Speech Software

Note: The following software packages are, or were at some point, maintained by members of the ICSI Speech Research Group. Unless stated otherwise in the documentation, you should assume that no support is available for this software.

cluster
A tool for performing hierarchical cluster analysis (HCA) on a set of vectors and outputing dendrograms in a variety of graphical formats. cluster also has an option to perform principal component analysis (PCA).
Author: A. Stolcke
Download source code

dpwelib
A collection of sound-processing libraries and applications, including:
sndf - a low-level multi-format soundfile interface library
snd - a higher-level integrated interface to sndf soundfiles
audIO - a portable audio hardware input/output interface library
sndplay - a simple soundfile play program
sndrecord - a simple program for recording soundfiles from audio hardware
sndcmp - for gathering statistics on a soundfile, or the difference between two soundfiles.
sndrfmt - change the format (sample type, number of channels) of a soundfile.
sndrsmp - change the sampling rate (resample/interpolate) a soundfile.
Author: D. Ellis
Download source code

feacalc
A feature calculation package to process audio files for statistical pattern recognition research.
Author: D. Ellis
Download source code,

feacat
A tool for "cat'ing" feature files, although it does many other things as well as just copying the input to the output, including a wide range of file format conversion funtions and subset extraction specifiers.
Author: D. Ellis
Download source code,

noway
Noway is a start-synchronous decoder, based on a stack decoder algorithm.
Author: S. Renals
Download source code

pfile_utils
A collection of utilities for manipulating "PFiles", the ICSI feature file archive format.
Author: J. Bilmes
Download source code

PraatLib
A C++ library wrapper to Praat, a prosodic feature calculation toolkit. PraatLib allows to calculate a number of prosodic features and takes care of interfeature depencies. This means, if a feature depends on another, it tries to make sure that each features is only calculates once to save CPU time. So the main strength of the library lies in calculating a set of features at once. Praatlib currently supports over 70 different features, which are thoroughly documented in the Praat help.
Authors: M. Feld, G. Friedland, et al.
License: GPL version 2
Download source code

QuickNet
A set of programs and libraries that can be used to train MLPs (multi-layer perceptrons, a type of "neural net"). QuickNet is focused on high-performance MLP training and use for audio processing, although it may be applicable to other domains.
Author: D. Johnson
Download source code, test data

rasta
Computes variants of the Perceptual Linear Prediction (PLP) analysis, including several forms of RASTA processing.
Author: N. Morgan
Download source code
ICSIlog()
ICSI log is a fast logarithm C code implementation which contains a C implementation fo the ICSI logarithm function. ICSI log implements a fast single precistion approximation of hte natural logarithm with adjustable accuracy.
The funct6ion uses bi level operations and a look up tabl of adjustable size. The accuracy is defined by the tabl isze. However, the execution speed of the logarithm reduces as the table grows and does not fit into the cache anymore.
Download source code