The preceding chapters have been concerned with the spectrum analysis of sinusoids and noise at a particular point in time (or a single spectrum for all time). This chapter introduces the Short-Time Fourier Transform (STFT)--a time-ordered sequence of spectral estimates, each using a finite-length analysis window. The STFT is used to compute the classic spectrogram, used extensively for speech and audio signals in general [132,57,199,162,226,74,81]. Finally, we point to methods for making spectrograms correspond better to audio perception, so that what you see is what you hear, to a greater extent. In particular, a loudness spectrogram based on a psychoacoustic model of time-varying loudness perception is described.
The Short-Time Fourier TransformThe Short-Time Fourier Transform (STFT) (or short-term Fourier transform) is a powerful general-purpose tool for audio signal processing [7,9,8]. It defines a particularly useful class of time-frequency distributions  which specify complex amplitude versus time and frequency for any signal. We are primarily concerned here with tuning the STFT parameters for the following applications:
- Approximating the time-frequency analysis performed by the ear for purposes of spectral display.
- Measuring model parameters in a short-time spectrum.
then the sum of the successive DTFTs over time equals the DTFT of the whole signal :
Practical Computation of the STFTWhile the definition of the STFT in (7.1) is useful for theoretical work, it is not really a specification of a practical STFT. In practice, the STFT is computed as a succession of FFTs of windowed data frames, where the window ``slides'' or ``hops'' forward through time. We now derive such an implementation of the STFT from its mathematical definition. The STFT in (7.1) can be rewritten, adding to , as
In this form, the data centered about time are translated to time 0, multiplied by the (let's assume zero-phase) window , and then the DTFT is performed. Since the nonzero portion of the windowed data is centered on time zero, the DTFT can be replaced by the DFT (or FFT). This effectively samples the DTFT in frequency. This sampling will not cause (time) aliasing if the number of samples around the unit circle is greater than the width (in samples) of the time interval including all nonzero datapoints. In other words, sampling the frequency axis is information-preserving when the signal is properly time limited.8.3Let denote the window length (typically an odd number) and be the DFT length (typically a power of 2). Then sampling (7.3) at , , and using the fact that the window is time-limited to less than samples centered about time zero, yields
Since indexing in the DFT is modulo , the sum over can be ``rotated'' to a sum from 0 to as is conventionally implemented for the DFT. In practice, this means that the right half of the windowed data frame goes at the beginning of the FFT input buffer, and the left half of the windowed frame goes at the end, with zero-padding in the middle (see Fig.2.6b on page for an illustration).
samples of the input signal
into a local buffer of
which is initially zeroed
- Multiply the data frame pointwise by a length
to obtain the
windowed data frame (time normalized):
with zeros on both sides to obtain a
where is chosen to be a power of two larger than . The number is the zero-padding factor. As discussed in §2.5.3, the zero-padding factor is the interpolation factor for the spectrum, i.e., each FFT bin is replaced by bins, interpolating the spectrum using ideal bandlimited interpolation , where the ``band'' in this case is the -sample nonzero duration of in the time domain.
- Take a length
to obtain the time-normalized,
frequency-sampled STFT at time
where , and is the sampling rate in Hz. As in any FFT, we call the bin number.
- If needed, time normalization may be removed using a
linear phase term to yield the sampled STFT:
The (continuous-frequency) STFT may be approached arbitrarily closely by using more zero padding and/or other interpolation methods. Note that there is no irreversible time-aliasing when the STFT frequency axis is sampled to the points , provided the FFT size is greater than or equal to the window length .
STFT can be viewed as a function of either frame-time or bin-frequency . We will develop both points of view in this book. At each frame time , the STFT can be regarded as producing a Fourier transform centered around that time. As advances, a sequence of spectral transforms is obtained. This is depicted graphically in Fig.9.1, and it forms the basis of the overlap-add method for Fourier analysis, modification, and resynthesis . It is also the basis for transform coders [16,284]. In an exact Fourier duality, each bin of the STFT can be regarded as a sample of the complex signal at the output of a lowpass filter whose input is . As discussed in §9.1.2, this signal is obtained from by frequency-shifting it so that frequency is translated down to 0 Hz. For each value of , the time-domain signal , for , is the output of the th ``filter bank channel,'' for . In this ``filter bank'' interpretation, the hop size can be interpreted as the downsampling factor applied to each bin-filter output, and the analysis window is seen as the impulse response of the anti-aliasing filter used prior to downsampling. The window transform is also the frequency response of each channel filter (translated to dc). This point of view is depicted graphically in Fig.9.2 and elaborated further in Chapter 9.
Short Time Fourier Transform (STFT) is a function of both time (frame number ) and frequency ( ). It is therefore an example of a time-frequency distribution. Others include 7.1. The window length is proportional to the resolution cell in time, indicated by the vertical lines in Fig.7.1. The width of the main-lobe of the window-transform is proportional to the resolution cell in frequency, indicated by the horizontal lines in Fig.7.1. As detailed in Chapter 3, choosing a window length and window type (Hamming, Blackman, etc.) chooses the ``aspect ratio'' and total area of the time-frequency resolution cells (rectangles in Fig.7.1). For an example of a non-uniform time-frequency tiling, see Fig.10.14.
matlab segment illustrates the above processing steps:
Xtwz = zeros(N,nframes); % pre-allocate STFT output array M = length(w); % M = window length, N = FFT length zp = zeros(N-M,1); % zero padding (to be inserted) xoff = 0; % current offset in input signal x Mo2 = (M-1)/2; % Assume M odd for simplicity here for m=1:nframes xt = x(xoff+1:xoff+M); % extract frame of input data xtw = w .* xt; % apply window to current frame xtwz = [xtw(Mo2+1:M); zp; xtw(1:Mo2)]; % windowed, zero padded Xtwz(:,m) = fft(xtwz); % STFT for frame m xoff = xoff + R; % advance in-pointer by hop-size R end
- The window w is implemented in zero-centered (``zero-phase'') form (see, e.g., §2.5.4 for discussion).
- The signal x should have at least Mo2 leading zeros for this (simplified) implementation.
- See §F.3 for a more detailed implementation.
Classic SpectrogramsThe spectrogram is a basic tool in audio spectral analysis and other applications. It has been used extensively in speech analysis . The spectrogram can be defined as an intensity plot (usually on a log scale, such as dB) of the Short-Time Fourier Transform (STFT) magnitude.8.4 As defined in the previous section, the STFT is simply a sequence of FFTs of windowed data segments, where the windows are allowed to overlap in time, typically by at least 50% . Parameters of the spectrogram include the
- window length ,
- window type (Hamming, Kaiser, etc.),
- hop-size , and
- FFT length .
Spectrogram of Speech7.2. It was generated using the Matlab code displayed in Fig.7.3. The function spectrogram is listed in §F.3. The spectrogram is computed as a sequence of FFTs of windowed data segments. The spectrogram is plotted within spectrogram using imagesc.
[y,fs,bits] = wavread('SpeechSample.wav'); soundsc(y,fs); % Let's hear it % for classic look: colormap('gray'); map = colormap; imap = flipud(map); M = round(0.02*fs); % 20 ms window is typical N = 2^nextpow2(4*M); % zero padding for interpolation w = hamming(M); spectrogram(y,N,fs,w,-M/8,1,60); title('Speech Sample Spectrogram'); colormap(imap);
- The harmonics should be resolved.
- Pitch and formant variations should be closely followed.
Audio SpectrogramsSince classic spectrograms  typically show log-magnitude intensity (dB) versus time and frequency, and since sound-pressure level in dB is roughly proportional to perceived loudness, at least at high levels [179,276,305], we can say that a classic spectrogram provides a reasonably good psychoacoustic display for sound, provided the window length has been chosen to be comparable to the ``integration time'' of the ear. However, there are several ways we can improve the classic spectrogram to obtain more psychoacoustically faithful displays of perceived loudness versus time and frequency:
- Loudness perception is closer to linearly related to amplitude at low loudness levels.
- Since the STFT offers only one ``integration time'' (the window length), it implements a uniform bandpass filter bank--i.e., spectral samples are uniformly spaced and correspond to equal bandwidths. The window transform gives the shape of each effective bandpass filter in the frequency domain. The choice of window length determines the common time- and frequency-resolution at all frequencies. Figure 9.14 shows a frequency-response overlay of all 5 channel filters created by a length 5 DFT using a zero-phase rectangular window. In the ear, however, time resolution increases and frequency resolution decreases at higher frequencies. Thus, the ear implements a non-uniform filter bank, with wider bandwidths at higher frequencies. In the time domain, the integration time (effective ``window length'') becomes shorter at higher frequencies.
Auditory Filter BanksAuditory filter banks are non-uniform bandpass filter banks designed to imitate the frequency resolution of human hearing [307,180,87,208,255]. Classical auditory filter banks include constant-Q filter banks such as the widely used third-octave filter bank. Digital constant-Q filter banks have also been developed for audio applications [29,30]. More recently, constant-Q filter banks for audio have been devised based on the wavelet transform, including the auditory wavelet filter bank . Auditory filter banks have also been based more directly on psychoacoustic measurements, leading to approximations of the auditory filter frequency response in terms of a Gaussian function , a ``rounded exponential'' , and more recently the gammatone (or ``Patterson-Holdsworth'') filter bank [208,255]. The gamma-chirp filter bank further adds a level-dependent asymmetric correction to the basic gammatone channel frequency response, thus providing a more accurate approximation to the auditory frequency response [112,111]. The output power from an auditory filter bank at a particular time defines the so-called excitation pattern versus frequency at that time [87,179,305]. It may be considered analogous to the average power of the physical excitation applied to the hair cells of the inner ear by the vibrating basilar membrane in the cochlea.8.6 The shape of the excitation pattern can thus be thought of as approximating the envelope of the basilar membrane vibration. The excitation pattern produced from an auditory filter bank, together with appropriate equalization (frequency-dependent gain) and nonlinear compression, can be used to define specific loudness as a function of time and frequency [306,305,177,182,88]. Because the channels of an auditory filter bank are distributed non-uniformly versus frequency, they can be regarded as a basis for a non-uniform sampling of the frequency axis. In this point of view, the auditory-filter frequency response becomes the (frequency-dependent) interpolation kernel used to extract a frequency sample at the filter's center frequency. See §7.3.3 below for further details.
loudness spectrogram is to display some psychoacoustic model of loudness versus time and frequency. Instead of specifying FFT window length and type, one specifies conditions of presentation, such as physical amplitude level in dB SPL, angle of arrival at the ears, etc. By default, it can be assumed that the signal is presented to both ears equally, and the listening level can be normalized to a ``comfortable'' value such as 70 dB SPL.8.7 A time-varying model of loudness perception has been developed by Moore and Glasberg et al. [87,182,88]. A loudness spectrogram based on this work may consist of the following processing steps:
- Compute a multiresolution STFT (MRSTFT) which approximates the frequency-dependent frequency and time resolution of the ear. Several FFTs of different lengths may be combined in such a way that time resolution is higher at high frequencies, and frequency resolution is higher at low frequencies, like in the ear. In each FFT, the frequency resolution must be greater than or equal to that of the ear in the frequency band it covers. (Even ``much greater'' is ok, since the resolution will be reduced down to what it should be by smoothing in Step 2.)
- Form the excitation pattern from the MRSTFT by resampling the FFTs of the previous step using interpolation kernels shaped like auditory filters. The new spectral sampling intervals should be proportional to the width of a critical band of hearing at each frequency. The shape of each interpolation kernel (auditory filter) should change with amplitude level as well as center frequency . This step effectively converts the uniform filter bank of the FFT to an auditory filter bank.8.8
- Compute the specific loudness from the excitation pattern for each frame. This step implements a compressive nonlinearity which depends on the frequency and level of the excitation pattern . The specific loudness can be interpreted as loudness per ERB.
- If desired, the instantaneous loudness can be computed as the the sum of the specific loudness over all frequency samples at a fixed time. Similarly, short- and long-term time-varying loudness estimates can be computed as lowpass-filterings of the instantaneous loudness over time .
Loudness Spectrogram ExamplesWe now illustrate a particular Matlab implementation of a loudness spectrogram developed by teaching assistant Patty Huang, following [87,182,88] with slight modifications.8.9
Multiresolution STFTFigure 7.4 shows a multiresolution STFT for the same speech signal that was analyzed to produce Fig.7.2. The bandlimits in Hz for the five combined FFTs were , where the last two (in parentheses) were not used due to the signal sampling rate being only kHz. The corresponding window lengths in milliseconds were , where, again, the last two are not needed for this example. Our hop size is chosen to be 1 ms, giving 75% overlap in the highest-frequency channel, and more overlap in lower-frequency channels. Thus, all frequency channels are oversampled along the time dimension. Since many frequency channels from each FFT will be combined via smoothing to form the ``excitation pattern'' (see next section), temporal oversampling is necessary in all channels to avoid uneven weighting of data in the time domain due to the hop size being too large for the shortened effective time-domain windows.
7.5 shows the result of converting the MRSTFT to an excitation pattern [87,182,108]. As mentioned above, this essentially converts the MRSTFT into a better approximation of an auditory filter bank by non-uniformly resampling the frequency axis using auditory interpolation kernels. Note that the harmonics are now clearly visible only up to approximately 20 ERBs, and only the first four or five harmonics are visible during voiced segments. During voiced segments, the formant structure is especially clearly visible at about 25 ERBs. Also note that ``pitch pulses'' are visible as very thin, alternating, dark and light vertical stripes above 25 ERBs or so; the dark lines occur just after glottal closure, when the voiced-speech period has a strong peak in the time domain.
Nonuniform Spectral ResamplingRecall sinc interpolation of a discrete-time signal :
- Width of interpolation kernel (main-lobe width) 1/width-in-other-domain
- Shape of interpolation kernel gain profile (window) in other domain
auditory filter banks was introduced in §7.3.1 above. In this implementation, the auditory filters were synthesized from the Equivalent Rectangular Bandwidth (ERB) frequency scale, discussed in §E.5. The auditory filter-bank shapes are a function of level, so, ideally, the true physical amplitude level of the input signal at the ear(s) should be known. The auditory filter shape at 1 kHz in response to a sinusoidal excitation for a variety of amplitudes is shown in Fig.7.6.
Specific LoudnessFigure 7.7 shows the specific loudness computed from the excitation pattern of Fig.7.5. As mentioned above, it is a compressive nonlinearity that depends on level and also frequency.
7.8 shows all four previous spectrogram figures in a two-by-two matrix for ease of cross-reference.
Instantaneous, Short-Term, and Long-Term LoudnessFinally, Fig.7.9 shows the instantaneous loudness, short-term loudness, and long-term loudness functions overlaid, for the same speech sample used in the previous plots. These are all single-valued functions of time which indicate the relative loudness of the signal on different time scales. See  for further discussion. While the lower plot looks reasonable, the upper plot (in sones) predicts only three audible time regions. Evidently, it corresponds to a very low listening level.8.10 The instantaneous loudness is simply the sum of the specific loudness over all frequencies. The short- and long-term loudnesses are derived by smoothing the instantaneous loudness with respect to time using various psychoacoustically motivated time constants . The smoothing is nonlinear because the loudness tracks a rising amplitude very quickly, while decaying with a slower time constant.8.11 The loudness of a brief sound is taken to be the local maximum of the short-term loudness curve. The long-term loudness is related to loudness memory over time. The upper plot gives loudness in sones, which is based on loudness perception experiments ; at 1 kHz and above, loudness perception is approximately logarithmic above 50 dB SPL or so, while below that, it tends toward being more linear. The lower plot is given in phons, which is simply sound pressure level (SPL) in dB at 1 kHz [276, p. 111]; at other frequencies, the amplitude in phons is defined by following an ``equal-loudness curve'' over to 1 kHz and reading off the level there in dB SPL. This means, e.g., that all pure tones have the same perceived loudness when they are at the same phon level, and the dB SPL at 1 kHz defines the loudness of such tones in phons.
signals. All were implemented in terms of the short-time Fourier transform (STFT). The classical spectrogram was reviewed, and its performance on a speech sample was illustrated. A loudness spectrogram based on a model of time-varying loudness perception  was discussed. In this model, the STFT (or a multiresolution STFT), is smoothed and non-uniformly resampled in frequency to approximate an auditory filter bank, whose power output is taken to be the excitation pattern. A compressive nonlinearity is then applied to produce the specific loudness, which we took as our loudness spectrogram. The specific loudness can be optionally smoothed with respect to time to form a short- or long-term loudness spectrogram. Summing over frequency yields the corresponding loudness functions versus time. FFT-based non-uniform filter banks, providing more efficient loudness spectrograms, are discussed in §10.7.
Overlap-Add (OLA) STFT Processing
Spectrum Analysis of Noise