`
`llcuuz/q
`
`ISemihax
`
`ORIGINAL
`
`SEMiNAR
`
`E£,3&9
`02/18/187
`
`TRANSCRIBED IER’OMZ VIDEO RECORDING.
`
`fE‘ORzg
`
`IKENYON ‘& KENIYION!
`
`
`
`1500 ‘K STREET,
`
`1N .~W.r
`
`SUITE 70a
`
`WASHINGTON, DC
`
`200316
`
`15
`
`16
`
`17
`
`18I
`
`l9.
`
`VH5’250N 262533
`‘Aldersion Reporting Company
`800—‘FOR—DEPO
`I] 111 mm Street,IN;W., Suite 400‘
`
`thhington,‘ DC 20005
`
`SIGHTSOUND TECHNOLOGIES
`EXHIBIT 2328
`
`CBM2013-00023 (APPLE v. SIGHTSOUND)
`PAGE 000001
`
`
`
`Seminar
`
`P R O C E E D I N G S
`
`Speaker: Good afternoon. Welcome to 380.
`
`Just
`
`a quick reminder.
`
`As the book goes around,
`
`those of you
`
`who were enrolled in the class, please sign it. And you
`
`might pick up the missing slots that you have not signed
`
`for in the past, annotating them appropriately.
`
`Our speaker for next week will be Tom Blank.
`
`He's going to speak on incremental CAD techniques for
`
`systems design. But it turns out
`
`to be a euphemism for
`
`three things that he thinks are very interesting and it
`
`promises to be a fairly exciting talk.
`
`care of the remaining introductions and the rest of the
`
`For those of you who are keeping track,
`
`the Sun
`
`Microsystem talk scheduled for March 11th has been
`
`canceled and we will have a new speaker for you. There
`
`are actually two possible speakers, one of whom is as yet
`
`uncommitted. But if it can be worked out, it will cover a
`
`particularly interesting area of the law and software at
`
`the moment.
`
`Our speaker for today is David Schwartz and John
`
`Stoffner. They're from Compusonics, a startup company
`
`here in Palo Alto. They're going to speak on
`
`multiprocessor computers for digital, audio and video
`
`recording and editing.
`
`It's a masterful sound and light
`
`show,
`
`I think. And I'll introduce Dave and he can take
`
`Alderson Reporting Company
`1111 14th Street, N.W., Suite 400
`800-FOR-DEPO
`
`Washington, DC 20005
`
`PAGE 000002
`
`
`
`Seminar
`
`
`
` afternoon.
`
`
`
`Thank you, Dennis. We're
`Mr. David Schwartz:
`
`
`The talk —— we Changed the
`going straight to slides.
` title of the talk, of course, without telling Dennis.
`
`
`
`Stoffner is heading up the video project, but we felt that
`
`
`It's now inside the DSP—lOOO audio computer.
`
`John
`
`trying to do two talks in one session, both audio and
`
`
`
`back and get more into the video processors.
`
`This is the machine we're going to talk about
`
`today.
`
`It's the DSP—lOOO.
`
`It's a computer that is an
`
`audio device.
`
`It records. There's a cartridge in there.
`
`I may as well start by this -— John, could you help me
`
`out?
`
`I'm tied down.
`
`Just hand the cartridges out.
`
`I'd
`
`like these back.
`
`Just hand them to the audience.
`
`Just
`
`pass them around.
`
`Those are the optical disks that this machine
`
`records on.
`
`They are write once optical disks. They're
`
`double sided, 233 megabytes per side. They're made by
`
`several companies.
`
`The ones that you're handing around
`
`are made by Dysel Chemical Company in Japan. They're also
`
`made by Phillips and 3M and Plasma in England.
`
`Do I ask for slides? Or do they automatically
`
`
`video, would just get to be too much.
`So we'll cover the
`
` audio. And if you're still interested, we can be invited
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`change them?
`
`
`10
`
`11
`
`12
`
`1‘I’ 13
`
`14
`
`15
`
`16
`
`17
`
`18
`
`19
`
`20
`
`21
`
`22
`
`23
`
`24
`
`025
`
`Alderson Reporting Company
`1111 14th Street, N.W., Suite 400
`800-FOR-DEPO
`
`Washington, DC 20005
`
`PAGE 000003
`
`
`
`Senfinar
`
`Speaker: Just ask for them.
`
`Mr. David Schwartz: Slide.
`
`The DSP 1000 has a
`
`number of outstanding points.
`
`I'll summarize the big
`
`ones.
`
`The recordings are permanent.
`
`I
`
`think you're all
`
`familiar with tape recorders and what happens with tapes
`
`and cassettes.
`
`The quality of the recording is equivalent when
`
`making an original recording with the machine that is
`
`using microphones and recording a live performance, you're
`
`effectively making a digital master,
`
`like the compact disk
`
`master. Although the format is different, we do sample at
`
`44.1 kilohertz, 16 bit linear samples in stereo.
`
`So we think the quality's very high. Of course,
`
`it's random access.
`
`Since it's a computer, its editing
`
`software is built in. We'll talk about that.
`
`The recording time varies. There
`
`It interfaces to any computer really over the RS
`
`232 port.
`
`The first software that we're bringing out for
`
`interfacing is MS DOS software because we all know the
`
`world is according to IBM at this point.
`
`And the error rate of the disk is very, very
`
`low, especially compared to compact disks or CD ROM.
`
`The
`
`error correction board, made here in Valley, corrects to
`
`10(—12) beta errors, which makes it suitable for any type
`
`of data storage.
`
`Slide please.
`
`Alderson Reporting Company
`1111 14th Street, N.W., Suite 400
`800-FOR-DEPO
`
`Washington, DC 20005
`
`PAGE 000004
`
`
`
`Seminar
`
`are two key technologies in here.
`
`One is the high
`
`capacity storage medium and a computer that can use it in
`
`real time. And the other is digital signal processing
`
`which is used to optimize the storage on the disk.
`
`It's
`
`also used to create what you might say is the emulation of
`
`tape speed.
`
`Mr. David Schwartz: Oh, boy. We're in trouble
`
`Of course,
`
`there's no tape.
`
`So there's no tape
`
`speed.
`
`So we have what we call short for Compusonics,
`
`CSX. There are three settings that are user selectable
`
`from the front panel, CSX2, very high fidelity of stereo,
`
`equivalent to master recording, 30 minutes per side of the
`
`disk.
`
`CSX4 is about what you'd get from a very high
`
`quality professional tape recorder, an analog tape
`
`recorder.
`
`Say a Studder or an Atari Pro analog half inch
`
`or Ampex half inch deck.
`
`CSX8 is a baud limited signal,
`
`but very clean up to six kilohertz.
`
`It's mainly intended
`
`for very old records which don't have any high frequency
`
`material at all other than surface noise or speech, very
`
`high fidelity speech. Or, of course, AM radio which runs
`
`on that bandwidth.
`
`Of course, if you turn the disk over, you have
`
`double the recording time too. Slide.
`
`Speaker: We're averaging a minute 30 seconds on
`
`the slide.
`
`10
`
`ll
`
`12
`
`l3
`
`14
`
`15
`
`16
`
`17
`
`18
`
`19
`
`20
`
`21
`
`22
`
`23
`
`24
`
`25
`
`Alderson Reporting Company
`1111 14th Street, N.W., Suite 400
`800-FOR-DEPO
`
`Washington, DC 20005
`
`PAGE 000005
`
`
`
`Seminar
`
`on the slides. We have 49 slides.
`
`So we're going to move
`
`along. This is the back of the machine.
`
`So you can see
`
`how to hook it up.
`
`You take your tape deck, unplug it from your
`
`system, give it to your little brother or little sister,
`
`put this on the shelf and plug in your -— here are your
`
`inputs, RCA jacks, outputs,
`
`remote control, DD25, parallel
`
`port for digital dubbing. And, yes, we will offer a
`
`digital converter for compact disk output to copy sonex
`
`digital input.
`
`DCE to DTE switch for your R8232, wires
`
`two and three. Your RSZ32C port. And, of course, power.
`
`This is obvious.
`
`You just plug it in.
`
`Slide.
`
`The applications for consumers of
`
`preserving records and tapes that have some value, either
`
`as collector pieces or just sentimental value to you.
`
`Every time you play a record or a tape now, you know
`
`you're wearing it out.
`
`Now, if you really care about
`
`having music 20 or 30 or 50 years from now,
`
`this is the
`
`only medium I think that can be called archival for audio
`
`at this point.
`
`If you are
`
`If you're into creating your own music, you sing
`
`in a barber shop quartet or play in a rock 'n roll band,
`
`it's ideal for making your own digital disks. Custom
`
`dubs. That is party disks.
`
`I don't know why we didn't
`
`just say party disks.
`
`Home music editing.
`
`Alderson Reporting Company
`1111 14th Street, N.W., Suite 400
`800-FOR-DEPO
`
`Washington, DC 20005
`
`PAGE 000006
`
`
`
`
`
`Seminar
`
`creating program material from your own noodlings around
`
`on the piano or whatever, very few amateur musicians,
`
`including myself, are perfect. We make lots of mistakes.
`
`So what you need to do is edit out
`
`the mistakes and try
`
`and put together a credible performance that you can then
`
`play at your next party with your friends.
`
`So editing is
`
`a key application.
`
`Slide. With a machine that costs $7,000, you
`
`know,
`
`there are going to be a lot of professionals using
`
`it.
`
`So, we have live recording, music editing,
`
`preservation of master tapes.
`
`I
`
`think you've heard all
`
`the stories about how many good master tapes have just
`
`decayed in the vault or accidentally been demagnetized by
`
`somebody who just needed another role of tape and figured,
`
`well, nobody wants those old Jimmy Hendrix masters anyway.
`
`those out of business.
`
`Libraries need to archive audio.
`
`Those old wax
`
`cylinders from the turn of the century do not withstand
`
`multiple playing very well. They're a candidate for being
`
`converted to optical disk.
`
`Broadcasting and studio sound effects.
`
`Sound
`
`effects generally are played off of tape cartridges these
`
`days which are 40 year old technology. They're kind of
`
`eight track, endless loop cartridges. We hope to put
`
`Alderson Reporting Company
`1111 14th Street, N.W., Suite 400
`800-FOR—DEPO
`
`Washington, DC 20005
`
`PAGE 000007
`
`
`
`Seminar
`
`Automated random access playback. That is to
`
`say if you plug your machine into your PC and use our
`
`software, you have in effect a big jukebox. Or you can
`
`use the PC screen and a mouse or cursor keys to select
`
`what you want to play next, cue up what you want to play
`
`next, make a log list and play them in the order. Or
`
`generate a random number and play everything on the disk
`
`in some desired random fashion.
`
`Slide. Editing.
`
`The built—in editor. We tried
`
`to make it intuitive.
`
`Now, since very few people have
`
`ever done electronic editing, what would intuitive be?
`
`And we thought, well, we can visualize making marks in a
`
`linear file.
`
`Those of you who have edited tape know that
`
`you take a razor blade and you cut it and then you stick
`
`it back together.
`
`So we tried to keep that analogy in the editor.
`
`So we've given you the two flags. While the sound file is
`
`playing,
`
`in real time you simply push flag one, drop the
`
`first flag, and then flag two.
`
`There are two function
`
`keys. These make little markers.
`
`Then the software gives you the choice of either
`
`saving the part that's between the two flags or saving the
`
`outside parts which creates two new sound files, but are
`
`the parts from the beginning to flag one and then flag two
`
`10
`
`11
`
`12
`
`1". 13
`
`14
`
`15
`
`16
`
`17
`
`18
`
`19
`
`20
`
`21
`
`22
`
`23
`
`24
`
`
`
`025
`
`to the end of the sound file.
`
`Alderson Reporting Company
`1111 14th Street, N.W., Suite 400
`800-FOR-DEPO
`
`Washington, DC 20005
`
`PAGE 000008
`
`
`
`Seminar
`
`And there save is actually two separate sound
`
`files.
`
`I know the slide says save to sound file 11.
`
`They're actually saved as 11 and 12.
`
`The inside, of
`
`course,
`
`is just saved as one new sound file. Once you
`
`have these new sound files, you can make a list.
`
`DMA
`
`Slide.
`
`So you essentially cut up the audio that
`
`you're interested in, cutting out
`
`the part you don't like
`
`or saving the part you do like.
`
`Then you assemble them in
`
`any sequence you like.
`
`The real time disk operating and
`
`editing system which makes a nice pneumonic RTDOES. RTDOES
`
`is fast enough that on playback, any sound file anywhere
`
`on the disk —— let's say you have an hour of stereo
`
`recorded and you chop it up into three second or four
`
`second phrases, little musical phrases.
`
`You can reassemble them and play them back in
`
`real time as you wish at will. And the lists themselves
`
`may be saved. We call these edit lists.
`
`Up to 12,600
`
`edit lists can be written into the directory space of the
`
`disk itself.
`
`Slide. Here the whole system in a gross
`
`overview,
`
`the underwrite section, A to D, D to A,
`
`converters, 16 bit linears,
`
`a buffer that buffers the data
`
`for the signal processors. These are Texas Instruments,
`
`TMS3205, 320-105. Data then is passed into RAM.
`
`The
`
`68,000 is handling the data transfers at this point.
`
`Alderson Reporting Company
`1111 14th Street, N.W., Suite 400
`800-FOR-DEPO
`
`Washington, DC 20005
`
`PAGE 000009
`
`
`
`Seminar
`
`10
`
`through the SCSI port to the disk. Or if there's a serial
`
`port which connects to the PC,
`
`the parallel port for
`
`dubbing,
`
`the system ROMS that contained the software,
`
`presently 64K bytes, a 20—key pad of which we're only
`
`using five keys, and the LCD display.
`
`Slide. Here is the data path as it's shown in
`
`one of our development systems.
`
`Now,
`
`this is getting into
`
`a little bit of history of how did this come about and
`
`where did it come from? We started developing these in
`
`'83 and started with a multibus chassis computer to which
`
`we added custom boards.
`
`So here is the audio in and out
`
`section, A to D. There are the FIFOs.
`
`The local audio
`
`data bus is just a ribbon cable to the signal processors,
`
`parallel data.
`
`here to help the whole development.
`
`Here is the Intel multibus,
`
`the back plane.
`
`So
`
`the data comes in the A to D, gets processed by the signal
`
`processors.
`
`The CPU manages this transfer of data into
`
`RAM, back out of RAM to the DNA board, down through the
`
`SCSI bus to the disk controller - this also handles error
`
`correction — to the disk.
`
`It says floppy disk. Yes, we
`
`do make a companion product that uses floppy disks and
`
`optical disks.
`
`The development systems also use Winchester disk
`
`drives for obvious reasons. There's a UNIX system living
`
`Alderson Reporting Company
`1111 14th Street, N.W., Suite 400
`800-FOR-DEPO
`
`Washington, DC 20005
`
`PAGE 000010
`
`
`
`Seminar
`
`11
`
`Next slide. Here is the machine in front of you
`
`gutted. There's the little keypad and LCD display.
`
`The
`
`front panel laid on its face,
`
`the main board not populated
`
`in this picture,
`
`the tray that holds the analog power
`
`supply,
`
`linear supply and the switcher for the digital
`
`electronics.
`
`They are separate for reasons of noise.
`
`The
`
`optical disk drive and the error correction and controller
`
`board back here, SCSI board.
`
`then we'd have
`
`Slide. Okay. Let's get into the nitty gritty.
`
`Let's start up here at the RCA jack. Here are the two
`
`pairs in and out. These are the filters, anti-imaging,
`
`anti—(inaudible) filters. There are 13 pair elliptical
`
`filters that are analog.
`
`The important thing about the
`
`filters. These are made in Sweden of all places. Because
`
`it turned out
`
`they were the best filters we could find.
`
`And what we were looking for, of course,
`
`is very low
`
`distortion at 90 dB down. And very important at 20
`
`kilohertz correction of any phase delay. Any group delays
`
`introduced by the filters have to be corrected.
`
`Some people say you can't hear this at 20
`
`kilohertz. Our contention is that you can hear it and you
`
`better have good filters to take care of it.
`
`The A to D converters are Bill Brown,
`
`PCM 755.
`
`Notice there are two entirely mono channels. We are not
`
`multiplexing the data.
`
`If we multiplex it,
`
`Alderson Reporting Company
`1111 14th Street, N.W., Suite 400
`800-FOR-DEPO
`
`Washington, DC 20005
`
`PAGE 00001 1
`
`
`
`Seminar
`
`the classic Sony problem of a one sample delay to deal
`
`with. We
`
`thought we'd rather not deal with the delay and
`
`handle the two mono channels.
`
`These are all FIFO chips,
`
`hardware chips, mostex.
`
`There's the signal processors.
`
`Sorry, we don't
`
`have two identical packages. There's the ceramic package
`
`and here's the plastic package for the 320.
`
`In this area,
`
`there are two banks of 8K bytes of staff static RAM.
`
`So
`
`the data after it's converted is held in these FIFOs and
`
`then grabbed into fast static RAM and processed about
`
`l/lOOth of a second, between 1/100th of a second and
`
`2/lOOth of a second at a time by the 3205.
`
`It's block
`
`processing.
`
`John Stoppler will be talking more about that
`
`in a moment.
`
`I don't
`
`Once the data is processed, it's sitting here in
`
`its static RAM.
`
`68,000 sees this RAM in its address space
`
`and moves the data out word—by—word because the 68,000 has
`
`no block movement instructions.
`
`Thank you, very much.
`
`Move it out a word at a time into the system RAM. There's
`
`half a megabyte of system RAM right here.
`
`From system RAM
`
`through the -— under DMA control, DMA chip, out
`
`through
`
`the SCSI chip. This is the NCR SCSI port in a chip. Very
`
`nice little thing. And here's the connector up to the
`
`disk drives, controller error correction board which lives
`
`upstairs on that plate that you saw. E-prompts.
`
`Alderson Reporting Company
`111 1 14th Street, N.W., Suite 400
`800-FOR-DEPO
`
`Washington, DC 20005
`
`PAGE 000012
`
`
`
`Seminar
`
`think I'll go through and point out every chip on here.
`
`We'll be here all day.
`
`Next slide. We'll have questions at the end of
`
`this.
`
`If you want to come back and start digging into
`
`this or that, we'll be glad to put it back up and dig into
`
`it.
`
`The software structure in general. Here's
`
`RTDOES. Real Time Disk Operating and Editing System is
`
`the overall structure that handles all of the commands of
`
`the user.
`
`or the AT&T accunet which is AT&T's high speed digital
`
`ACL stands for Audio Control Language. This is
`
`an internal language that RTDOES knows about that we wrote
`
`that have high level commands, much like a graphic
`
`language for controlling graphics in a graphic system.
`
`The audio command, audio control language, has
`
`an R8232 protocol set that we publish so that you can talk
`
`to the machine via a dumb terminal or any PC. We show an
`
`IBM PC right here.
`
`ACL also handles the remote switches for the
`
`remote control and, of course,
`
`the front panel commands.
`
`Disk operations, signal processing, which is
`
`audio I0 is actually handled under the signal processors
`
`running their own code. And digital IO which talks either
`
`to another DSP 1000 or another piece of digital audio gear
`
`Alderson Reporting Company
`111 1 14th Street, N.W., Suite 400
`800-FOR-DEPO
`
`Washington, DC 20005
`
`PAGE 000013
`
`
`
`Seminar
`
`transmission system. That would be a good topic for
`
`someone else to talk about.
`
`Next slide. Okay. Let's talk about the disks
`
`for a minute. Here's the disk drive. There's a compact
`
`disk which we all know about. Here's the optical disk
`
`you've been handing around. Here's the shell that it
`
`lives in. And you can see it has a stainless steel hub.
`
`The disks are believed to be very durable.
`
`If they do get
`
`damaged,
`
`they can be repolished. They're lexon.
`
`Slide.
`
`The cross section is slightly in error.
`
`We got this from the manufacturer of the disk drive,
`
`Optotech. We do not have a glass substrate. There is no
`
`laquer. Yes,
`
`there is tellurium. And, yes,
`
`there is air
`
`in here.
`
`I think this is pretty self-explanatory. What
`
`you do is you heat up the terrarium enough so that it
`
`curls up.
`
`It sort of shrivels the way plastic does when
`
`you touch a hot match to it.
`
`I'll show you that in the
`
`writes to verify that it did burn a good crater. And this
`
`next slide, please.
`
`Here is the right signal. This is power versus
`
`time.
`
`So if you give the disk drive this kind of signal,
`
`it burns this kind of hole in the terrarium. What
`
`actually happens is the terrarium kind of crawls back and
`
`forms these like a volcanic crater kind of thing.
`
`It
`
`piles up around the edges.
`
`The drive reads while it
`
`Alderson Reporting Company
`1111 14th Street, N.W., Suite 400
`800-FOR-DEPO
`
`Washington, DC 20005
`
`PAGE 000014
`
`
`
`Seminar
`
`1 5
`
`is the signal you read back from the head. Again, power
`
`and time.
`
`So you get a negative going signal for the fact
`
`that there is no reflection, very little reflection,
`
`from
`
`the polycarbonate.
`
`MIT working on the signal processing before I met him,
`
`Slide. These are the actual holes.
`
`They don't
`
`look much like holes here.
`
`I'm not sure what
`
`they look
`
`like, white chocolate covered pretzels.
`
`You don't see the
`
`difference in background reflectivity between the shiny
`
`tellurium here and the polycarbonate here simply because
`
`this SEM was made looking for the high edge of the crater.
`
`So here's the edge of the crater.
`
`These are about a
`
`micrometer.
`
`Slide. That's your music you were looking at.
`
`Well,
`
`I mentioned that there are two parts of the machine,
`
`two key components,
`
`the hardware and the computer
`
`essentially which I've been working with very closely.
`
`And then there's the software,
`
`the signal processing.
`
`The
`
`co—founder of the company who's most closely related to
`
`developing the software, writing the software,
`
`for signal
`
`processing which is really one of the two key elements
`
`here is John Stoppler,
`
`the fellow who is presently heading
`
`the video project.
`
`I
`
`invited John here to talk to you
`
`about the signal processing.
`
`So,
`
`John, come on up.
`
`I might mention that John started this work at
`
`Alderson Reporting Company
`1111 14th Street, N.W., Suite 400
`800-FOR-DEPO
`
`Washington, DC 20005
`
`PAGE 000015
`
`
`
`Seminar
`
`probably going back as far as 1981.
`
`Mr. John Stoppler: The disk drive alone which
`
`David just showed you would be capable of storing about 15
`
`minutes of stereo material without any signal processing.
`
`That is if you could get the data onto the disk drive that
`
`fast which you can't. And it turns out that the only
`
`thing which makes this device work and the only thing
`
`which is likely to make any devices like this work is in
`
`the next several years is digital signal processing for
`
`the of reducing the amount of digital information required
`
`to store what is otherwise a vast amount of audio data.
`
`Basically,
`
`the goals are pretty clear. We want
`
`to minimize the storage memory requirements in the
`
`recorder. We'd like to minimize the disk or the data
`
`transfer rate which also relates to the transmission
`
`bandwidth. And while we're doing all that, we want to
`
`retain some semblance of high audio fidelity, if possible
`
`the highest audio fidelity. And at the same time, we also
`
`want
`
`to retain signal acceptability.
`
`In other words, we
`
`don't want to make such a muddy bunch of bits after we're
`
`done that we can't do something simple like start in the
`
`middle of a song, for example, without having to go all
`
`the way to the start and reconstruct all the history.
`
`Next slide. There's a few observations which
`
`16
`
`
`
`10
`
`11
`
`12
`
`‘I’ 13
`
`14
`
`15
`
`16
`
`17
`
`18
`
`19
`
`20
`
`21
`
`22
`
`23
`
`24
`
`.25
`
`one can make about music. First of all, as I'm sure
`
`Alderson Reporting Company
`1111 14th Street, N.W., Suite 400
`800-FOR-DEPO
`
`Washington, DC 20005
`
`PAGE 000016
`
`
`
`Seminar
`
`you‘re probably aware, music has a very broadly varying
`
`spectrum which varies with time. Spectral concentrations
`
`of energy show up everywhere in the spectrum. Oftentimes,
`
`they alternate between broad band and very narrow band
`
`behavior. And there's also very wide ranging dynamic
`
`range.
`
`Sometimes going up greater than 90 DB, sometimes
`
`staying within a very narrow range.
`
`Next slide. This is a long time avid spectrum
`
`of some orchestral music.
`
`Judging from this slide or from
`
`this data which was taken, one would think, well, why do
`
`you even really need to record much information above five
`
`kilohertz? That's the frequency access on the bottom.
`
`Those five kilohertz in the long-term average spectrum,
`
`it's about 30 DB down.
`
`Well,
`
`the fact is that for some finite
`
`There's something special about this particular
`
`percentage of time, you do get energy up there which the
`
`earth latches onto and finds very important. But a lot of
`
`the time,
`
`the energy is very narrow band.
`
`Next slide. This is a clot which was made on
`
`one of our audio computers at Compusonics of a time and
`
`amplitude spectrum of some violin strokes.
`
`The amplitude
`
`runs up into the page or up into the screen.
`
`The
`
`frequency is across the bottom.
`
`And then amplitude and
`
`the log rhythmic domain.
`
`Alderson Reporting Company
`1111 14th Street, N.W., Suite 400
`800-FOR-DEPO
`
`Washington, DC 20005
`
`PAGE 000017
`
`
`
`Seminar
`
`picture. And that is that it was made with a process
`
`which simulates as closely as possible the auditory sense
`
`of hearing,
`
`the musical sensitivity of hearing. And
`
`therefore, each analysis frequency is actually following
`
`what's known as the critical bandwidth of hearing. And
`
`we'll see that in the next slide.
`
`Next slide please.
`
`The one point
`
`I wanted to
`
`make about that last slide —— back up one slide for a
`
`second. Thanks.
`
`Is that you can see large percentages of
`
`the data,
`
`the musical data, has no energy in it, both in
`
`time and in frequency.
`
`Every now and then you get an
`
`attack which extends across the entire frequency range.
`
`But
`
`then for a large percentage of the data,
`
`there's
`
`nothing there.
`
`These curves show that.
`
`Okay. Next slide. What one can do if one were
`
`able to run an algorithm like this in real time -— which
`
`we don't currently —— we've only simulated this in
`
`non—real
`
`time -— is you could take advantage of the fact
`
`that the ear tends to hear things in what's known as
`
`critical bands.
`
`If you have a source of sound near a
`
`particular frequency region,
`
`in this case near 400 hertz,
`
`then it is very difficult if you present a second sound
`
`such as a sine sound or any other narrow band source, near
`
`that primary source, it's very hard to pick that up unless
`
`it's played as loud or louder.
`
`Alderson Reporting Company
`111 1 14th Street, N.W., Suite 400
`800-FOR-DEPO
`
`Washington, DC 20005
`
`PAGE 000018
`
`
`
`Senfinm
`
`Next slide.
`
`The pendency or critical band width
`
`which is the distance from which you'd have to go from the
`
`primary source before it becomes audible to a certain
`
`extent is known as the critical bandwidth. And this is a
`
`plot of what
`
`the critical bandwidth relationship looks
`
`like in frequency versus bandwidth -- bandwidth versus
`
`frequency.
`
`And it takes on somewhat of a third octave
`
`characteristic which is to say that the critical bands
`
`themselves get wider and wider as frequency increases.
`
`Thereby,
`
`the ear becomes more sensitive to wider frequency
`
`expanses as you go higher and higher frequency.
`
`Next slide.
`
`Now,
`
`if we could take advantage of
`
`all of that information and pack that into a signal
`
`processor and run it in real time, we would have a fairly
`
`complex algorithm which would take 40 or 50 million
`
`floating point operations per second or fixed point if you
`
`took care to code it properly to run in real time for the
`
`analysis and the synthesis.- I'm not going to go into all
`
`the details of how that works, but that is the ultimate in
`
`data reduction which is where we started looking at the
`
`problem.
`
`Next slide. And this is the basic approach you
`
`would do.
`
`You would run your input
`
`through an auditory
`
`10
`
`11
`
`12
`
`l‘l’ 13
`
`14
`
`15
`
`16
`
`17
`
`18
`
`19
`
`20
`
`21
`
`22
`
`23
`
`24
`
`.25
`
`filter bank.
`
`Then based on what you know about the
`
`19
`
`
`
`Alderson Reporting Company
`1111 14th Street, N.W., Suite 400
`800-FOR-DEPO
`
`Washington, DC 20005
`
`PAGE 000019
`
`
`
`Seminar
`
`20
`
`auditory system and based on the fact that you're
`
`splitting out the data as in that plot that we saw
`
`earlier, you would take out any particular features that
`
`you would find and pull out the silences and dead spots
`
`and then go ahead and code all that information.
`
`Next slide.
`
`Now,
`
`the sorts of things we
`
`actually did in practice and because we had to come down
`
`to the real world and run in real time in order to make a
`
`And another pathway is to do that and also build a
`
`product which we can get on the market
`
`in a reasonable
`
`amount of time is do some kind of approximation to that.
`
`One of these algorithms which is our -- which
`
`we're currently using in our lowest data rate mode is a
`
`sampling rate and 80 OCM algorithm which simulates to a
`
`certain extent those observations of musical signals and
`
`hearing characteristics.
`
`You have an input signal and you have four
`
`different possible pathways of analysis.
`
`One of those is
`
`to do nothing which is the upper pathway. One of those is
`
`to model
`
`the signal with the short-term recursor filter
`
`model. Another one is to decimate the signal which is to
`
`say cut the sample rate in half,
`
`thereby making the
`
`assumption that while there's nothing in the upper half of
`
`the spectrum as we saw in that long-term average spectrum
`
`and in those plots. Which is true for much of the time.
`
`10
`
`ll
`
`12
`
`‘I’ l3
`
`14
`
`15
`
`16
`
`17
`
`18
`
`19
`
`20
`
`21
`
`22
`
`23
`
`24
`
`.25
`
`Alderson Reporting Company
`1111 14th Street, N.W., Suite 400
`800-FOR-DEPO
`
`Washington, DC 20005
`
`PAGE 000020
`
`
`
`Senfinm
`
`21
`
`short—term model.
`
`When you go backwards through the whole process,
`
`rebuild them all, and then say, okay.
`
`Now I've rebuilt
`
`those four different possible pathways. Which one of them
`
`subject to some psychoacoustic criterion is the best one
`
`to use? And you choose the one.
`
`Well, at this point in time,
`
`there's not a lot
`
`of energy in the higher octave. Let's pick one of the
`
`lower branches. This also is a very computational
`
`intensive way of going about things.
`
`It begins to come
`
`into the realm of reality in terms of what we can do with
`
`a few TMS 320 chips. There's also some very expensive
`
`signal processors available now which could do this that
`
`would make this product even more expensive. But we are
`
`using some versions of this algorithm in the current
`
`to another one
`
`product.
`
`Next slide. This is a plot of the error signal
`
`attained with each of the four algorithms running in
`
`parallel. Every now and then, one of the algorithms such
`
`as this one gives you a very large error. What happens,
`
`if you're following along on this algorithm,
`
`this happens
`
`to be the full band ADPCM. When it blows up like this,
`
`you know that you're getting noise. And so all you do is
`
`you branch from that particular algorithm -— which all
`
`four are running in parallel remember —-
`
`Alderson Reporting Company
`1111 14th Street, N.W., Suite 400
`800—FOR-DEPO
`
`Washington, DC 20005
`
`
`
`PAGE 000021
`
`10
`
`ll
`
`12
`
`‘I’ 13
`
`14
`
`15
`
`16
`
`17
`
`18
`
`19
`
`20
`
`21
`
`22
`
`23
`
`24
`
`.25
`
`
`
`Seminar
`
`which is giving you a lower noise level.
`
`Next slide.
`
`Now,
`
`the signal processing
`
`algorithm which we use in our highest fidelity mode is
`
`simpler yet. Because that comes down to being able to
`
`essentially run the algorithm at the full data rate and
`
`retain the highest possible quality.
`
`And in fact,
`
`that
`
`algorithm has the attribute that it's completely lossless.
`
`What that means is everything I've shown you up until now,
`
`when you had floating point or operations associated with
`
`it, which incur some kind of loss or bit loss, when you
`
`reconstruct the data if you compare the output with the
`
`input,
`
`they're not identical.
`
`This algorithm which we run in our highest
`
`fidelity situations has the attribute that while it
`
`thereby the data required the most. And then pack those
`
`reduces data, what you record is exactly what you get back
`
`when you play. And I don't think I want to go through
`
`each of the steps now. We can talk a little bit about it
`
`later in detail if you have more questions.
`
`The basic idea of the algorithm is to go back
`
`again to the idea of the spectral dynamic range and
`
`concentrations of spectral energy and run filters, several
`
`different filters, over the music, again in parallel,
`
`just
`
`like the ADPCM algorithm, several of them in parallel and
`
`then choose the one which reduces the dynamic range and
`
`Alderson Reporting Company
`1111 14th Street, N.W., Suite 400
`800-FOR-DEPO
`
`Washington, DC 20005
`
`PAGE 000022
`
`
`
`Seminar
`
`into the minimum word sizes.
`
`Next slide.
`
`The synthesis algorithm just does
`
`the opposite.
`
`It runs the inverse filter. And if you
`
`choose the filters right, namely if you choose them to be
`
`filters with integer coefficients, you can work out the
`
`math and the program in such a way that the entire
`
`operation is 100 percent reversible.
`
`Next slide. Okay.
`
`The background,
`
`the
`
`kilohertz.
`
`research, which was done on some of these algorithms was
`
`conducted on our own DSP 2002 audio computer. We'll see
`
`picture of that. We use a 16 bit A to D and D to A
`
`conversion. This was while we were in the process of
`
`developing the system.
`
`So we couldn't do the research
`
`itself on the system. We had real time signal processing
`
`using the Texas Instruments TMS 3203. And we recorded
`
`everything onto magnetic disk. And we could also do a lot
`
`of our research in non—real
`
`time using a sort of a batch
`
`mode processing.
`
`Okay. Next. Okay.
`
`So we tested lots of
`
`different music, classical,
`
`rock 'n roll,
`
`jazz. There's
`
`two examples here.
`
`The first one is a classical example.
`
`The second one is a rock 'n roll example by the Police.
`
`We tried all kinds of different sampling rates.
`
`I believe
`
`these were done at either 50 or 44.1.
`
`I
`
`think it was 50
`
`Alderson Reporting Company
`111 1 14th Street, N.W., Suite 400
`800-FOR-DEPO
`
`Washington, DC 20005
`
`PAGE 000023
`
`
`
`
`
`Seminar
`
`24
`
`Next slide. And this gives you some idea —— is
`
`it possible to raise that a little bit? Well, you can see
`
`it.
`
`The upper curve here is the classical music example.
`
`This is the distribution of word sizes required for the
`
`entire piece.
`
`The average word size —— well,
`
`the highest
`
`word size, it's very hard to see there, but there's a tiny
`
`histogram there.
`
`It occupied the entire 16 bits.
`
`The
`
`center is pretty much at 12 bits.
`
`And that's the DSP 2002 audio
`
`After proce