Express Mail mailing label no.EL212283154US
`Date of Deposit: June 8, 2000
`I hereby certify that this paper or fee is being deposited with the United States Postal Service "Express Mail Post
`Office to Addressee" service under 37 CFR 1.10 on the date indicated above and is addressed to the Assistant
`Cor missio er for PktentsrlaQX PROVISIONAL APPLICATION, Washington, D. C. 20231.
`
`„."
`
`_ 0 0
`
`Anizov
`
`By:
`
`arrie Martin
`
`Attorney Docket No. AGLE0001PR
`
`IN THE U.S. PATENT AND TRADEMARK OFFICE
`Provisional Application Cover Sheet
`
`Assistant Commissioner for Patents
`BOX PROVISIONAL APPLICATION
`Washington, D.C. 20231
`
`Sir:
`
`1-1
`
`This is a request for filing a PROVISIONAL APPLICATION FOR PATENT under 37 CFR 1.53(b)(2).
`
`INVENTOR(s)/APPLICANT(s)
`
`Middle Initial Residence (City and Either State or Foreign Country
`First Name
`Last Name
`San Carlos, California
`Ted
`Calderone
`M.
`Woodside, California
`Paul
`Cook
`Palo Alto, California
`J.
`Mark
`Foster
`Additional inventors are being named on separately numbered sheets attached hereto.
`
`Title of the Invention
`
`METHOD AND APARATUS FOR CENTRALIZED VOICE-DRIVEN NATURAL
`LANGUAGE PROCESSING IN MULTI-MEDIA & HIGH BAND WIDTH
`APPLICATIONS
`
`Correspondence Address
`
`Michael A. Glenn
`3475 Edison Way, Ste. L
`Menlo Park, CA 94025
`
`Telephone No. (650) 474-8400
`
`Enclosed Application Parts (check all that apply)
`
`(X) Specification
`Number of Pages
`and 3 Drawing(s)
`
`( ) Other (specify)
`
`24
`
`( X) Small Entity Statement -( specify - IND or BUS)
`
`Filing Fee and Method of Payment
`
`X $75.00 for Small Entity
`
`$150 for Large Entity
`
`The Commissioner is authorized to charge the filing fee of $75.00 and any additional fees or credit any
`overpayment to Deposit Account No. 07-1445 (Order No. AGLE0001PR). A copy is enclosed for this
`purpose.
`
`spectfu y Submitted,
`
`EA E. JENNINGS,
`Reg. No. 44,804
`
`Comcast - Exhibit 1003, page 1
`
`

`

`Attorney Docket No. /4 GLC000 I F.
`
`Applicant/Patentee: AGILETV CORPORATION
`Serial or Patent No.
`Herewith
`Filed or Issued:
`Title: METHOD & APARATUS FOR CENTRALIZED VOICE-DRIVEN NATURAL LANGUAGE
`PROCESSING IN MULTI-MEDIA & HIGH BAND WIDTH APPLICATIONS
`
`Atty Docket No.
`
`AGLE0001PR
`
`VERIFIED STA I F,MENT (DECLARATION) CLAIMING SMALL ENTITY STATUS
`37 CFR 1.9(f) and 1.27(c)--SMALL BUSINESS CONCERN
`
`I hereby declare that I am
`[ the owner of the small business concern identified below:
`[ X ] an official empowered to act on behalf of the small business concern identified below:
`
`AGILETV CORPORATION
`ONAME OF CONCERN:
`ii!ADDRESS: 333 Ravenswood Ave.. Bldg. 202, Menlo Park, CA 94025
`
`1.7I hereby declare that the above identified small business concern qualifies as a small business concern as defined in 13
`'.:!CFR 121.3-18, and reproduced in 37 CFR 1.9(d), for purposes of paying reduced fees under 41(a) and (b) of Title
`''I‘735, U.S. Code, in that the number of employees of the concern, including those of its affiliates, does not exceed 500
`:persons. For purposes of this statement, (1) the number of employees of the business concern is the average over the
`'—'previous fiscal year of the concern of the persons employed on a full-time, part-time or temporary basis during each of
`the pay periods of the fiscal year, and (2) concerns are affiliates of each other when either, directly or indirectly, one
`.,concern controls or has the power to control the other, or a third party or parties controls or has the power to control
`
`DI hereby declare that rights under contract or law have been conveyed to and remain with the small business
`invention entitled: METHOD & APARATUS FOR
`the
`D concern identified above with regard to
`F:10ENTRALIZED VOICE-DRIVEN NATURAL LANGUAGE PROCESSING IN MULTI-MEDIA & HIGH
`, by inventor(s) Ted Calderone, Paul M.
`BAND WIDTH APPLICATIONS
`Cook, Mark J. Foster , described in
`
`[ X ] the specification filed herewith.
`[ ] application Serial No.
`[ ] patent #
`
`filed
`issued
`
`
`
`
`
`If the rights held by the above-identified small business concern are not exclusive, each individual, concern or
`organization having rights to the invention is listed below* and no rights to the invention are held by any person, other
`than the inventor, who could not qualify as a small business concern under 37 CFR 1.9(d) or by any concern which
`would not qualify as a small business concern under 37 CFR 1.9(d) or a nonprofit organization under 37 CFR 1.9(e).
`
`*Note: separate verified statements are required from each named person, concern or organization having rights to the
`invention averring to their status as small entities. (37 CFR 1.27)
`
`Name:
`Address:
`
`[ ] individual
`
`[ small business concern
`
`
`
`[ ] nonprofit organization
`
`Comcast - Exhibit 1003, page 2
`
`

`

`Attorney Docket No. A & L Ca 0 ? ( s
`
`I acknowledge the duty to file, in this application or patent, notification of any change in status resulting in loss of
`entitlement to small entity status prior to paying, or at the time of paying, the earliest of the issue fee or any
`maintenance fee due after the date on which status as a small entity is no longer appropriate. (37 CFR 1.28(b)).
`
`I hereby declare that all statements made herein of my own knowledge are true and that all statements made on
`information and belief are believed to be true; and further, that these statements were made with the knowledge that
`willful false statements and the like so made are punishable by fine or imprisonment, or both, under 1001 of Title 18
`of the U.S. Code, and that such willful false statements may jeopardize the validity of the application, any patent
`issuing thereon, or any patent to which this verified statement is directed.
`
`NAME OF PERSON SIGNING:
`James E. Jervis
`TITLE IN ORGANIZATION:
`Vice President, Intellectual Property
`ADDRESS OF PERSO SIGNING: 333 Rav swood Ave., Bldg. 202 Menlo Park, Ca 94025
`
`• SIGNATURE
`
`DA
`
`Comcast - Exhibit 1003, page 3
`
`

`

`Attorney Docket: AGLE0001PR
`
`AgileTv System
`
`Current Practice
`
`Currently voice operated functions using the latest voice recognition technologies
`have been limited to only a hand full of applications such as toys, appliances,
`some computers, voice dictation, cellular phones and voice control of one's
`home. Most of these applications use voice recognition technology that runs on
`a computer or voice recognition chip technology. These voice recognition
`systems typically offer only a limited number of commands and the recognition
`efficiency could be considered only fair and often requires voice training. There
`is however, another class of voice recognition technology called "Natural
`Language" which requires state of the art processing software and hundreds of
`megabytes of RAM to support. "Natural Language" voice recognition is currently
`being used in high end systems such as billing systems for the utility companies
`and the New York stock exchange because of it's ability to recognize spoken
`words from any voice. Some natural language systems claim to be totally user
`independent and are also capable of recognizing speech in several different
`languages. The system described in this disclosure uses a "Natural Language"
`voice recognition engine for both speech recognition and voice identification.
`
`With the exception of dictation systems that use voice recognition and show text
`
`as recognized on the monitor screen, all the systems mentioned above do not
`provide immediate feedback to voice input.
`
`In cable systems, several downstream data channels to receive channel and
`synchronization information are typically transmitted in a band of frequencies
`that, in the past, were reserved for re-broadcasting FM channels over cable.
`Currently most cable systems reserve some of the 88 to 108 MHz FM spectrum
`for set-top data transmission leaving the unused portion of that spectrum for
`
`AG LE0001P R_v 2
`
`Page 1 of 24
`
`Comcast - Exhibit 1003, page 4
`
`

`

`Attorney Docket: AGLE0001PR
`
`barker channels or additional video channels. The Open Cable Standard
`requires that the 70 to 130 MHz band be available for what's called Out-of-Band
`or (00B) or Downstream transmission.
`
`Most cable systems of today use the popular Hybrid Fiber Coax type architecture
`so that the downstream video signals, digital or analog, are sent to "hubs" or
`"nodes" via fiber optic cable. At the receiving side of the node, the optical signal
`from the fiber gets converted to an electrical signal containing all the analog and
`digital video RF carriers. This signal, in turn, is amplified and distributed via
`coaxial cable to all the subscribers in the node with a typical node consisting of
`anywhere from 500 to 1000 subscribers. Also, the 5 to 40 MHz upstream signal
`from each subscriber in the node is collected, combined and then sent to the
`headend via the same fiber used for the downstream video carriers or a separate
`fiber is used.
`
`Summary Introduction
`
`Certain embodiments include a multi-user control system for audio visual devices
`incorporating a voice recognition system that is centrally located, Certain further
`embodiments include centrally locating the voice recognition system in or near
`Cable Television (CATV) headend. Certain other further embodiments include
`centrally locating the voice recognition system in or near a server farm. Certain
`other further embodiments include centrally locating the voice recognition system
`in or near a web-site. Certain other further embodiments include centrally
`locating the voice recognition system in or near a gateway.
`
`Certain embodiments are capable of recognizing the vocal commands from a
`cable subscriber and then enacting upon those commands to control the delivery
`of entertainment and information services such as Video On Demand, Pay Per
`View, Channel control and the Internet. This system is unique in that the voice
`
`AG LE0001PR_v 2
`
`Page 2 of 24
`
`Comcast - Exhibit 1003, page 5
`
`

`

`Attorney Docket: AGLE0001PR
`
`command which is originated in the home of the subscriber is then sent upstream
`via a 5 to 40 MHz return path in the cable system to a central voice recognition
`The voice recognition and identification engine
`and processing engine.
`described in this disclosure is capable of processing thousands of voice
`commands simultaneously and therefore can offer a low latency entertainment
`and information experience to the Subscriber.
`
`Functional Description (what it does)
`
`In certain embodiments, overall media control system consists of several
`functional blocks with the first function being that of inputting a subscriber's voice
`into the system.
`
`Figure 1 depicts a remote control unit 1000 coupled with set-top apparatus 1100
`communicating via a two-stage wireline communications system containing a first
`wireline physical transport 1200 distributor node 1300 and a high speed physical
`transport 1400, possessing various delivery points 1500 and entry points 1510-
`1518 to a tightly coupled server farm 2000 with one or more gateways 2100, one
`or more tightly coupled server arrays 2200, in accordance with certain
`embodiments.
`
`Certain embodiments include a remote control unit 1000 fitted with microphone.
`Certain further embodiments include a remote control unit 1000 fitted with a
`special noise canceling microphone. Certain other further embodiments include
`a remote control unit 1000 fitted with microphone and a push-to-talk button.
`Certain further embodiments include a remote control unit 1000 fitted with a
`special noise canceling microphone and a push-to-talk button.
`
`The purpose of the microphone in the remote is to relay the subscriber's voice
`commands to the central voice recognition engine. The purpose of the push-to-
`
`AGLE0001PR_v_2
`
`Page 3 of 24
`
`Comcast - Exhibit 1003, page 6
`
`

`

`Attorney Docket: AGLE0001PR
`
`talk button is to begin the process of voice recognition by informing the system
`
`that the Subscriber is about to speak and also to provide immediate address
`information.
`
`In certain embodiments, voice commands from the subscriber are then
`preprocessed, that is the analog signals picked up from the microphone are
`converted to digital signals where they undergo additional processing before
`being transmitted to the voice recognition and identification engine located in the
`cable headend or other centralized location.
`
`The preprocessing function can also take place in the remote control 1000 itself
`
`before being transmitted to the set-top box 1100 or set-top appliance 1100, in
`certain embodiments.
`
`The voice signal from the remote 1000 is a digitally modulated RF signal whose
`properties comply with part 15 of the FCC rules in certain embodiments. The
`set-top box 1100 or set-top appliance 1100 receives the voice signal from the
`remote 1000 and performs the preprocessing function mentioned above.
`
`It 1100 is also used to transmit voice and sub'scriber address data to the
`centralized location or headend for voice recognition and identification. The RF
`signal from the remote 1000 is received by the set-top appliance 1100 and then
`re-modulated for upstream transmission 1200 on the 5 to 40 MHz cable return
`path. If a commercial set-top box 1100 is used to transmit the upstream voice
`data then the upstream channel allocation and transmission protocol are then
`controlled by the bi-directional communication system which is resident in the
`set-top box.
`
`In certain alternative embodiments, a commercial set-top box 1100 is not being
`used to transmit the digitized voice data upstream, the set-top appliance 1100 is
`
`AGLE0001PR_v 2
`
`Page 4 of 24
`
`Comcast - Exhibit 1003, page 7
`
`

`

`Attorney Docket: AGLE0001PR
`
`the upstream channel allocation and
`receiving
`for
`responsible
`then
`synchronization information. The data receiver in the set-top appliance 1100 is
`frequency agile, that is, it can be tuned to any one of several downstream data
`channels to receive channel and synchronization information.
`
`The system described specifically uses the subscriber's address information as a
`means by which the centrally located Agile Voice Processor can fetch a particular
`subscriber's parameter file. The parameter file contains voice training parameter
`data, voice identification parameters and user profiles for each member of the
`family at that address. This file can also contain parental control information and
`other specifics to that particular household such as language preferences or
`movie preferences or even internet preferences.
`
`The Addressed Subscriber Parameter File (ASPF) is what gives the system an
`identification and voice recognition.
`extremely high probability of user
`Addressing is an important feature when considering secure transactions such as
`banking because the speech recognition and identification system has to only
`identify an average of 4 parameter files for any one physical address, which of
`course, results in a very high probability of recognizing a specific speaker's voice.
`
`Financial level transactional security (Voice Banking) can be realized with this
`system and with the addition of simple voice encryption processing in the Voice
`Preprocessor even higher levels of security can be attained. This directly
`supports a method of contracting based upon an offer perceived by the user, a
`recognizable acceptance of the offer by an identified user.
`
`The set-top appliance 1100 is also capable of receiving and decoding data in the
`downstream path. This function is required in order to organize and synchronize
`the transmission of upstream data. Downstream data can consist of upstream
`
`AGLE0001PR_v 2
`
`Page 5 of 24
`
`Comcast - Exhibit 1003, page 8
`
`

`

`Attorney Docket: AGLE0001PR
`
`channel allocation information and voice verification overlay information coded as
`text.
`
`For embodiments where the set-top box 1100 is used for both upstream and
`downstream communication for the described voice command function, the
`function of the Set-top appliance is only to receive the RF signal from the remote
`control and then digitize and compress the voice signal, further preparing it for
`upstream transmission.
`
`New RF protocol standards such as "Blue Tooth" allow the remote control's RF
`signal to transmit the voice signal directly to the set-top box where again, the
`
`preprocessing can either be done in the remote control 1000 or in firmware within
`the set-top box 1100.
`
`Set-top boxes 1100 that employ the DOCSIS type cable modems such as Open
`Cable set-top boxes or the so called "Heavy Set-top boxes" from Scientific
`Atlanta and General Instruments are capable of sending and receiving voice data
`using efficient data transmission protocols.
`The DOCSIS protocol also
`incorporates error detection and correction capabilities as well as other
`transmission enhancements such as pre-equalization for more efficient and error
`free transmission.
`
`The voice signal transmitted from Subscriber's set-top box or set-top appliance
`1100 is received 1510 by the 5 to 40 MHz data receiving equipment 2100 in the
`cable Headend.
`
`If the digitized voice signal comes from a commercial set-top box such as a
`General Instruments or a Scientific Atlanta set-top, then the return path receiving
`equipment in the headend is specific to that type of box. Therefore, the data
`coming from this equipment, which will contain other upstream traffic, is parsed in
`
`AGLE0001PR_v 2
`
`Page 6 of 24
`
`Comcast - Exhibit 1003, page 9
`
`

`

`Attorney Docket: AGLE0001 PR
`
`such a way that only the voice commands and address information from the
`subscriber are used by the AgileTV Voice Recognition Engine in the headend.
`
`If the digitized voice signal that's being sent upstream comes from the AgileTV's
`
`set-top appliance then the upstream data receiver in the headend is a separate
`standalone unit designed to receive only voice command signals from the
`AgileTV's set-top appliance in the subscribers home. Using the set-top
`appliance as the upstream transmitter allows the use of custom upstream
`protocols such as FM, AM, PSK or spread spectrum digital transmission. Digital
`transmission techniques such as QPSK or QAM can also be employed but
`
`require more costly transmission and receiver equipment.
`
`Upon receiving the digitized and preprocessed voice signal from the subscriber's
`set-top box or set-top appliance, the received upstream signal will be in the form
`of an Ethernet data stream containing voice and address information. Since the
`Agile Voice Processing Unit (AVPU) is a high speed voice processing unit
`capable of processing the data from several nodes, the digital voice signals from
`each of these nodes are combined into a single high speed digital bit stream in
`the input multiplexer of the AVPU)
`
`Voice Processing Engine description
`
`Upstream signals 1510 are received at the Agile Voice Processor Unit (AVPU)
`RPD 2100, in certain embodiments.
`
`1. Voice and data signals are received from commercial return path data
`receivers, or:
`
`2. Voice and data signals are received and decoded by custom return path
`receivers using at least one of the following protocol options: FM or AM
`
`AGLE0001 PR_v_2
`
`Page 7 of 24
`
`Comcast - Exhibit 1003, page 10
`
`

`

`Attorney Docket: AGLE0001 PR
`
`modulation/demodulation, FDMA, TDMA, FSK, PSK, or QPSK digital
`Spectrum modulation/demodulation,
`
`modulation/demodulation,
`
`Spread
`
`Telephony or cellular return or Wireless
`
`• Application Introduction
`
`The AVPU Engine is not an application service, in and of itself. While AgileTV
`may provide new end user applications, the primary function of the AVPU Engine
`
`is to provide voice control services for existing applications, such as Interactive
`Program Guides, and Video On Demand services.
`
`• Application Registration
`
`At system initialization time, applications such as VOD or Interactive Program
`Guides that wish to utilize voice recognition services must first register with the
`AVPU system. A standard program interface is utilized to enable each
`application to specify its complete menu hierarchy in the form of a tree structure.
`This tree contains "labels" for each menu, along with the text of each "button" on
`each menu screen. This provides the information to the AVPU engine to enable
`it to independently provide voice navigation services through the menu hierarchy
`on behalf of the application. This menu hierarchy represents the "static" portion
`of the application's data.
`
`In addition to the static menu structure, it is also the responsibility of the
`application to inform AVPU of "dynamic" content — for example, the names of
`movies in a VOD system, or program names and times in an interactive program
`guide. Each time a user enters a menu context in which dynamic content
`appears, the application will inform the system of this context by passing a
`"handle" associated with the list of names that comprise the dynamic content.
`The system will combine the static menu content with the augmented dynamic
`
`AGLE0001 PR_v 2
`
`Page 8 of 24
`
`Comcast - Exhibit 1003, page 11
`
`

`

`Attorney Docket: AGLE0001PR
`
`content (see Similarity Searching below), as well as application-independent
`keywords such as HELP, in order to form a complete "grammar". This construct
`is then passed to the voice recognition engine to maximize recognition accuracy.
`
`Given that dynamic content, by definition, varies, it is the application's
`responsibility to inform the system whenever the content changes.
`In an
`interactive TV guide application, for example, the application may register a new
`set of dynamic content every one-half hour. For a VOD system, this registration
`would be performed whenever the database of movies offered changes.
`
`• Application Interface
`
`Once registration has been completed, and the system is being used, recognition
`of an utterance will cause a signal to be sent back to the application. This signal
`will inform the application to perform the requested action, and/or to update the
`contents of the screen as a result of the user's request.
`In this manner, the
`application can utilize the system's voice recognition services with minimal
`modifications to the application's code, while retaining the same graphical "look
`and feel" that users have become accustomed to.
`
`• Using the System
`
`In the subscriber's home, AgileTV supplies a voice-enabled remote control,
`which contains both a microphone and, in certain further embodiments, a Push-
`To-Talk (PTT) switch, as well as traditional universal remote control functionality.
`While the conventional remote control functions are transmitted via IR, voice
`output is transmitted as RF to a small "VoiceLink" pod located at the set-top box.
`
`When the PTT button is pushed by the user, the remote control sends a "PTT
`active" command to the VoiceLink, which then informs the set-top box to place an
`
`AGLE0001 PR_v 2
`
`Page 9 of 24
`
`Comcast - Exhibit 1003, page 12
`
`

`

`Attorney Docket: AGLE0001 PR
`
`icon on the screen, indicating to the user that the system is "listening" to them.
`Next, as the user speaks into the microphone, the speech is digitized,
`compressed, and transmitted to the VoiceLink.
`
`The VoiceLink encrypts the speech sample to provide security, then adds
`
`subscriber address information, a length code, and a Cyclical Redundancy Code
`(CRC) to enable data transmission errors to be detected.
`
`In homes with "heavy" set-top boxes, the VoiceLink will transmit this voice
`information to the set-top box, which will then transmit it to the headend as a
`series of packets.
`
`Otherwise, the VoiceLink will directly transmit the voice stream to the headend
`itself. This process continues until the VoiceLink receives a "PTT Release" from
`the remote, indicating end of speech. This information is also transmitted to the
`headend, signaling end of utterance.
`
`AGLE0001PR_v_2
`
`Page 10 of 24
`
`Comcast - Exhibit 1003, page 13
`
`

`

`Attorney Docket: AGLE0001PR
`
`• Address Decoding
`
`Each individual consumer's interface (i.e. set-top box or set-top appliance) will
`have a unique address that is determined during the manufacturing process. As
`voice packets are transmitted upstream, this address information is pre-pended
`to the voice packets, enabling rapid determination of which household the voice
`sample is being received from. This address information is key to improving the
`efficiency of several different headend processing stages. The first address
`decode is used to assign an input buffer address to the sample. This input buffer
`is used to collect incoming voice packets until the final packet of a speech
`utterance has been received.
`
`Certain further embodiments use on-the-fly Cyclical Redundancy Code (CRC)
`error checking generation. Each time a packet is read in, CRC in CPU registers
`is computed as each byte is read, then the partial CRC is stored at the end of the
`stored packet. When the next packet arrives, the partial CRC is read from where
`it was stored, and the new packet data is appended to the end of the previous
`packet, overwriting the temporary CRC. This continues until a complete voice
`sample has been received. By doing this, memory accesses is cut in half
`compared to first storing the string, then making a second pass generating the
`CRC.
`
`Once a complete speech utterance has been received, the Input CPU will use
`the sample's source address to target the speech data to a specific voice
`processing CPU. This direct correspondence between the source address and a
`specific voice CPU is important, since it allows voice CPUs to efficiently cache
`user-specific parameters for the households they serve. Without this mapping,
`the bandwidth necessary to move household-specific data to each voice CPU
`would be prohibitive. In certain further embodiments, a translation table is
`actually used: this allows voice->CPU assignments to be changed dynamically in
`
`AGLE0001 PR_v 2
`
`Page 11 of 24
`
`Comcast - Exhibit 1003, page 14
`
`

`

`Attorney Docket: AGLE0001PR
`
`the event of a hardware failure, while retaining the efficiency advantages of direct
`mapping.
`
`• Grammar File is Loaded
`
`In order for a speech sample to be processed by the voice recognition engine, it
`is necessary for a voice CPU to first contain in its local memory a copy of the
`
`grammar definition associated with the household's set-top box state. Speech
`recognition is most effective when the speech engine is aware of which words
`are most likely to be spoken, the order in which these words may appear, and the
`meaning of various sequences of words — this information is contained in a
`construct known as a grammar.
`
`Before transferring the new speech sample to a voice CPU, the grammar
`associated with the speech sample is transferred to the target speech CPU,
`using a simple LRU queue.
`If the voice CPU contains empty space in its
`grammar buffer memory, then the indicated grammar is transferred directly to the
`empty buffer area from disk. If not, then the least-recently-used grammar buffer
`entry is discarded, and the new grammar information is loaded into the vacated
`buffer memory.
`
`• Household Parameter File is Loaded
`
`The next step in processing the voice sample is to ensure that the parameters
`associated with this household are already cached in the specific voice CPU's
`RAM.
`If these parameters are not present, then the least-recently-used
`parameter cache entry is evicted from the cache.
`
`To do this, the oldest cache entry on this voice CPU is first examined to see if it
`has been modified. If so, the cache entry will be written to disk, and the cache
`
`AGLE0001PR_v 2
`
`Page 12 of 24
`
`Comcast - Exhibit 1003, page 15
`
`

`

`Attorney Docket: AGLE0001PR
`
`slot is then declared vacant. Next, the household speech parameters associated
`with the new speech sample are loaded into the vacated cache block. During the
`relatively long access times needed to load a new set of household parameters
`from disk (and optionally to write the old parameters to disk), the current voice
`sample will be held in the input buffer in a "waiting" state.
`
`Only after the new household speech parameters have been loaded into the
`targeted voice CPU will the voice sample be moved into the work queue for the
`voice CPU. In this manner, the voice CPU is not held off from processing other
`voice requests during lengthy disk accesses.
`Instead, the voice CPU will
`continue to process other voice samples associated with households whose
`
`parameters are already in cache.
`
`• Assignment to a Voice CPU
`
`Once the voice parameters associated with a speech sample are finally cached
`in a voice CPU, the speech sample is assigned to the voice CPU by placing a
`descriptor for the voice sample on the target voice CPU's "work" queue. As
`speech samples are processed, they are removed from the front of the work
`queue by the voice CPU.
`
`Eventually, the voice CPU will reach the location of the current input sample.
`Once this occurs, the speech sample is transferred into the voice CPU's local
`memory under DMA control, and the status of this voice sample is changed to
`"Next". This transfer occurs in parallel with the processing of the prior speech
`sample, ensuring that voice CPU utilization is maximized.
`
`Once this transfer is complete, and the voice CPU completes processing of the
`prior sample, the status of this voice sample is changed to "Current", and the
`voice recognition engine begins processing this sample.
`
`AGLE0001PR_v 2
`
`Page 13 of 24
`
`Comcast - Exhibit 1003, page 16
`
`

`

`Attorney Docket: AGLE0001PR
`
`• Deadlock Elimination
`
`Successful processing of a speech sample requires both the proper grammar
`and the proper household parameter information be simultaneously loaded into a
`voice CPU. The possibility exists of having a race condition in which a recently-
`loaded grammar or speech parameter file is evicted prior to its use, in the
`process of loading the grammar or speech parameters for the current voice
`sample.
`
`To eliminate this race condition, the total number of speech samples sitting in the
`waiting and working queues of a voice CPU may not exceed the number of
`cache entries in the voice CPU.
`
`• Speaker Identification
`
`The first step in recognizing the current speech sample is to determine which
`individual person pronounced the current utterance. To do this, the "Speaker
`Identification" software module running on the targeted voice CPU compares the
`vocal characteristics of this speech sample with the characteristics of the
`speakers who have been previously identified in this household — these voice
`"templates" are an important component of the speech parameters that are
`cached in the CPU.
`
`In the vast majority of utterances, the incoming speech sample will match the
`characteristics of a previously-identified speaker. When this occurs, the speech
`sample is passed on to the next phase, speech recognition.
`
`If the speech sample is not identified with an existing speaker, then a "new user"
`routine is invoked, enabling a new user to be associated with this household.
`This routine records the new individual's speech parameters in this household's
`
`AGLE0001PR_v_2
`
`Page 14 of 24
`
`Comcast - Exhibit 1003, page 17
`
`

`

`Attorney Docket: AGLE0001PR
`
`speech parameters, so that during subsequent utterances, the new speaker will
`be identified by the speaker identification process.
`
`• Speech Recognition
`
`The inputs to the speech recognition software module are a speech sample, an
`individual user's speech parameters, and the grammar to be recognized. The
`speech engine determines the most likely utterance based on statistical analysis,
`and returns a text string corresponding to the utterance. As a statistical process,
`this matching process is probabilistic: along with the returned text string, the
`speech engine will also return a "percentage of match likelihood": this enables
`different applications to respond differently based on the calculated confidence in
`the recognition result.
`
`For recognition results having a low "cost", such as a request to display listings
`for a particular movie, lower confidence criteria need apply. For recognition
`results with a high cost to the user, such as a request to purchase a movie,
`higher confidence thresholds may be required (furthermore, purchase verification
`will be requested).
`
`When recognition accuracy is particularly low, and the voice recognition engine
`determined partial matches to more than one possible phrase, the engine will
`return the text of several possible matches. This process, known as "N-Best"
`enables an application, or the user, to select from several alternative recognition
`results.
`
`• Voice recording
`
`In cases where a transaction will result in a charge to the use