VOICEBOX: Speech Processing Toolbox for MATLAB

Introduction

VOICEBOX is a speech processing toolbox consists of MATLAB routines that are maintained by and mostly written by Mike Brookes, Department of Electrical & Electronic Engineering, Imperial College, Exhibition Road, London SW7 2BT, UK. Several of the routines require MATLAB V6.5 or above and require (normally slight) modification to work with earlier veresions.

The routines are available as a zip archive and are made available under the terms of the GNU Public License.

The routine VOICEBOX.M contains various installation-dependent parameters which may need to be altered before using the toolbox. In particular it contains a number of default directory paths indicating where temporary files should be created, where speech data normally resides, etc. See the comments in voicebox.m for a fuller description.

For reading compressed SPHERE format files, you will need the SHORTEN program written by Tony Robinson and SoftSound Limited www.softsound.com. The path to the shorten executable must be set in voicebox.m.

Please send any comments, suggestions, bug reports etc to mike.brookes@ic.ac.uk.

Audio File Input/Output: Read and write WAV and other speech file formats
Frequency Scales: Convert between Hz, Mel, Erb and MIDI frequency scales
Fourier/DCT/Hartley Transforms: Various related transforms
Random Number and Probability Distributions: Generate random vectors and noise signals
Vector Distances: Calculate distances between vector lists
Speech Analysis: Active level estimation, Spectrograms
LPC Analysis of Speech: Linear Predictive Coding routines
Speech Synthesis: Glottal waveform models
Speech Enhancement: Spectral noise subtraction
Speech Coding: PCM coding, Vector quantisation
Speech Recognition: Front-end processing for recognition
Signal Processing: Miscellaneous signal processing functions
Information Theory: Routines for entropy calculation and symbol codes
Computer Vision: Routines for 3D rotation
Printing and Display Functions: Utilities for printing and graphics
Voicebox Parameters and System Interface: Get or set VOICEBOX and WINDOWS system parameters
Utility Functions: Miscellaneous utility functions

Audio File Input/Output

Routines are available to read and, in some cases write, a variety of file formats:

Read Write Suffix

readwav writewav .wav These routines allow an arbitrary number of channels and can deal with linear PCM (any precision up to 32 bits), A-law PCM and Mu-law PCM. Large files can be read and written in small chunks.

readhtk writehtk .htk Read and write waveform and parameter files used by Microsoft's Hidden Markov Toolkit.

readsfs .sfs Speech Filing system files from Mark Huckvale at UCL.

readsph .sph NIST Sphere format files (including TIMIT). Needs SHORTEN for compressed files.

readaif .aif AIFF format (Audio Interchange File Format) used by Mac users.

readcnx cnx Read Connex database files (from BT)

readau au Read AV audio files (from Sun)

Frequency Scale Conversion

From f	To f	Scale
frq2mel	mel2frq	mel	The mel scale is based on the human perception of sinewave pitch.
frq2erb	erb2frq	erb	The erb scale is based on the equivalent rectangular bandwidths of the human ear.
frq2erb	erb2frq	bark	The bark scale is based on critical bands and masking in the human ear.
frq2midi	midi2frq	midi	The midi standard specifies a numbering of semitones with middle C being 60. They can use the normal equal tempered scale or else the pythagorean scale of just intonation. They will in addition output note names in a character format.

Fourier, DCT and Hartley Transforms

Forward	Inverse
rfft	irfft	Forward and inverse discrete fourier transforms on real data. Only the first half of the conjugate symmetric transform is generated. For even length data, the inverse routine is asumptotically twice as fast as the built-in MATLAB routine.
rsfft		Forward transform of real, symmetric data to give the first half only of the real, symmetric transform.
zoomfft		Calculate the discrete fourier transform at an arbitrary set of linearly spaced frequencies. Can be used to zoom into a subset of the full frequency range.
rdct	irdct	Forward and inverse discrete cosine transform on real data.
rhartley	rhartley	Hartley transform on real data (forward and inverse transforms are the same).

Random Numbers and Probability Distributions

Random Number Generation

randvec	generates random vectors from gaussian or lognormal mixture distributions.
randiscr	generates discrete random values with a specified probability vector
stdspectrum	generates noise samples or filter coefficients for a variety of standard spectra including: A, B, C or BS468 weighting, USASI noise, POTS spectrum, LTASS, Internal masking noise (from SII spec)
randfilt	generates filtered gaussian noise without any startup transients.
rnsubset	selects a random subset of k elements from the numbers 1:n

Probability Density Functions

lognmpdf	calculates the pdf of a lognormal distribution
gaussmix	generates a multivariate Gaussian mixture model (GMM) from training data
gaussmixp	calculates full and marginal log probability and relative mixture probabilities from a GMM
gaussmixd	determines marginal and conditional distributions from a GMM and can be used to perform inference on unobserved variables.

Miscellaneous

histndim	calculates an n-dimensional histogram (and plots a 2-D one)
gausprod	calculates the product of two gaussian distributions
maxgauss	calculates the mean and variance of the maximum element of a gaussian vector

Vector Distance

disteusq calculates the squared euclidean distance between all pairs of rows of two matrices.

distitar calculates the Itakura spectral distances between sets of AR coefficients.

distitpf calculates the Itakura spectral distances between power spectra.

distisar calculates the Itakura-Saito spectral distances between sets of AR coefficients.

distispf calculates the Itakura-Saito spectral distances between power spectra.

distchar calculates the COSH spectral distances between sets of AR coefficients.

distchpf calculates the COSH spectral distances between power spectra.

Speech Analysis

enframe can be used to split a signal up into frames. It can optionally apply a window to each frame.

overlapadd Join frames up using overlap-add processing. Commonly used with enframe.

fram2wav interpolates a sequence of frame-based value into a waveform

ewgrpdel calculates the energy-weighted group delay waveform.

activlev calculates the active level of a speech segment according to ITU-T recommendation P.56.

spgrambw draws a monochrome spectrogram with a dB scale.

txalign finds the best alignment (in a least squares sense) between two sets of time markers (e.g. glottal closure instants).

dypsa estimates the glottal closure instants from the speech waveform.

fxrapt is an implementation of the RAPT pitch tracker by David Talkin.

soundspeed gives the speed of sound as a function of temperature

importsii calculate the SII importance function

LPC Analysis of Speech

lpcauto &
lpccovar perform linear predictive coding (LPC) analysis. The routines relating to LPC are described in more detail on another page. A large number of conversion routines are included for changing the form of the LPC coefficients (e.g. AR coefficients, reflection coefficients etc.): these are of the form lpcxx2yy where xx and yy denote the coefficient sets.

lpcrr2am calculates LPC filters for all orders up to a given maximum.

lpcbwexp performs bandwidth expansion on an LPC filter.

ccwarpf performs frequency warping in the complex cepstrum domain.

lpcifilt performs inverse filtering to estimate the glottal waveform from the speech signal and the lpc coefficients.

lpcrand can be used to generate random, stable filters for testing purposes.

Speech Synthesis

glotros Calculates the Rosenberg model of the glottal flow waveform

glotlf Calculates the Liljencrants-Fant model of the glottal flow waveform

Speech Enhancement

estnoisem uses a minimum-statistics algorithm to estimate the noise spectrum from a noisy speech signal that has been divided into frames.

specsub performs speech enhancement using spectral subtraction

ssubmmse performs speech enhancement using the MMSE or log MMSE criteria

Speech Coding

lin2pcma converts an audio waveform to 8-bit A-law PCM format

lin2pcmu converts an audio waveform to 8-bit mu-law PCM format

pcma2lin converts 8-bit A-law PCM to a waveform

pcmu2lin converts 8-bit mu-law PCM to a waveform

kmeans vector quantisation using the K-means algorithm

kmeanlbg vector quantisation using the LBG algorithm

kmeanhar vector quantisation using the K-harmonic means algorithm

potsband calculates a bandpass filter corresponding to the standard telephone passband.

Speech Recognition

melcepst implements a mel-cepstrum front end for a recogniser

melbankm constructs a bandpass filterbank with mel-spaced centre frequencies

cep2pow converts multivariate Gaussian means and covariances from the log power or cepstral domain to the power domain

pow2cep converts multivariate Gaussian means and covariances from the power domain to the log power or cepstral domain

ldatrace performs Linear Discriminant Analysis with optional constraints on the transform matrix

Signal Processing

findpeaks finds the peaks in a signal

maxfilt performs running maximum filter

meansqtf calculates the output power of a rational filter with a white noise input

windows generates window functions

windinfo calculate window properties and figures of merit

zerocros finds the zero crossings of a signal with interpolation

ditherq adds dither and quantizes a signal

schmitt passes a signal through a schmitt trigger having hysteresis

dlyapsq solves the discrete lyapunov equation using an efficient square root algorithm

momfilt generate running moments from a signal

Information Theory

huffman	calculates optimum D-ary symbol code from a probability mass vector
entropy	calculates entropy and conditional entropy for discrete and continuous distributions

Computer Vision

rot--2--	converts between the following representations of rotations: rotation matrix (ro), euler angles (eu), axis of rotation (ax), plane of rotation (pl), real quaternion vector (qr), real quaternion matrix (mr), complex quaternion vector (qc), complex quaternion matrix (mc). A detailed description is given here.
peak2dquad	find a quadratically-interpolated peak in a 2D array by fitting a biquadratic function to the array values
polygonarea	Calculates the area of a polygon
polygonwind	Determines whether points are inside or outside a polygon
polygonxline	Determines where a line crosses a polygon

Printing and Display Functions

figbolden	makes the lines on a figure bold and enlarges font sizes for printing clearly
xticksi	Label the x-axis tick marks using SI multipliers for large and small values. Particularly useful for logarithmic plots.
yticksi	Label the y-axis tick marks using SI multipliers for large and small values. Particularly useful for logarithmic plots.
sprintsi	prints a value with the correct standard SI multiplier (e.g. 2100 prints as 2.1 k)
bitsprec	rounds values to a precision of n bits
frac2bin	converts numbers to fixed-point binary strings

Voicebox Parameters and System Interface

voicebox	contains a number of installation-dependent global parameters and is likely to need editing for each particular setup.
unixwhich	searches the WINDOWS system path for an executable (like UNIX which command)
winenvar	Obtains WINDOWS environment variables

Utility Functions

zerotrim removes from a matrix any trailing rows and columns that are all zero.

logsum calculates log(sum(exp(x))) without overflow problems.

dualdiag simultaneously diagonalises two matrices: this is useful in computing LDA or IMELDA transforms.

permutes all possible permutations of the numbers 1:n

choosenk all possible ways of choosing k elements out of the numbers 1:n without duplications

choosrnk all possible ways of choosing k elements out of the numbers 1:n with duplications allowed

rotation generates rotation matrices

skew3d manipulates 3#3 skew symmetric matrices

atan2sc arctangent function that returns the sin and cos of the angle

bitsprec Rounds values to a precision of n bits

dlyapsq Solve the discrete lyapunov equation

finishat Estimate the finishing time of a long loop

m2htmlpwd Create HTML documentation of matlab routines in the current directory

Read	Write	Suffix
readwav	writewav	.wav	These routines allow an arbitrary number of channels and can deal with linear PCM (any precision up to 32 bits), A-law PCM and Mu-law PCM. Large files can be read and written in small chunks.
readhtk	writehtk	.htk	Read and write waveform and parameter files used by Microsoft's Hidden Markov Toolkit.
readsfs		.sfs	Speech Filing system files from Mark Huckvale at UCL.
readsph		.sph	NIST Sphere format files (including TIMIT). Needs SHORTEN for compressed files.
readaif		.aif	AIFF format (Audio Interchange File Format) used by Mac users.
readcnx		cnx	Read Connex database files (from BT)
readau		au	Read AV audio files (from Sun)

disteusq	calculates the squared euclidean distance between all pairs of rows of two matrices.
distitar	calculates the Itakura spectral distances between sets of AR coefficients.
distitpf	calculates the Itakura spectral distances between power spectra.
distisar	calculates the Itakura-Saito spectral distances between sets of AR coefficients.
distispf	calculates the Itakura-Saito spectral distances between power spectra.
distchar	calculates the COSH spectral distances between sets of AR coefficients.
distchpf	calculates the COSH spectral distances between power spectra.

enframe	can be used to split a signal up into frames. It can optionally apply a window to each frame.
overlapadd	Join frames up using overlap-add processing. Commonly used with enframe.
fram2wav	interpolates a sequence of frame-based value into a waveform
ewgrpdel	calculates the energy-weighted group delay waveform.
activlev	calculates the active level of a speech segment according to ITU-T recommendation P.56.
spgrambw	draws a monochrome spectrogram with a dB scale.
txalign	finds the best alignment (in a least squares sense) between two sets of time markers (e.g. glottal closure instants).
dypsa	estimates the glottal closure instants from the speech waveform.
fxrapt	is an implementation of the RAPT pitch tracker by David Talkin.
soundspeed	gives the speed of sound as a function of temperature
importsii	calculate the SII importance function

lpcauto & lpccovar	perform linear predictive coding (LPC) analysis. The routines relating to LPC are described in more detail on another page. A large number of conversion routines are included for changing the form of the LPC coefficients (e.g. AR coefficients, reflection coefficients etc.): these are of the form lpcxx2yy where xx and yy denote the coefficient sets.
lpcrr2am	calculates LPC filters for all orders up to a given maximum.
lpcbwexp	performs bandwidth expansion on an LPC filter.
ccwarpf	performs frequency warping in the complex cepstrum domain.
lpcifilt	performs inverse filtering to estimate the glottal waveform from the speech signal and the lpc coefficients.
lpcrand	can be used to generate random, stable filters for testing purposes.

glotros	Calculates the Rosenberg model of the glottal flow waveform
glotlf	Calculates the Liljencrants-Fant model of the glottal flow waveform

estnoisem	uses a minimum-statistics algorithm to estimate the noise spectrum from a noisy speech signal that has been divided into frames.
specsub	performs speech enhancement using spectral subtraction
ssubmmse	performs speech enhancement using the MMSE or log MMSE criteria

lin2pcma	converts an audio waveform to 8-bit A-law PCM format
lin2pcmu	converts an audio waveform to 8-bit mu-law PCM format
pcma2lin	converts 8-bit A-law PCM to a waveform
pcmu2lin	converts 8-bit mu-law PCM to a waveform
kmeans	vector quantisation using the K-means algorithm
kmeanlbg	vector quantisation using the LBG algorithm
kmeanhar	vector quantisation using the K-harmonic means algorithm
potsband	calculates a bandpass filter corresponding to the standard telephone passband.

melcepst	implements a mel-cepstrum front end for a recogniser
melbankm	constructs a bandpass filterbank with mel-spaced centre frequencies
cep2pow	converts multivariate Gaussian means and covariances from the log power or cepstral domain to the power domain
pow2cep	converts multivariate Gaussian means and covariances from the power domain to the log power or cepstral domain
ldatrace	performs Linear Discriminant Analysis with optional constraints on the transform matrix

findpeaks	finds the peaks in a signal
maxfilt	performs running maximum filter
meansqtf	calculates the output power of a rational filter with a white noise input
windows	generates window functions
windinfo	calculate window properties and figures of merit
zerocros	finds the zero crossings of a signal with interpolation
ditherq	adds dither and quantizes a signal
schmitt	passes a signal through a schmitt trigger having hysteresis
dlyapsq	solves the discrete lyapunov equation using an efficient square root algorithm
momfilt	generate running moments from a signal

zerotrim	removes from a matrix any trailing rows and columns that are all zero.
logsum	calculates log(sum(exp(x))) without overflow problems.
dualdiag	simultaneously diagonalises two matrices: this is useful in computing LDA or IMELDA transforms.
permutes	all possible permutations of the numbers 1:n
choosenk	all possible ways of choosing k elements out of the numbers 1:n without duplications
choosrnk	all possible ways of choosing k elements out of the numbers 1:n with duplications allowed
rotation	generates rotation matrices
skew3d	manipulates 3#3 skew symmetric matrices
atan2sc	arctangent function that returns the sin and cos of the angle
bitsprec	Rounds values to a precision of n bits
dlyapsq	Solve the discrete lyapunov equation
finishat	Estimate the finishing time of a long loop
m2htmlpwd	Create HTML documentation of matlab routines in the current directory