Signal Spectra

Periodic Signals

D.R Campbell

School of Computing

University of Paisley 1

Signal Spectra The magnitude spectrum of an audio signal describes the distribution of magnitudes with frequency i.e. what frequencies (of pure tones) are present and at what amplitudes. The phase spectrum can display in what way the phase relationship between two signals varies with frequency. Only the magnitude spectrum will be considered in the following notes.

Pure Tone (Sinusoid ) 1kHz 0dB

Pure tone amplitude 0dB

Spectral line Amplitude 0dB

1/Tp = 1kHz

Spectral line at 1000 Hz

Tp = 1 ms

Time history
D.R Campbell School of Computing

Amplitude Spectrum
University of Paisley 2

Spectrum Of Pure Tone Signal The magnitude spectrum of a pure tone (sinusoid) is a single spectral line at the frequency of the tone, and with an amplitude equal to that of the tone.

Triangular wave -3dB 160 Hz

Amplitude -3dB Tp = 6.25 ms

Fundamental 3rd Harmonic

D.R Campbell

1/Tp = 160Hz

School of Computing

University of Paisley 3

Spectrum Of a Complex Periodic Signal The magnitude spectrum of a complex periodic signal e.g. square wave, triangular wave, consists of a combination of pure tones based on a fundamental (which has a period equal to that of the complex periodic signal) plus higher harmonics with frequencies which are integer multiples of the fundamental frequency. The spectrum of a complex periodic signal thus consists of a spectral line at the fundamental frequency and a spectral line at each of the harmonic frequencies equally spaced on a linear frequency scale. Some periodic signals may be composed of a fundamental plus odd harmonics e.g. square wave, or a fundamental plus only even harmonics, or a fundamental plus an incomplete range of odd and even harmonics.

Periodic Pulse train -3dB 160 Hz

Fundamental

1/ = 1600Hz

2/ = 3200Hz

Sin(x)/x Envelope

1/Tp = 160Hz

D.R Campbell

School of Computing

University of Paisley 4

Spectrum Of a Periodic Pulse Signal The magnitude spectrum of a periodic pulse train also consists of a combination of pure tones based on a fundamental (which has a period equal to that of the complex periodic signal) plus higher harmonics with frequencies which are integer multiples of the fundamental frequency. The spectrum of a the pulse train signal consists of a spectral line at the fundamental frequency and a spectral line at each of the harmonic frequencies equally spaced by 1/T on a linear frequency scale. The envelope of the magnitude spectrum has a SINC function ( sin(x)/x ) shape which decreases to zero at intervals of 1/ , where is the pulse width. If the pulses are made narrower is smaller and the value 1/ becomes larger i.e. the zero locations of the SINC shaped envelope will move to higher frequencies. In the limit as becomes very small compared with T the spectrum will begin to look flat with frequency. This is an important property which is implicitly made use of in sampling signals to convert them to digital data, and explicitly made use of when using starting pistols to characterise acoustic spaces or when generating musical sounds from synthsisers.

Signal Spectra
Random Signals

D.R Campbell

School of Computing

University of Paisley 5

Spectra Of Random Signals Another magnitude spectrum which approaches a flat distribution with frequency is that of the random noise signal known as white noise. Since random signals are by their nature non-periodic their spectra will vary from moment to moment and a large number of samples or a long time period is necessary to characterise their average spectral properties. A white noise signal may be a random signal with a Gaussian amplitude distribution or it may have some other distribution. That is, the statistical properties do not determine the shape of the spectrum e.g. low-pass filtered Gaussian white noise signal will still have a Gaussian probability distribution of amplitudes. That said, most noise sources provided in synthesisers and audio signal generation software provide Gaussian amplitude distribution. Computer generated random signals are not truly random, they are termed pseudorandom. This means that they are in fact periodic signals, but with a very long period within which the amplitudes are changing effectively randomly for most practical purposes. Their periodic nature can be very useful when it is desired to repeat an experiment exactly e.g. the measurement of the acoustic properties of a room after changing some feature of the room. A truly random source would not allow this. Many modern computer based acoustics measuring systems use pseudo-random noise generators.

Typical White Noise Spectra

1024 data points

4096 data points

44100 data points

The spectra become flatter with frequency as larger numbers of samples are used to compute it i.e. for white noise, all frequencies are present at equal amplitude on a linear frequency scale.
D.R Campbell School of Computing University of Paisley 6

White Noise A random noise source is described as being white if its long term spectrum contains equal amplitudes of all frequencies. This means all frequencies between 0 Hz and half the sampling frequency when considering digital signal processing noise sources. The noise spectrum is termed white by analogy with white light i.e. light from a source which radiates all colours equally. The figure above shows the effect of including more and more samples in the calculation of the spectrum of a noise signal. The displayed spectrum initially looks as if there are greater amplitudes of some frequencies than others. This is shown not to be true in the long term as the number of samples used is increased. Another way of expressing this flatness with frequency property, is to say that white noise has equal power in equal bandwidths eg. the power in the 200 Hz bandwith between 100 Hz and 300 Hz equals that in the 200 Hz bandwidth between 1000 Hz and 1200 Hz.

Insufficient Averaging of Noise Spectra

White noise apparent Spectral envelope

Log frequency scale

Insufficient averaging of the noise spectrum can mislead as to the long term spectrum shape of the noise signals shown. The white noise spectrum appears to rise with increasing frequency, yet it should be flat (in theory all frequency components at equal amplitude).
D.R Campbell School of Computing University of Paisley 7

White and Pink Noise Spectra

10 dB White noise long term Spectrum

One decade

Log frequency scale

When sufficient averages are taken the flat long term average spectrum of white noise is revealed and here, viewed on a log frequency scale, the power spectrum of pink noise rolls off at 3dB/octave (10 dB/decade) i.e. it contains a greater proportion of low frequencies than white noise and sounds less hissy.
D.R Campbell School of Computing University of Paisley 8

Pink Noise A related random noise signal also used for audio purposes is pink noise which can be formed from white noise by low-pass filtering. Pink noise has a spectral envelope which is not flat with frequency but rolls off at higher frequencies. Since it contains a greater relative proportion of low frequency energy than white noise and sounds less hissy. Its power is proportional to 1/f i.e. P = K/f, which when plotted on a log (dB) scale gives a slope of 3dB/Octave. So instead of equal power for equal frequency increment, as with white noise, pink noise has equal power per octave. Suppose a pink noise has noise power P1 and P2 at frequencies f1 and f2 where P1 = K/f1 and P2 = K/f2 , if f2 = 2f1 (i.e. one octave above) then from the definition of dB the relative power is, 10log10(P2/P1) = 10log10((K/f2)/(K/f1)) = 10log10(f1/f2) = 10log10(f1/(2f1)) = 10log10(0.5) = -3dB So every time frequency is doubled the power or energy goes down 3dB. Pink noise is particularly useful when making measurements to assess the acoustic properties of a performance space using a 1/3 octave spectrum analyser. This type of analyser operates with a constant percentage bandwidth, and the bandwidth of its filters gets wider towards the high frequencies in a way which is similar to human hearing. If white noise was used with such an analyser, the higher frequency readings would be unnaturally dominant and give a false impression of what an average human listener would experience.