Spectralextractor

Spectralextractor uses frequency variation to discriminate between pitch and noise. The frequency of each bin is tracked and used to create a signal measuring the rate at which the bin is changing frequency; the higher this rate of frequency of change the more the bin is associated with noise rather than pitch components—that is, the bin’s frequency instability becomes a correlate for noise. A frequency change threshold is set to extract pitch or noise;  when the rate of change falls below the threshold the bin is identified with pitch components, above, with noise. A response time control (lowpass filter) slows the rate at which the signal is allowed to cross the threshold, thereby preventing gurgle noise. A lowpass filter (threshold accumulator) is applied to the tracked bin frequency as well to smooth out artifacts from the process. While specrtalextractor functions differently than spectwarper, its results are very, very similar. I prefer spectwarper which has less grit to it, although both are interesting. Spectralextractor is newer—a work-in-progress—and requires more refinement—its parameters of control are not yet general enough, changing their effect when the FFT size is modified (beware).

Back to Main


Amplitude Change Response Time in Seconds
Amplitude Reports Print Mode
Analysis Frames per Second
Begin Time in Seconds
End Time in Seconds
EQ - Low Shelf Gain
EQ - High Shelf Gain
EQ - Low Shelf Frequency
EQ - High Shelf Frequency
Extract Periodic or Noise
FFT Length
Frequency Change Threshold
Frequency Shift Factor
Gain in Decibels
Oscillator Resynthesis Threshold in Decibels
Output Format
Peak Rescale Level
Pitch Transposition in Semitones
Resynthesis Channel
Spectrum Warpshape Index
Threshold Accumulator Response Time in Seconds
Time Expansion/Contraction Factor
Time Interval Between Reports
Window Size in Samples
Window Type

 

Amplitude Change Response Time in Seconds


Amplitude Reports Print Mode

Two flags are provided for controlling the output amplitude statistics; one turns the statistics on or off, and the other sets how often they will be reported. The statistics provide the peak output level in amplitude and decibels. With integer format output files, output values exceeding the normalized peak amplitude of 1. (0 dB) are clipped to a value of 1.0, and the statistics placed in clip mode; in clip mode reports are made only for frames where clipping occurs. The peak amplitude, its time, and the number of clipped samples are reported at the end of processing. With floating-point format output files, output values exceeding the normalized peak amplitude of 1. are not clipped since they will be rescaled in the second pass; output statistics proceed normally throughout. The levels before and after rescaling are reported at the end of processing.

0 turns amplitude reports off, 1 turns them on.


Analysis Frames per Second

This controls how often the phase vocoder will perform an analysis on the signal. It is a translation of the classic decimation control that specifies how many samples to skip between analysis frames. More frames increases the resolution of time but decrease speed. 200 frames per second is a good reference point. If you expand time you should increase this proportionately to maintain about 200 or more frames per second.


Begin Time in Seconds

The time, in seconds, at which to begin processing the soundfile.


End Time in Seconds

The time, in seconds, at which to stop processing the soundfile. 0 or less is equivalent to the duration of the soundfile.


Low/High Shelf Equalization

Equalization has been provided at various points in routines to allow for the needed adjustment of spectra. The EQ consists of low and hi shelf segments, whose width is adjusted through control of the shelf breakpoint frequency. The region between the shelf segments is represented by a linear decibel gradient between the decibel levels of the two shelves. Some routines implement the EQ before pitch changes, others after. EQ placed before pitch changes (pre-transpose/shift) will cause the EQ to be transposed with the pitch changes, whereas afterwards (post-transpose/shift) will keep them fixed as shifts and transpositions occur.

Low Shelf Gain

Determines how the amplitude of sounds below the low shelf frequency will be affected.

High Shelf Gain

Determines how the amplitude of sounds above the high shelf frequency will be affected.

Low Shelf Frequency

Determines the frequency below which the low shelf gain will be used.

High Shelf Frequency

Determines the frequency above which the high shelf gain will be used.


FFT Length

The FFT size must be a power of 2. Larger FFT sizes resolve frequencies better but transient behavior more poorly. Choose your FFT size according to the sound you are working with. A size of 1024 or 2048 works well in most cases.


Frequency Change Threshold


Frequency Shift Factor

With the frequency shift control, a constant or function value is added to all the bin frequencies to produce a nonlinear pitch domain translation of the spectrum. Frequency shift is related to things like ring modulation and their similarly nonlinear shifts of pitch characteristics. Use this to create small distortions of the harmonic integrity of a sound.


Gain in Decibels

The output and other components can be gained. 0 dB represents unity gain, no change. A change of +/- 6 dB represents a doubling or halving of the amplitude. Increments of 10 dB are loosely associated with one change in dynamic level.


Oscillator Resynthesis Threshold in Decibels

The phase vocoder resynthesizes the signal using one of two methods, depending on the type of changes made to the FFT. If the changes are only to the magnitudes (amplitudes), then the faster overlap/add method is used. If however changes in frequency are made, then the FFT integrity is compromised, necessitating use of the oscillator bank method in which each bin is synthesized as a sine wave changing in frequency and amplitude. This method is slower, although a resynthesis threshold is available that can be used to increase the computation speed by turning off bins whose amplitude falls below the threshold. A threshold of -60dB is appropriate, although safety warrants using a lower threshold if the spectrum is thin and its decays exposed; use your ear.


Output Format

The output sound file is written as a NeXT/Sun format sound file in either 16-bit short or 32-bit floating point format, of one or more channels. The channels are processed one at a time beginning with the first channel. The first pass writes zeros in the channels yet to be processed, replacing them when processing proceeds to those channels.

0 tells PVCX to use the format of the input file, 1 equals integer format, and 2 equals rescaled floats.


Peak Rescale Level

Selection of the floating-point, output-file format invokes an amplitude rescaling feature. Once processing is complete, a second pass through the sound file is made to rescale the values to the decibel level specified. A dB rescale level of 1 causes rescaling to the level of the original input file.


Pitch Transposition in Semitones

With the pitch transposition control, a constant or function value is multiplied against all bin frequncies. This is classic transposition, here specified in semitones of transposition (12 semitones equals an octave). Conversion is made to produce the appropriate frequency multiplier.


Resynthesis Channel

All routines allow both monophonic and multi-channel input files to be processed. With multi-channelled files, you can either select one channel and produce a monophonic output file, or process all the channels. Channels are numbered beginning with 1. Processing of multi-channelled files is done one channel at a time beginning with channel 1, with zeros written to channels which have yet to be processed. Processing one channel at a time requires less memory and allows you to audition the output sooner than if you did all channels at once.

Use 0 to process all channels.


Spectrum Warpshape Index

Many of the routines employ the principle of warping in which a distribution of values is transformed by an identity function. In these places an exponential function is employed to remap a 0-1 range of values into a new orientation that preserves the minima (0) and maxima (1) while bringing the distribution closer to either extreme as a result of the curvature of the exponential function selected. The curvature of the exponential function is selected through a warp index. Specifically, warp index w will reorient the input x through the function below (^ = exponentiation).

y = (1. - (e^(x * w))) / (1. - (e^w))

In this function, the warp index of 0 produces a linear function and an untransformed output. Positive warp index values of increasing magnitude produce curves of increasing concavity (increasing slope) that draw values towards the 0-valued minima, and reduce the function integral. Negative values do the opposite, drawing values towards the maxima of 1, increasing the integral.

The practical use of this mechanism is found in various places. One such place is the reshaping of the frequency response distribution characteristics. In this, positive warp indeces cause the peaks of the response to be accentuated while the weaker frequencies are expanded out (i.e. pushed towards 0). Negative values have the opposite effect as they compress the dynamic range of the response and raise the relative level of the weaker noise components. Another place where warp applies is in the remapping of FFT amplitudes through the spectrum warpshape. In this, the sucessive FFT frames have their amplitudes remapped by the identity function, similiarly expanding or compressing the dynamic range depending upon the warp specified; 0 (linear warp function) leaves the amplitudes unchanged.


Threshold Accumulator Response Time in Seconds


Time Expansion/Contraction Factor

Once the spectral modifications are made to the FFT analysis, an inverse FFT is invoked to produce the samples of a time-domain signal. The classic phase vocoder paradigm controls the number of samples through the interpolation value and its relation to the decimation. The arcane relationship of decimation and interpolation is here translated into the parameter of time expansion/contraction, allowing for the direct scaling of time. Use values greater than 1 to expand time, less than 1 contract it.


Time Interval Between Reports

Determines the interval in seconds of the soundfile between amplitude reports. See Amplitude Reports Print Mode for a further explaination.


Window Size in Samples

The window size is a less opaque parameter; like the FFT, it must be a power of 2. Windows twice the size of the FFT work well. Larger window sizes may resolve frequencies better. Specifying 0 for the window size will automatically set the window to twice the FFT size.


Window Type

The FFT and inverse FFT are computed using a window. Like the FFT size, the shape of the window used can effect the quality of the analysis and resynthesis. (See F.R.Moore, Stieglitz, or Roads for further explanation.) A variety of windows are available including: Hamming, Rectangular, Blackman, Triangular, and Kaiser (in 8 different forms as related to 8 different alpha values). Blackman (-w2) or Kaiser (-w8) are recommended for most applications. In some unusual cases where transient behavior is being lost, consider using other windows such as the Rectangular, although take care to assure that it is not producing pops or a buzzy sound.