poser@csli.Stanford.EDU (Bill Poser) (04/08/90)
I have sent a fairly lengthy reply to this query by email. If anyone else is interested, please send me mail.
bryant@oakhill.UUCP (Bryant Wilder) (04/09/90)
>From: doug@wacsvax.OZ (Doug Robb) >Some researchers in the Psychology Department are looking for >software that will help them to do analysis of speech. >The sorts of things they would like to do is to capture >and edit speech (mark start and end of words or whatever) >and probably do spectral analysis. i just came back from icassp 90 in albuquerque, new mexico. in the ariel_ booth there was a very good tool for speech analysis. i don't remember the price, but it sounded low to me. the software is from a speech professor at mit. the dsp card used is the ariel_ dsp56001 half-height card for the pc. this is a slick system. ariel_'s phone number is (201) 249-2900. it's in new jersey. bryant wilder motorola dsp in oak hill, texas
poser@csli.Stanford.EDU (Bill Poser) (04/10/90)
I got so many requests for this information that I have
decided to post after all and have expanded and reorganized what
I originally sent out. I didn't think there would be so much interest.
There are three different sorts of products mentioned:
Audio i/o boards
DSP boards
speech analysis software
I'm not directly going to say anything about DSP boards. There
are a lot of them out there, and I'm not all that knowledgable about them.
They come up only because some of the audio i/o boards also have
DSP capability. For most people doing linguistic or
psychological work, DSP hardware is not very important, as the
work of designing the materials to be analyzed, obtaining the
recordings, making the necessary measurements, and so forth
completely dominates the time spent in actually performing
signal processing. For example, on any workstation these days
you can compute a single spectral slice virtually instanteously
in software. DSP hardware is more important to people doing
work in speech recognition (who do a lot of computing with the
same data and often don't measure much of anything) or using
computationally intensive signal processing (e.g. higher order
spectral analysis).
AUDIO I/O HARDWARE
One company to check out is Data Translations (addresses of all
companies mentioned are at the end of the message). They make a wide range of
data acquisition products, covering PCs and VME-bus machines (e.g. SUNs).
I know that they were planning to provide a UNIX device driver for their
VME-bus hardware but don't know what they actually have available.
Nirvonics at least used to sell drivers for Data Translations boards
and may still do so. (Nirvonics is a two-man company that may not be
very active anymore.)
Another source is Ariel, which has boards for PCs and for
a variety of workstations, some in hand, some in the works. At ICASSP
in Albuquerque last week they told me, for example, that they were working
on a board for HP 9000/300 series machines. Ariel provides both
audio i/o and DSP.
Software Research Corporation sells a board for PCs,
intended to go with their analysis software. I don't know if you can
get it unbundled.
For the Macintosh, you can get two sorts of hardware. At
least two companies sell very cheap (~US$100) ADDACs. These have only
8 bits resolution and are therefore not really appropriate for most
research purposes, at least research involving spectral analysis. GW
Instruments sells a 14-bit ADDAC together with a basic software
package (play, record, etc.). The hardware is very good, and provides
for various inputs in addition to audio. It looks like it might be
good for experimental control, reaction time, and that sort of thing.
The NeXT machine has a builtin DSP chip and builtin 8 bit
A/D. Metaresearch sells a 16-bit ADDAC board for a little under
US$1000. This board seems to be aimed more at the computer music
market so the sampling rates aren't typical ones for speech, but
you can get sampling rates in the right range, and I understand that
they are willing to customize.
AT&T has a combined ADDAC/DSP board for the VME-bus (SUN).
You can contact AT&T directly or deal with Entropic Processing
which sells the AT&T board together with their software.
ANALYSIS SOFTWARE
For PC class machines, I know of four sources.
Nirvonics, Inc. produces some very fine software, but it isn't
comprehensive. I know they have nice time series editors, play and
record software, and a program for controlling certain kinds of
experiments, but to my knowledge they don't have signal
processing programs (e.g. spectral analysis). Source is available.
Software Research Corporation sells an integrated
hardware/software package that is more comprehensive. I have seen
their ads but have no direct knowledge of their system. I don't
know if they will provide source.
Bill Henke (WLH) has a system called MITSYN that has evolved over
many years of speech work at MIT. It is an unusual system, consisting
of a command interpreter and a DSP block compiler. I have seen
hardcopy generated by it and read a description of it, but have not
actually seen it. It runs on PCs and on VAXen under VMS. I don't know
if he will provide source.
Ariel has got what they call a "SpeechStation", which is a
combination of speech analysis software and one of their boards.
I don't know if they provide source.
MACINTOSHen
GW Instruments has software in addition to their hardware.
When I saw the software, I was not impressed. It provided the basic
analyses, but had inferior facilities for making measurements and
logging them and was not very flexible. However, this was several
years ago and they may have improved it, although my understanding is
that they are mainly interested in hardware. When last I heard, they
would not provide source.
Info Signal sells a program called Signalyze, a comprehensive
display and signal processing package that has the virtue of costing
only US$250. They will send you a demo diskette for US$10. (However,
they will only sell manuals with their software, not separately.) I
believe that they will provide source for part but not all of their
system (the idea being to allow extension but not access to the core
system).
Signal Technologies' ILS, discussed below, is available for
the Macintosh.
ESPS/Waves, discussed below, is available for the Macintosh.
UNIX WORKSTATIONS
Signal Technologies' ILS has been a standard for quite
a while in the speech and signal processing communities. It
provides a whole slew of signal processing and display
facilities. Source (FORTRAN) is available. ILS comes out of the
VAX/VMS world but has become available for a variety of UNIX
machines and for the Macintosh.
A newcomer to this market is Entropic Processing's
ESPS/Waves package. ESPS is a comprehensive signal processing
package comparable in scope to ILS that relies on communication
via pipes and other UNIX facilities to be very extensible and
flexible. Waves is an editor/analysis system developed by Dave
Talkin at AT&T Bell Labs that has now been integrated with ESPS.
ESPS/Waves don't have to be used with any particular hardware,
but if you have a SUN you can get an ADDAC/DSP board from AT&T
that Waves was designed to work with. ESPS/Waves is designed
for UNIX systems with the X window system, but there is supposed
to be a Macintosh version available (though I don't see how one
can take full advantage of the design philosophy on a Mac).
Source (C) is definitely available but costs significant extra
money.
I think that ESPS is much nicer than ILS, in terms
of modularity, absence of unmotivated restrictions, and so
forth. To put it simply, ESPS comes out of a C/UNIX mentality,
ILS out of a FORTRAN/VMS mentality. Of course, some people may
prefer the latter.
Although I think that ILS is rather crufty and that ESPS is
clearly preferable, Signal Technology has a new product called
Nth-Power. I don't know much about it, but it is probably
worth checking out.
Nirvonics' software (discussed above under PCs) is also
available for SUN 3s. I don't know if it is available for SUN 4s.
Finally, with some trepidation now that this is going
out to the net rather than a few people, I will mention that I
myself have written a programmable time series editor that runs
on UNIX systems under X. It is sort of an EMACS for phonetics.
It can handle multiple time series and other functions of the
integers (e.g. spectral slices) but does not at present handle
spectrograms. It does not do things like spectral analysis or
audio i/o itself but relies on other programs to do this. It can
execute other programs as child processes, both locally and on
remote machines. This means that you need other programs to
perform these tasks, but that it is very portable (e.g. it knows
nothing about any particular audio i/o system) and easily
extensible and changeable (since you can run whatever other
programs you like). It is superior to other systems that I know
of in customizability (since it is truly programmable and
communicates easily with other programs) and in its facilities
for making measurements and logging them automatically, but is
not sufficient stand-alone (since it does not do
signal-processing). There are a few respects in which it could
be more robust, and at present, although you can associate mouse
button presses with any action at all, it does not make use of
menus and all of the graphical interface tools. I have
distributed it to a few other labs so far, with source, at
minimal cost, but would prefer to wait a while before
distributing it very widely. I will entertain requests from
people seriously interested who can put up with an editor that is
very powerful but is still a one-man show under development.
MY RECOMMENDATIONS
Taking into account the fact that there are some systems
that I don't know very much about, here is my view of the matter.
The PC-based systems that I have seen are not terribly
interesting, except possibly for Henke's software. The choice
is probably between a Macintosh and the various UNIX workstations. If it
is to be the Mac, I would go with the GW Instruments ADDAC board and
either Signalyze or ESPS/Waves for the software. In a general way
I would prefer ESPS/Waves to Signalyze, but I am not sure that
the benefits of the ESPS/Waves philosophy are realized on the Mac.
I also don't know to what extent the ESPS/Waves port to the Mac
is optimized for the Mac architecture, whereas Signalyze
is designed specifically for the Mac.
For a UNIX workstation, I would choose ESPS/Waves
together with my own editor.
The choice between the Macintosh and the UNIX systems
depends on what you want to do and what sort of users you have.
Naive users tend to prefer the Macintosh, and that may be the
right choice if you will have a lot of naive users (e.g. students
doing exercises), if the tasks fit the software well (so that
customization and communication with other programs is not important),
and if the need to do things in a purely interactive fashion will
not be inefficient.
I myself strongly prefer a UNIX environment, the most
important reason being that the typical Mac interface, while nice
for the novice user, is very inefficient and cumbersome for doing
real research, and that the lack of real multitasking and of
things like UNIX pipes impedes extensibility and modularity.
This is to some extent a matter of religion (Mac partisans
are requested to send flames to alt.religion.computers), but my own
choice would be the combination of ESPS/Waves (and my editor) on a
UNIX workstation.
Bill Poser
Financial Disclaimer:
I have no connection with any of the companies mentioned.
In 1982-1983 I worked for AT&T Bell Labs.
Ideological Disclaimer:
I love UNIX and dislike Macintoshen.
Addresses:
Ariel Corporation
433 River Road
Highland Park, NJ 08904
USA
Tel: 201-249-2900
Telex: 4997279
FAX: 201-249-2123
DSP BBS: 201-249-2124
AT&T
Jim Kingsbury
Allentown, PA
USA
215-770-2945
215-770-2978 (FAX)
Data Translation, Inc.
100 Locke Drive
Marlboro, MA 01752-1192
USA
508-481-3700
Entropic Processing, Inc.
Washington Research Laboratory
600 Pennsylvania Ave., S.E., Suite 202
Washington, D.C. 20003
(800-446-0028)
(202-547-1420)
(202-546-6648) (FAX)
esps@wrl.epi.com
...uunet!epiwrl!esps
GW Instruments
P.O. Box 2145
264 Monsignor O'Brien Highway No. 8
Cambridge, MA 02141
USA
617-625-4096
InfoSignal Inc.
231 Belair E.
ROSEMERE, Quebec
J7A 1A9
Canada
CompuServe 76357,1213
76357.1213@COMPUSERVE.COM
FAX: 1 - (514) 430-8522
Metaresearch, Inc.
516 SE Morrison
Suite M-1
Portland, OR 97214
USA
Nirvonics Inc.
P.O. Box 5062
Plainfield, NJ 07061
USA
Signal Technology, Inc.
5951 Encina Road
Goleta, CA 93117-2274
USA
800-235-5787
805-683-3771
Software Research Corporation
3939 Quadra Street
Victoria, B.C.
V8X 1J5
Canada
WLH (= William L. Henke)
133 Bright Road
Belmont, MA 02178
USA
617-484-2486