2,180 reads
2,180 reads

Guitar Tuner: Pitch Detection for Dummies

by Jeremy GustineJuly 6th, 2021
Read on Terminal Reader
Read this story w/o Javascript
tldt arrow

Too Long; Didn't Read

The ultimate goal is to build a real-time guitar tuner that operates on sound waves picked up through microphones. At the heart of most guitar tuners is some sort of pitch detection algorithm. This post will focus on exploring 3 of those algorithms - zero-crossing, fast fourier transform, and autocorrelation. We will only be concerned with a six-string, standard-tuned guitar. The western musical system has 12 notes, which are along the top row. The frequency of each octave is double the frequency of the previous octave.

People Mentioned

Mention Thumbnail

Company Mentioned

Mention Thumbnail

Coin Mentioned

Mention Thumbnail
featured image - Guitar Tuner: Pitch Detection for Dummies
Jeremy Gustine HackerNoon profile picture

I recently spent some time working on a passion project of mine. As a guitar player (that is an incredibly generous description) and a software developer, I have always been interested in attempting to develop a guitar tuner myself.


I have a goal to create a simple guitar tuner that will work by analyzing a live audio signal via microphone - perfect for a laptop/desktop computer or mobile phone. At the heart of most guitar tuners is some sort of pitch detection algorithm. This post will focus on exploring 3 of those algorithms - zero-crossing, fast fourier transform, and autocorrelation.

Guitar Basics

Before diving into pitch detection, we need to understand some basics about the guitar. We will only be concerned with a six-string, standard-tuned guitar. The notes of the six strings, from low pitch to high, are E, A, D, G, B, and E. Below is a chart with the frequencies of every note on a standard-turned guitar (assuming it has 22 frets):



C

C#

D

Eb

E

F

F#

G

G#

A

Bb

B

2





82.41

87.31

92.50

98.00

103.8

110.0

116.5

123.5

3

130.8

138.6

146.8

155.6

164.8

174.6

185.0

196.0

207.7

220.0

233.1

246.9

4

261.6

277.2

293.7

311.1

329.6

349.2

370.0

392.0

415.3

440.0

466.2

493.9

5

523.3

554.4

587.3

622.3

659.3

698.5

740.0

784.0

830.6

880.0

932.3

987.8

6

1047

1109

1175











The numbers 2-6 in the left-most column are octaves. The western musical system has 12 notes, which are along the top row. The 12 notes continuously repeat across octaves. The first note on the first string of the standard-tuned guitar is E2, while the first note on the sixth string is E4.


There are a few interesting things to observe in the chart above. The first is that the difference in frequency between notes gradually increases as the pitch of notes gets higher. The second is that the frequency of each octave is double the frequency of the previous octave. For example, E3 has a frequency of 164.8, which is double E2's frequency of 82.41.


There are a couple of other concepts we should define that will help us as we look at these algorithms:

  • Fundamental frequency: "The lowest frequency of any vibrating object is called the fundamental frequency. The fundamental frequency provides the sound with its strongest audible pitch reference - it is the predominant frequency in any complex waveform" (teachmeaudio.com).
  • Harmonic: A harmonic is an integer multiple of the fundamental frequency. A vibrating guitar string will produce the fundamental frequencies as well as numerous harmonics (double the fundamental frequency, triple the fundamental frequency, etc.).

Goals and Constraints

The ultimate goal is to build a real-time guitar tuner that operates on sound waves picked up through microphones (as opposed to analyzing an audio clip, for example). To make our task simpler, we are not interested in distinguishing between octaves. It will be sufficient for us to identify an E note, for example, rather than knowing if a note is E2 or E3.

Zero Crossing

Perhaps the most basic algorithm that can be considered for pitch detection is the zero crossing method. As the name implies, this technique works by analyzing an audio signal in the time domain and counting the number of times the amplitude of the wave changes from a positive to a negative value.


Plot of a 440 Hz clean sine wave

Above is a picture of a nice, clean sine wave. With a clean signal, we can easily calculate the frequency of the signal. "The frequency of a sine wave is the number of complete cycles that happen every second" (mathopenref.com).


“Frequency is equal to the number of cycles divided by the time“ (softschools.com). With this simple formula, we can create an implementation of a zero-crossing algorithm in Javascript:


function getNumZeroCrossings (samples) {
  var numZeroCrossings = 0
  for (var i = 1; i < samples.length; i++) {
    if (samples[i] * samples[i - 1] < 0) {
      numZeroCrossings++
    }
  }
  return numZeroCrossings
}

function getNumCycles (numZeroCrossings) {
  return Math.floor((numZeroCrossings - 1) / 2)
}

function calculateFrequency (signalDurationSeconds, numCycles) {
  return numCycles / signalDurationSeconds
}

function start () {
  var signalDurationSeconds = 1
  var arr = [6, 2, -2, -6, -2, 2, 6, 2, -2, -6, -2, 2, 6, 2, -2]
  var numZeroCrossings = getNumZeroCrossings(arr)
  var numCycles = getNumCycles(numZeroCrossings)
  var freq = calculateFrequency(1, numCycles)
  console.log(freq + " Hz") // outputs "2 Hz"
}


The zero crossing method of pitch detection is computationally inexpensive and easy to understand. It works very well for clean audio signals. Unfortunately, clean sine waves, as pictured above, are hard to come by when building a guitar tuner, especially when receiving input from a microphone.


A time-Domain wave of the low E guitar string being plucked

The image pictured above is a more realistic example of the kinds of waveforms that we will be dealing with. The wave is much more complex than a simple sine wave due to a variety of factors such as harmonics, and more importantly, noise. As you can see, the zero crossings in this wave are much more unpredictable. A naive zero-crossing algorithm will not suffice for such complex audio signals.


Fast Fourier Transform

What is a fast Fourier transform? According to Wikipedia:

A fast Fourier transform (FFT) is an algorithm that computes the discrete Fourier transform (DFT) of a sequence, or its inverse (IDFT). Fourier analysis converts > a signal from its original domain (often time or space) to a representation in the frequency domain and vice versa. The DFT is obtained by decomposing a sequence > of values into components of different frequencies.


Ok...so what does that mean? Real-world audio signals are complex and contain a variety of frequency information. For our purposes, the FFT will convert the signal into a set of numbers that we can use to figure out which frequencies are the most prominent in the signal. Let's take a look at a few pictures to make this more concrete.


A time-domain plot of a 440 Hz clean sine wave

FFT plot of a 440 Hz clean sine wave


The two images above show a time-domain graph of a clean 440 Hz sine wave and a frequency-domain graph of the Fast Fourier Transform of the same wave. Since the signal is a constant 440 Hz, the FFT diagram shows a single spike at 440 Hz. This data is very easy to interpret. Let's take a look at a slightly more complex example.


A slightly less clean time-domain signal the FFT counterpart

Let’s say that the image above (from Tony DiCola on learn.adafruit.com) shows the time-domain sound wave of a whistle recorded through a microphone, as well as the fast Fourier transform of that wave. The time-domain wave is not nearly as clean as the previous example. How can we detect the pitch of the signal? There is not a single frequency present in the wave. By passing this signal through an FFT, however, we can see that there is a particular frequency that is significantly more prominent than the rest. The peak in the FFT diagram shows us what the pitch of the whistle is. The FFT's ability to divide an audio signal into frequency components makes it a powerful technique that is frequently used in audio analysis.


Interpreting FFT output data

After passing a time-domain signal through an FFT, you will receive back an array of numbers that represents the frequency domain.


fft_output = [23, 43, 65, 443, 321, 54, 56 ... ]


Each index of the array is called a bin. You can think of each bin as a bucket for a range of frequencies. The number in each bin is the amplitude, or "strength" of the frequencies in that bin in the signal. The higher the amplitude, then the more those frequencies are represented in the signal. The size of the range of frequencies that each bin represents depends on the resolution.


resolution = sample_rate / fft_size


We can see by the formula that we can increase our resolution by either reducing the sample_rate or increasing fft_size. The fft_size can be thought of as a buffer that needs to be filled with sound samples. If we desire increased resolution, we can reduce the sample_rate, but it will take longer to fill the buffer.


To determine which frequencies are in which bin, we can use the following formula:


starting_frequency = bin_number * resolution


The range of frequencies for a given bin_number would then be from starting_frequency to starting_frequency + resolution.


For example, if we have a sample rate of 48000 and a buffer (FFT) size of 1024, then we will have a resolution of 46.875. This means that each bin represents a range of frequencies 46.875 Hz wide. If we would like to determine the frequencies represented in bin number 12, we can see that the starting_frequency will be 562.5. This means that bin 12 represents the frequencies from 562.5 to 609.375.


The importance of resolution

In the above example, we have a resolution of 46.875. As is shown in the frequency chart in the Guitar Basics section, the difference in frequencies between the first two notes on a guitar (where the difference between frequencies is the lowest) are (F - E), or 87.31 Hz - 82.41 Hz = 4.9 Hz. The resolution of 46.875 that we achieved with a sample rate of 48000 and an FFT size of 1024 is not even close to giving us enough precision to distinguish between those two notes. In addition, according to the paper A Digital Guitar Tuner, a trained musician can distinguish differences as small as 0.5 Hz. In order to decrease our resolution, we need to reduce our sampling rate and/or increase the FFT size. This comes at the cost of additional computational complexity and more time required to fill the sample buffer, which reduces how real-time the results are. Achieving a resolution of 0.5 Hz is incredibly non-performant, but we can get passable results with a sampling rate of 48000 and an FFT size of 16384, which gives us a resolution of 2.92.

Applying the FFT to a guitar signal

Given the above information, we should be able to pluck a guitar string, then pass the audio signal through an FFT, and then look for a spike in the data to figure out the frequency, right? Let's take a look at an FFT plot of plucking the low E string.


Time-Domain wave of the low E guitar string being plucked

Frequency-domain plot of the low E guitar string being plucked

The data in these images are more interesting and complex than in the previous images shown. When plucking a guitar's low E string, we get numerous spikes in our FFT plot. These spikes are called harmonics (or overtones). Harmonics are spikes in frequency at multiples of the fundamental frequency. Can we simply pick the largest spike in the plot and consider that the frequency?


For a low E string, we are expecting to see a spike at the fundamental frequency of 82.41 Hz - and we do. The problem is that, while this is a significant spike in the data, it is not the largest one. This is a common problem when using pitch detection algorithms on audio signals from musical instruments. Depending on the data, the fundamental frequency may be the first spike, or the second, or the third, and so on. It may be the largest spike, and it may not. Given these problems, how can we reliably detect the pitch?


The best answer is to use a more robust pitch detection algorithm! The FFT, by itself, is simply not the best tool for the job when trying to build a pitch-detecting guitar tuner. There are, however, some additional techniques that can be used to extract the pitch from FFT output.


Making lemonade out of lemons

There are various algorithms and heuristics that can be used on the imperfect data to attempt to get decent results in this scenario. One of the most well-known methods is called the Harmonic Product Spectrum, which will not be covered in this article, but is definitely worth investigating.


For my demo, I leveraged the fact that harmonics are multiples are the fundamental frequency and that we do not need to distinguish between octaves. The simple heuristic finds the frequency of the largest spike in the data set and then divides the frequency by increasingly large integers from 1..n until the number is within the range of notes on a typical guitar fretboard. This method works since we are not interested in distinguishing between octaves.


In practice, it works okay on the higher strings of the guitar, where the difference in frequency between notes is large. The problems with the FFT's resolution mean that a more robust algorithm is needed to produce a reliable guitar tuner.


Check out a demo of FFT-based pitch detection here. The visualizations are based on some awesome work by George Gally.

Autocorrelation

What is autocorrelation? According to Wikipedia:

Autocorrelation, also known as serial correlation, is the correlation of a signal with a delayed copy of itself as a function of delay. Informally, it is the similarity between observations as a function of the time lag between them. The analysis of autocorrelation is a mathematical tool for finding repeating patterns, such as the presence of a periodic signal obscured by noise, or identifying the missing fundamental frequency in a signal implied by its harmonic frequencies. It is often used in signal processing for analyzing functions or series of values, such as time domain signals.


In a nutshell, autocorrelation can be used to extract the otherwise difficult-to-recognize repeating patterns within a noisy signal. We can then determine the frequency by counting the repeating patterns.


Autocorrelation is a special case of cross-correlation. "Cross-correlation is a measurement that tracks the movements of two or more sets of time series data relative to one another” (Investopedia). In digital signal processing, cross-correlation can be used to compare two signals to each other to determine how similar - or correlated - they are. Autocorrelation is similar, except that we are going to compare a signal to a time-shifted version of the same signal.


The equation

When researching autocorrelation methods, the first thing you will discover is that there is not a single definition. Different disciplines (statistics, finance, digital signal processing, etc.) define the function slightly differently. The autocorrelation can be calculated on continuous and discrete signals. We will be using a typical DSP definition for a discrete-time signal.


The autocorrelation function

In this equation, l represents the lag, x represents the sequence of data that we will be autocorrelating, and N is the number of data points in the sequence. The idea here is that we will call this function repeatedly for an increasing value of l, where the value will increase from 0 to N.


Most definitions of autocorrelation will include normalization of the data so that the maximum value of the output is scaled to 1. This is not strictly necessary to design a guitar tuner, and in fact, many definitions of autocorrelation within the DSP discipline will omit this step. We will normalize the data, however, since this is included in many autocorrelation definitions and because it makes visualizing the results easier to reason about. Many definitions of autocorrelation will use the standardization or min-max scaling methods of normalization, but we will use maximum absolute scaling to normalize our data so that it is all scaled between -1 and 1.

Maximum absolute scaling

The code

function rxx(l, N, x) {
  var sum = 0;
  for (var n = 0; n <= N - l - 1; n++) {
    sum += (x[n] * x [n + l])
  }
  return sum;
}

function autocorrelationWithShiftingLag(samples) {
  var autocorrelation = []
  for (var lag = 0; lag < samples.length; lag++) {
    autocorrelation[lag] = rxx(lag, samples.length, samples)
  }
  return autocorrelation
}

function maxAbsoluteScaling(data) {
  var xMax = Math.abs(Math.max(...data))
  return data.map(x => x / xMax)
}

var audioData = [...] // time domain data collected from microphone
var result = maxAbsoluteScaling(autocorrelationWithShiftingLag(audioData))


Building intuition

Autocorrelation: comparing a signal to the time-shifted version of itself


The fantastic animation above (by Qingkai Kong) shows how the autocorrelation method works. The blue signal is the original signal. The red one is the time-shifted version, where we are continuously increasing the lag for each invocation of the autocorrelation function. The image on the bottom shows the result of the autocorrelation. Notice that at a time lag of 0, the result of the autocorrelation is 1. This is because the signal compared with itself is identical - it is perfectly correlated. As we continue to time-shift the red signal, the correlation function will output larger numbers when the signals are very similar and smaller numbers when they are not.


The peaks of the autocorrelation output represent the periods of the signal. By counting the time between each peak, we can determine the frequency of the signal. Qingkai Kong’s example above shows the comparison of a clean sine wave with itself. This method is especially useful for noisier signals with harder to recognize periodicity patterns. This fantastic video by David Dorran does a great job of visualizing how this method works with more complex signals.


After building the autocorrelation result, a simple peak detection algorithm (very similar to the zero-crossing method described above) can be used to extract the frequency from the data.


Check out a demo of autocorrelation-based pitch detection here. The visualizations are built with p5.js, based on work by Jason Sigal and Golan Levin.


Summary

The three-pitch detection methods described in this article each have their strengths and weaknesses and are best used for particular use-cases. The zero-crossing method is computationally inexpensive and easy to understand, but it does not work well with noisy signals. The fast fourier transform has the advantage of being a well-known method of analyzing signals in the frequency domain. It can work well for distinguishing between musical notes at higher frequencies. However, it is difficult to achieve the resolution required to detect the small differences in frequency required for a reliable guitar tuner. The autocorrelation algorithm is the most reliable of the three for a guitar tuner, but, like the FFT, it is vulnerable to harmonic interference.


While each of these methods has enough drawbacks to make it difficult to design a robust guitar tuner by themselves, they form the basis for many more advanced algorithms. For example, an FFT combined with a Harmonic Product Spectrum can help minimize overtone problems. With autocorrelation, peak detection can be improved with parabolic interpolation. The YIN algorithm is a popular pitch detection method for which the autocorrelation function is a step in the process. This page describes several more pitch estimation methods that were not discussed in this post.