(19)
(11)EP 3 114 680 B1

(12)EUROPEAN PATENT SPECIFICATION

(45)Mention of the grant of the patent:
24.06.2020 Bulletin 2020/26

(21)Application number: 15717284.2

(22)Date of filing:  27.02.2015
(51)Int. Cl.: 
G06N 3/08  (2006.01)
G10L 15/16  (2006.01)
G10L 15/07  (2013.01)
(86)International application number:
PCT/US2015/017872
(87)International publication number:
WO 2015/134294 (11.09.2015 Gazette  2015/36)

(54)

LOW-FOOTPRINT ADAPTATION AND PERSONALIZATION FOR A DEEP NEURAL NETWORK

ADAPTION UND PERSONALISIERUNG MIT GERINGEM FUSSABDRUCK EINES TIEFEN NEURONALEN NETZWERKES

ADAPTATION ET PERSONNALISATION À FAIBLE MÉMOIRE D'UN RÉSEAU NEURONAL PROFOND


(84)Designated Contracting States:
AL AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HR HU IE IS IT LI LT LU LV MC MK MT NL NO PL PT RO RS SE SI SK SM TR

(30)Priority: 07.03.2014 US 201414201704

(43)Date of publication of application:
11.01.2017 Bulletin 2017/02

(73)Proprietor: Microsoft Technology Licensing, LLC
Redmond, WA 98052-6399 (US)

(72)Inventors:
  • XUE, Jian
    Redmond, Washington 98052-6399 (US)
  • LI, Jinyu
    Redmond, Washington 98052-6399 (US)
  • YU, Dong
    Redmond, Washington 98052-6399 (US)
  • SELTZER, Michael L.
    Redmond, Washington 98052-6399 (US)
  • GONG, Yifan
    Redmond, Washington 98052-6399 (US)

(74)Representative: Goddar, Heinz J. 
Boehmert & Boehmert Anwaltspartnerschaft mbB Pettenkoferstrasse 22
80336 München
80336 München (DE)


(56)References cited: : 
  
  • Jian Xue ET AL: "Restructuring of Deep Neural Network Acoustic Models with Singular Value Decomposition", Proc. Interspeech 2013, 25 August 2013 (2013-08-25), pages 2365-2369, XP055193450, Lyon, France Retrieved from the Internet: URL:http://www.isca-speech.org/archive/arc hive_papers/interspeech_2013/i13_2365.pdf [retrieved on 2015-06-03]
  • YU DONG ET AL: "KL-divergence regularized deep neural network adaptation for improved large vocabulary speech recognition", 2013 IEEE INTERNATIONAL CONFERENCE ON ACOUSTICS, SPEECH AND SIGNAL PROCESSING (ICASSP); VANCOUCER, BC; 26-31 MAY 2013, INSTITUTE OF ELECTRICAL AND ELECTRONICS ENGINEERS, PISCATAWAY, NJ, US, 26 May 2013 (2013-05-26), pages 7893-7897, XP032508812, ISSN: 1520-6149, DOI: 10.1109/ICASSP.2013.6639201 [retrieved on 2013-10-18]
  • GEOFFREY HINTON ET AL: "Deep Neural Networks for Acoustic Modeling in Speech Recognition: The Shared Views of Four Research Groups", IEEE SIGNAL PROCESSING MAGAZINE, IEEE SERVICE CENTER, PISCATAWAY, NJ, US, vol. 29, no. 6, 1 November 2012 (2012-11-01), pages 82-97, XP011469727, ISSN: 1053-5888, DOI: 10.1109/MSP.2012.2205597
  • XUE JIAN ET AL: "Singular value decomposition based low-footprint speaker adaptation and personalization for deep neural network", 2014 IEEE INTERNATIONAL CONFERENCE ON ACOUSTICS, SPEECH AND SIGNAL PROCESSING (ICASSP), IEEE, 4 May 2014 (2014-05-04), pages 6359-6363, XP032617895, DOI: 10.1109/ICASSP.2014.6854828 [retrieved on 2014-07-11]
  
Note: Within nine months from the publication of the mention of the grant of the European patent, any person may give notice to the European Patent Office of opposition to the European patent granted. Notice of opposition shall be filed in a written reasoned statement. It shall not be deemed to have been filed until the opposition fee has been paid. (Art. 99(1) European Patent Convention).


Description

BACKGROUND



[0001] Many computing devices, such as smartphones, desktops, laptops, tablets, game consoles, and the like, utilize automatic speech recognition (ASR) for performing a number of tasks including voice search and short message dictation. In an effort to improve the accuracy of ASR, the use of deep neural networks (DNNs) has been proposed. DNNs are artificial neural networks with more than one hidden layer between input and output layers and may model complex non-linear relationships. The hidden layers in DNNs provide additional levels of abstraction, thus increasing its modeling capability. DNNs when utilized in ASR however, suffer from a number of drawbacks associated with adaption and personalization. For example, the use of DNNs, while increasing ASR accuracy, also is accompanied by a very large number of parameters making the adaptation of DNN models very challenging. Furthermore, the cost associated with using DNNs in personalized ASR applications (i.e., multiple individual speakers) is prohibitive due to the need to store very large DNN models for each individual speaker during deployment. This problem is e.g. discussed in the conference paper of J.Xue, J. Li and Y. Gong, "Restructuring of Deep Neural Network Acoustic Models with Singular Value Decomposition", Interspeech, 2013, Lyon, France pp. 2365-2369. It is with respect to these considerations and others that the various embodiments of the present invention have been made.

SUMMARY



[0002] This summary is provided to introduce a selection of concepts in a simplified form that are further described below in the Detailed Description.

[0003] Embodiments provide for the adaptation and personalization of a deep neural network (DNN) model for automatic speech recognition. Utterances which include speech features for many speakers are used to train the DNN model. A decomposition approach such as low-rank factorization may then be applied to an original weight matrix in the DNN model. As a result of applying the decomposition approach, the original weight matrix may be converted into multiplications of multiple new matrices which are smaller than the original matrix. A square matrix may then be added to these matrices. Speaker-specific parameters may then be stored in the square matrix which is initialized as an identity matrix. The DNN model may then be adapted by updating the square matrix. The adapted DNN model may include a reduced number of model parameters than those received in the original DNN model. This process may be applied to all of a number of original matrices in the DNN model.

[0004] These and other features and advantages will be apparent from a reading of the following detailed description and a review of the associated drawings. It is to be understood that both the foregoing general description and the following detailed description are illustrative only and are not restrictive of the invention as claimed.

BRIEF DESCRIPTION OF THE DRAWINGS



[0005] 

FIGURE 1 is a block diagram illustrating a system for adapting and personalizing a deep neural network model for automatic speech recognition, in accordance with an example;

FIGURE 2 is a block diagram illustrating the adaptation of a deep neural network model, in accordance with an example;

FIGURE 3 is a flow diagram illustrating a routine for adapting and personalizing a deep neural network model for automatic speech recognition, in accordance with an embodiment;

FIGURE 4 is a flow diagram illustrating a routine for adapting and personalizing a deep neural network model for automatic speech recognition, in accordance with another embodiment;

FIGURE 5 is a simplified block diagram of a computing device with which various examples may be practiced;

FIGURE 6A is a simplified block diagram of a mobile computing device with which various examples may be practiced;

FIGURE 6B is a simplified block diagram of a mobile computing device with which various examples may be practiced; and

FIGURE 7 is a simplified block diagram of a distributed computing system in which various examples may be practiced.


DETAILED DESCRIPTION



[0006] Embodiments provide for the adaptation and personalization of a deep neural network (DNN) model for automatic speech recognition. Utterances which include speech features for many speakers are used to train the DNN model. A decomposition approach such as low-rank factorization may then be applied to an original weight matrix in the DNN model. As a result of applying the decomposition approach, the original weight matrix may be converted into multiplications of multiple new matrices which are smaller than the original matrix. A square matrix may then be added to these matrices. The speaker-specific parameters may then be stored in the square matrix which is initialized as the identity matrix. The DNN model may then be adapted by updating the square matrix. This process may be applied to all of a number of original matrices in the DNN model. The adapted DNN model may include a reduced number of model parameters than those received in the original DNN model.

[0007] In the following detailed description, references are made to the accompanying drawings that form a part hereof, and in which are shown by way of illustrations specific embodiments or examples. These embodiments may be combined, other embodiments may be utilized, and structural changes may be made without departing from the scope of the present invention. The following detailed description is therefore not to be taken in a limiting sense, and the scope of the present invention is defined by the appended claims and their equivalents.

[0008] Referring now to the drawings, in which like numerals represent like elements through the several figures, various aspects of the present invention will be described. FIGURE 1 is a block diagram illustrating a system 100 which may be utilized for adapting and personalizing a deep neural network model for automatic speech recognition, in accordance with an example. The system 100, which may comprise a conversational dialog system, includes a computing device 150 which may receive utterances 125 and 130 from one or more speakers 120 (e.g., Speakers 1 through N). The utterances 125 and 130 received by the computing device 150 may comprise a short message dictation 160 and/or a voice search query 165 which may be displayed to a user in user interface 155. The utterances 125 and 130 may include speech features 135 and 140 which may be received from one or more ASR tasks such as the short message dictation 155 and/or a voice search query 165.

[0009] In accordance with various examples, the computing device 150 may comprise, without limitation, a desktop computer, laptop computer, smartphone, video game console or a television. The computing device 150 may also comprise or be in communication with one or more recording devices (not shown) used to detect speech and receive video/pictures (e.g., MICROSOFT KINECT, microphone(s), and the like). The computing device 150 may store an application 170 which, as will be described in greater detail below, may be configured to receive the speech features 135 and 140 contained in the utterances 125 and 130.

[0010] As will further be described in greater detail below, the application 170 may also be configured to generate deep neural network (DNN) models 105 for use in automatic speech recognition. In accordance with examples described herein, the DNN models 105 may include one or more unadapted DNN models 110, each of which includes parameters 111 (i.e., model parameters) and one or more adapted DNN models 115, each of which may include parameters 116 (i.e., speaker-specific model parameters). As will be described in greater detail below with respect to FIGURES 2-4, a decomposition approach 114 (hereinafter referred to as SVD 114) may be applied to original matrices 112 in the unadapted DNN model 110 for adaptation such that the number of parameters 116 in the adapted DNN model 115 is reduced when compared to the number of parameters 111 in the unadapted DNN model 110. The adapted DNN model 115 may include matrices 117. The matrices 117 may comprise additional matrices converted from the original matrices 112 in the unadapted DNN model, as a result of the applied SVD 114, during adaptation operations performed by the application 170 which are described in detail below with respect to FIGURE 3. Furthermore, as a result of adaptation operations performed by the application 170 which are described in detail below with respect to FIGURE 4, the adapted DNN model 115 may also include delta matrices 118 and small matrices 119. In accordance with an example, the application 130 may comprise the BING VOICE SEARCH, WINDOWS PHONE SHORT MESSAGE DICTATION and XBOX MARKET PLACE VOICE SEARCH applications from MICROSOFT CORPORATION of Redmond Washington. It should be understood, however, that other applications (including operating systems) from other manufacturers may alternatively be utilized in accordance with the various examples described herein.

[0011] FIGURE 2 is a block diagram illustrating the adaptation of a DNN model, in accordance with an example. The adapted DNN model 115 may include layers 207, 209 and 211. The layer 207, which may comprise a non- linear function 205 and a matrix 210, is a non- linear layer while the layers 209 and 211 (which may comprise linear functions 214 and 225 as well as matrices 220 and 230, respectively) are linear layers. The adapted DNN model 115 may also comprise an additional linear function 235. As will be described in greater detail below with respect to FIGURE 3, the matrix 220 may comprise a square or identity matrix which is the only matrix updated during adaption and which contains a smaller number of parameters than the unadapted (i.e., original) DNN model 110.

[0012] FIGURE 3 is a flow diagram illustrating a routine 300 for adapting and personalizing a deep neural network model for automatic speech recognition, in accordance with an embodiment. When reading the discussion of the routines presented herein, it should be appreciated that the logical operations of various embodiments of the present invention are implemented (1) as a sequence of computer implemented acts or program modules running on a computing system and/or (2) as interconnected machine logical circuits or circuit modules within the computing system. The implementation is a matter of choice dependent on the performance requirements of the computing system implementing the invention. Accordingly, the logical operations illustrated in FIGURES 3-4 and making up the various embodiments described herein are referred to variously as operations, structural devices, acts or modules. It will be recognized by one skilled in the art that these operations, structural devices, acts and modules may be implemented in software, in hardware, in firmware, in special purpose digital logic, and any combination thereof without deviating from the scope of the present invention as recited within the claims set forth herein.

[0013] The routine 300 begins at operation 305, where the application 170 executing on the computing device 150 may receive one or more of the utterances 125, 130 including the speech features 135, 140, from the speakers 120. For example, the utterances 125, 130 may comprise short message dictation 160 and/or voice search query 165.

[0014] From operation 305, the routine 300 continues to operation 310, where the application 170 executing on the computing device 150 may apply the SVD 114 to an original matrix (i.e., the one or more of the original matrices 112) in the unadapted DNN model 110.

[0015] From operation 310, the routine 300 continues to operation 315, where the application 170 executing on the computing device 150 may convert the original matrix (i.e., the original matrices 112) into new matrices (i.e., the matrices 117) and replace an original layer with new layers (i.e., the layers 207, 209 and 211).

[0016] In accordance with an embodiment, the original matrix may comprise an m x n weight matrix A in the unadapted DNN model 110 and the operations 305 and 310 may be mathematically represented as follows when the SVD 114 has been applied:

where Σ may be a diagonal matrix with A's singular values on the diagonal. In accordance with an embodiment, if A is a sparse matrix, then the number of A's non-zero singular values may be much smaller than n, which may be represented by the variable k. In this case, the aforementioned expression may be rewritten as follows:

where U and N represent the new matrices 117. It should be understood that the new matrices 117 may be smaller than any of the original matrices 112. For example, the new matrices 117 may be smaller in one or more dimensions or be lower ranked than the original matrices 112.

[0017] From operation 315, the routine 300 continues to operation 320, where the application 170 executing on the computing device 150 may add a square matrix (i.e., the matrix 220) to the new matrices 117. As discussed above, the square matrix may comprise an identity matrix I. The aforementioned operation may be mathematically represented as follows:

It should be understood that as a result of the operations 305-315, the matrix A is converted into three matrices U, N and I and, correspondingly an original layer in the unadapted DNN model 110 is replaced with three layers (i.e., the layers 207, 209 and 211) with two of the layers being linear layers and the other layer being non-linear.

[0018] From operation 320, the routine 300 continues to operation 325, where the application 170 executing on the computing device 150 may adapt the unadapted DNN model 110 by updating the square matrix 220 to generate the adapted DNN model 115. In particular, the application 170 may only update the identity matrix I in the formula discussed above with respect to operation 320. It should be understood that the number of parameters for the identity matrix I may be k^2which represent a much smaller number of parameters than those originally received by the unadapted (i.e., original) DNN model 110 when k is small (as discussed above). It should be further understood that the adapted DNN model 115 represents a reduction in the overall number of speech parameters received by the original unadapted DNN model 110. Moreover, it should be appreciated that the routine 300 may be applied to all of the original matrices 112 in the unadapted DNN model 110 such that multiple square matrices are applied for adaptation. From operation 325, the routine 300 then ends.

[0019] FIGURE 4 is a flow diagram illustrating a routine 400 for adapting and personalizing a deep neural network model for automatic speech recognition, in accordance with an embodiment. The routine 400 begins at operation 405, where the application 170 executing on the computing device 150 may receive one or more of the utterances 125, 130 including the speech features 135, 140, from the speakers 120. For example, the utterances 125, 130 may comprise short message dictation 160 and/or voice search query 165.

[0020] From operation 405, the routine 400 continues to operation 410, where the application 170 executing on the computing device 150 may apply determine the adapted DNN model 115 from the unadapted DNN model 110 as described above with respect to operations 305-325 of FIGURE 3.

[0021] From operation 410, the routine 400 continues to operation 415, where the application 170 executing on the computing device 150 may calculate the difference between the matrices in the adapted DNN model 115 and unadapted DNN model 110 to determine the delta matrices 118. As should be understood by those skilled in the art, in some embodiments, the adapted DNN model may have the same format and structure as the unadapted DNN model 110 which may be very large. As a result, the storing of large DNN models for individual speakers during model deployment (i.e., personalization) may be very costly. Delta matrices however, comprise sparse matrices and have a very low rank. Thus, the small matrices 119 converted from the delta matrices 118 have a low footprint and enable the personalization of smaller and less costly DNN models for use in automatic speech recognition.

[0022] From operation 415, the routine 400 continues to operation 420, where the application 170 executing on the computing device 150 may apply the SVD 114 to each of the calculated delta matrices 118 while only maintaining a small amount of non-zero singular values for each delta matrix. As a result, after the application of the SVD 114, the delta matrices 118 may be largely unchanged.

[0023] From operation 420, the routine 400 continues to operation 425, where the application 170 executing on the computing device 150 may convert the delta matrices 118 into a subset of small matrices 119. In particular, each delta matrix 118 may be converted into the product of two low-rank matrices. In accordance with an embodiment, the operations 415-425 may be mathematically represented as follows:



[0024] From operation 425, the routine 400 continues to operation 430, where the application 170 executing on the computing device 150 may store the subset of small matrices 119. In particular, the application 170 may only store the matrices U and N (from the above mathematical expression) for each individual speaker. It should be understood that, in accordance with an embodiment, the total number of parameters in the above mathematical expression may change from mn to (m+n)k. From operation 430, the routine 400 then ends.

[0025] In accordance with some embodiments, the routines 300 and 400 discussed above with respect to FIGURES 3 and 4 may be combined. For example, the routine 300 may be utilized to perform SVD adaptation such that the difference between the adapted DNN model and the unadapted DNN model are square matrices between each of a pair of linear layers. SVD may further be applied on delta matrices to save DNN model costs even further.

[0026] FIGURE 5-7 and the associated descriptions provide a discussion of a variety of operating environments in which examples of the invention may be practiced. However, the devices and systems illustrated and discussed with respect to FIGURES 5-7 are for purposes of example and illustration and are not limiting of a vast number of computing device configurations that may be utilized for practicing examples of the invention, described herein.

[0027] FIGURE 5 is a block diagram illustrating example physical components of a computing device 500 with which various examples may be practiced. In a basic configuration, the computing device 500 may include at least one processing unit 502 and a system memory 504. Depending on the configuration and type of computing device, system memory 504 may comprise, but is not limited to, volatile (e.g. random access memory (RAM)), non-volatile (e.g. read-only memory (ROM)), flash memory, or any combination. System memory 504 may include an operating system 505 and application 170. Operating system 505, for example, may be suitable for controlling the computing device 500's operation and, in accordance with an example, may comprise the WINDOWS operating systems from MICROSOFT CORPORATION of Redmond, Washington. The application 170 (which, in some examples, may be included in the operating system 505) may comprise functionality for performing routines including, for example, the adaptation and personalization of a deep neural network (DNN) model for automatic speech recognition, as described above with respect to the operations in routines 300-400 of FIGURES 3-4.

[0028] The computing device 500 may have additional features or functionality. For example, the computing device 500 may also include additional data storage devices (removable and/or non-removable) such as, for example, magnetic disks, optical disks, solid state storage devices ("SSD"), flash memory or tape. Such additional storage is illustrated in FIGURE 5 by a removable storage 509 and a non-removable storage 510. The computing device 500 may also have input device(s) 512 such as a keyboard, a mouse, a pen, a sound input device (e.g., a microphone), a touch input device for receiving gestures, an accelerometer or rotational sensor, etc. Output device(s) 514 such as a display, speakers, a printer, etc. may also be included. The aforementioned devices are examples and others may be used. The computing device 500 may include one or more communication connections 516 allowing communications with other computing devices 518. Examples of suitable communication connections 516 include, but are not limited to, RF transmitter, receiver, and/or transceiver circuitry; universal serial bus (USB), parallel, and/or serial ports.

[0029] Furthermore, various examples may be practiced in an electrical circuit comprising discrete electronic elements, packaged or integrated electronic chips containing logic gates, a circuit utilizing a microprocessor, or on a single chip containing electronic elements or microprocessors. For example, various examples may be practiced via a system-on-a-chip ("SOC") where each or many of the components illustrated in FIGURE 5 may be integrated onto a single integrated circuit. Such an SOC device may include one or more processing units, graphics units, communications units, system virtualization units and various application functionality all of which are integrated (or "burned") onto the chip substrate as a single integrated circuit. When operating via an SOC, the functionality, described herein may operate via application-specific logic integrated with other components of the computing device/system 500 on the single integrated circuit (chip). Examples may also be practiced using other technologies capable of performing logical operations such as, for example, AND, OR, and NOT, including but not limited to mechanical, optical, fluidic, and quantum technologies. In addition, examples may be practiced within a general purpose computer or in any other circuits or systems.

[0030] The term computer readable media as used herein may include computer storage media. Computer storage media may include volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information, such as computer readable instructions, data structures, or program modules. The system memory 504, the removable storage device 509, and the non-removable storage device 510 are all computer storage media examples (i.e., memory storage.) Computer storage media may include RAM, ROM, electrically erasable read-only memory (EEPROM), flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other article of manufacture which can be used to store information and which can be accessed by the computing device 500. Any such computer storage media may be part of the computing device 500. Computer storage media does not include a carrier wave or other propagated or modulated data signal.

[0031] Communication media may be embodied by computer readable instructions, data structures, program modules, or other data in a modulated data signal, such as a carrier wave or other transport mechanism, and includes any information delivery media. The term "modulated data signal" may describe a signal that has one or more characteristics set or changed in such a manner as to encode information in the signal. By way of example, and not limitation, communication media may include wired media such as a wired network or direct- wired connection, and wireless media such as acoustic, radio frequency (RF), infrared, and other wireless media.

[0032] FIGURES 6A and 6B illustrate a suitable mobile computing environment, for example, a mobile computing device 650 which may include, without limitation, a smartphone, a tablet personal computer, a laptop computer and the like, with which various examples may be practiced. With reference to FIG. 6A, an example mobile computing device 650 for implementing the examples is illustrated. In a basic configuration, mobile computing device 650 is a handheld computer having both input elements and output elements. Input elements may include touch screen display 625 and input buttons 610 that allow the user to enter information into mobile computing device 650. Mobile computing device 650 may also incorporate an optional side input element 620 allowing further user input. Optional side input element 620 may be a rotary switch, a button, or any other type of manual input element. In alternative examples, mobile computing device 650 may incorporate more or less input elements. In yet another alternative example, the mobile computing device is a portable telephone system, such as a cellular phone having display 625 and input buttons 610. Mobile computing device 650 may also include an optional keypad 605. Optional keypad 605 may be a physical keypad or a "soft" keypad generated on the touch screen display.

[0033] Mobile computing device 650 incorporates output elements, such as display 625, which can display a graphical user interface (GUI). Other output elements include speaker 630 and LED 680. Additionally, mobile computing device 650 may incorporate a vibration module (not shown), which causes mobile computing device 650 to vibrate to notify the user of an event. In yet another example, mobile computing device 650 may incorporate a headphone jack (not shown) for providing another means of providing output signals.

[0034] Although described herein in combination with mobile computing device 650, in alternative examples may be used in combination with any number of computer systems, such as in desktop environments, laptop or notebook computer systems, multiprocessor systems, micro-processor based or programmable consumer electronics, network PCs, mini computers, main frame computers and the like. Various examples may also be practiced in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network in a distributed computing environment; programs may be located in both local and remote memory storage devices. To summarize, any computer system having a plurality of environment sensors, a plurality of output elements to provide notifications to a user and a plurality of notification event types may incorporate the various examples described herein.

[0035] FIG. 6B is a block diagram illustrating components of a mobile computing device used in one example, such as the mobile computing device 650 shown in FIG. 6A. That is, mobile computing device 650 can incorporate a system 602 to implement some examples. For example, system 602 can be used in implementing a "smartphone" that can run one or more applications similar to those of a desktop or notebook computer. In some examples, the system 602 is integrated as a computing device, such as an integrated personal digital assistant (PDA) and wireless phone.

[0036] Application 170 may be loaded into memory 662 and run on or in association with an operating system 664. The system 602 also includes non-volatile storage 668 within memory the 662. Non-volatile storage 668 may be used to store persistent information that should not be lost if system 602 is powered down. The application 170 may use and store information in the non- volatile storage 668. The application 170, for example, may comprise functionality for performing routines including, for example, the adaptation and personalization of a deep neural network (DNN) model for automatic speech recognition, as described above with respect to the operations in routines 300-400 of FIGURES 3-4.

[0037] A synchronization application (not shown) also resides on system 602 and is programmed to interact with a corresponding synchronization application resident on a host computer to keep the information stored in the non-volatile storage 668 synchronized with corresponding information stored at the host computer. As should be appreciated, other applications may also be loaded into the memory 662 and run on the mobile computing device 650.

[0038] The system 602 has a power supply 670, which may be implemented as one or more batteries. The power supply 670 might further include an external power source, such as an AC adapter or a powered docking cradle that supplements or recharges the batteries.

[0039] The system 602 may also include a radio 672 (i.e., radio interface layer) that performs the function of transmitting and receiving radio frequency communications. The radio 672 facilitates wireless connectivity between the system 602 and the "outside world," via a communications carrier or service provider. Transmissions to and from the radio 672 are conducted under control of OS 664. In other words, communications received by the radio 672 may be disseminated to the application 170 via OS 664, and vice versa.

[0040] The radio 672 allows the system 602 to communicate with other computing devices, such as over a network. The radio 672 is one example of communication media. The example of the system 602 is shown with two types of notification output devices: the LED 680 that can be used to provide visual notifications and an audio interface 674 that can be used with speaker 630 to provide audio notifications. These devices may be directly coupled to the power supply 670 so that when activated, they remain on for a duration dictated by the notification mechanism even though processor 660 and other components might shut down for conserving battery power. The LED 680 may be programmed to remain on indefinitely until the user takes action to indicate the powered-on status of the device. The audio interface 674 is used to provide audible signals to and receive audible signals from the user. For example, in addition to being coupled to speaker 630, the audio interface 674 may also be coupled to a microphone (not shown) to receive audible (e.g., voice) input, such as to facilitate a telephone conversation. In accordance with examples, the microphone may also serve as an audio sensor to facilitate control of notifications. The system 602 may further include a video interface 676 that enables an operation of on-board camera 640 to record still images, video streams, and the like.

[0041] A mobile computing device implementing the system 602 may have additional features or functionality. For example, the device may also include additional data storage devices (removable and/or non-removable) such as, magnetic disks, optical disks, or tape. Such additional storage is illustrated in FIG. 6B by storage 668.

[0042] Data/information generated or captured by the mobile computing device 650 and stored via the system 602 may be stored locally on the mobile computing device 650, as described above, or the data may be stored on any number of storage media that may be accessed by the device via the radio 672 or via a wired connection between the mobile computing device 650 and a separate computing device associated with the mobile computing device 650, for example, a server computer in a distributed computing network such as the Internet. As should be appreciated such data/information may be accessed via the mobile computing device 650 via the radio 672 or via a distributed computing network. Similarly, such data/information may be readily transferred between computing devices for storage and use according to well-known data/information transfer and storage means, including electronic mail and collaborative data/information sharing systems.

[0043] FIGURE 7 is a simplified block diagram of a distributed computing system in which various examples may be practiced. The distributed computing system may include number of client devices such as a computing device 703, a tablet computing device 705 and a mobile computing device 710. The client devices 703, 705 and 710 may be in communication with a distributed computing network 715 (e.g., the Internet). A server 720 is in communication with the client devices 703, 705 and 710 over the network 715. The server 720 may store application 170 which may be perform routines including, for example, the adaptation and personalization of a deep neural network (D N) model for automatic speech recognition, as described above with respect to the operations in routines 300-400 of FIGURES 3-4.

[0044] Content developed, interacted with, or edited in association with the application 170 may be stored in different communication channels or other storage types. For example, various documents may be stored using a directory service 722, a web portal 724, a mailbox service 726, an instant messaging store 728, or a social networking site 730.

[0045] The application 170 may use any of these types of systems or the like for enabling data utilization, as described herein. The server 720 may provide the proximity application 170 to clients. As one example, the server 720 may be a web server providing the application 170 over the web. The server 720 may provide the application 170 over the web to clients through the network 715. By way of example, the computing device 10 may be implemented as the computing device 703 and embodied in a personal computer, the tablet computing device 705 and/or the mobile computing device 710 (e.g., a smart phone). Any of these examples of the computing devices 703, 705 and 710 may obtain content from the store 716.

[0046] Various embodiments are described above with reference to block diagrams and/or operational illustrations of methods, systems, and computer program products. The functions/acts noted in the blocks may occur out of the order as shown in any flow diagram. For example, two blocks shown in succession may in fact be executed substantially concurrently or the blocks may sometimes be executed in the reverse order, depending upon the functionality/acts involved.

[0047] The description and illustration of one or more embodiments provided in this application are not intended to limit or restrict the scope of the invention as claimed in any way. The embodiments, examples, and details provided in this application are considered sufficient to convey possession and enable others to make and use the best mode of claimed invention. The claimed invention should not be construed as being limited to any embodiment, example, or detail provided in this application.


Claims

1. A method of adapting and personalizing a deep neural network (DNN) model for automatic speech recognition (ASR), comprising:

receiving (305), by a computing device, at least one utterance comprising a plurality of speech features for one or more speakers from one or more ASR tasks;

applying (310), by the computing device, a decomposition approach to an original matrix in the DNN model;

in response to applying the decomposition approach, converting (315) the original matrix into a plurality of new matrices, each of the plurality of new matrices being smaller than the original matrix;

adding (320), by the computing device, a square matrix to the plurality of new matrices, the square matrix storing speaker-specific parameters; and

adapting (325), by the computing device, the DNN model by updating the square matrix, the adapted DNN model comprising a reduction in a number of parameters in the DNN model.


 
2. The method of claim 1, further comprising replacing an original layer in the DNN model with a plurality of new layers.
 
3. The method of claim 2, wherein at least one of the plurality of new layers comprises a non-linear layer.
 
4. The method of claim 1, wherein applying, by the computing device, a decomposition approach to an original matrix in the DNN model comprises applying singular value decomposition (SVD) to the original matrix in the DNN model.
 
5. A system for adapting and personalizing a deep neural network (DNN) model for automatic speech recognition (ASR), comprising:

a memory for storing executable program code; and

a processor, functionally coupled to the memory, the processor being responsive to computer-executable instructions contained in the program code and operative to:
receive (405) at least one utterance comprising a plurality of speech features for one or more speakers from one or more ASR tasks;

determine (410) an adapted DNN model from the DNN model, the DNN model comprising a plurality of unadapted matrices and the adapted DNN model comprising a plurality of adapted matrices;

calculate (415) a difference between the plurality adapted matrices and the plurality of unadapted matrices to determine a plurality of delta matrices;

apply (420) a decomposition approach to each of the plurality of delta matrices;

convert each of the plurality of delta matrices into a subset of small matrices; and

store (430) the subset of small matrices, the subset of small matrices comprising a small percentage of a plurality of parameters in the DNN model.


 
6. The system of claim 5, wherein the processor, in applying a decomposition approach to each of the plurality of delta matrices, is operative to apply singular value decomposition (SVD) to each of the plurality of delta matrices.
 
7. The system of claim 5, wherein the processor, in converting each of the plurality of delta matrices into a subset of small matrices, is operative to convert the product of two low-rank matrices.
 
8. A computer-readable storage medium storing computer executable instructions which, when executed by a computer, will cause computer to perform a method of adapting and personalizing a deep neural network (DNN) model for automatic speech recognition (ASR), the method comprising:

receiving (305) at least one utterance comprising a plurality of speech features for one or more speakers from one or more ASR tasks;

applying (310) a decomposition approach to an original matrix in the DNN model;

in response to applying the decomposition approach, converting (315) the original matrix into a plurality of new matrices, each of the plurality of new matrices being smaller than the original matrix;

adding (320) a square matrix to the plurality of new matrices, the square matrix storing speaker-specific parameters; and

adapting (325) the DNN model by only updating the square matrix, the adapted DNN model comprising a reduction in a number of the plurality of parameters in the DNN model.


 
9. The computer-readable storage medium of claim 8, further comprising replacing an original layer in the DNN model with a plurality of new layers.
 
10. The computer-readable storage medium of claim 9, wherein at least one of the plurality of new layers comprises a non-linear layer, the non-linear layer comprising a non-linear function.
 


Ansprüche

1. Ein Verfahren zur Anpassung und Personalisierung eines Modells eines tiefen neuronalen Netzes (DNN) für die automatische Spracherkennung (ASR), umfassend:

Empfangen (305) mindestens einer Äußerung, die eine Vielzahl von Sprachmerkmalen für einen oder mehrere Sprecher aus einer oder mehreren ASR-Aufgaben umfasst, durch ein Computergerät;

Anwenden (310) eines Zerlegungsansatzes auf eine Originalmatrix im DNN-Modell durch die Rechenvorrichtung;

als Reaktion auf die Anwendung des Zerlegungsansatzes, Umwandeln (315) der ursprünglichen Matrix in eine Vielzahl von neuen Matrizen, wobei jede der Vielzahl von neuen Matrizen kleiner als die ursprüngliche Matrix ist;

Hinzufügen (320) einer quadratischen Matrix zu der Vielzahl neuer Matrizen durch die Rechenvorrichtung, wobei die quadratische Matrix sprecherspezifische Parameter speichert; und

Anpassung (325) des DNN-Modells durch die Rechenvorrichtung durch Aktualisierung der Quadratmatrix, wobei das angepasste DNN-Modell eine Reduzierung einer Anzahl von Parametern im DNN-Modell umfasst.


 
2. Das Verfahren nach Anspruch 1, ferner umfassend das Ersetzen einer ursprünglichen Schicht im DNN-Modell durch eine Vielzahl neuer Schichten.
 
3. Das Verfahren nach Anspruch 2, wobei mindestens eine der mehreren neuen Schichten eine nichtlineare Schicht umfasst.
 
4. Das Verfahren nach Anspruch 1, wobei die Anwendung eines Zerlegungsansatzes auf eine ursprüngliche Matrix im DNN-Modell durch die Rechenvorrichtung die Anwendung der Singulärwertzerlegung (SVD) auf die ursprüngliche Matrix im DNN-Modell umfasst.
 
5. Ein System zum Anpassen und Personalisieren eines Modells eines tiefen neuronalen Netzes (DNN) für die automatische Spracherkennung (ASR), umfassend:

einen Speicher zum Speichern von ausführbarem Programmcode; und

einen Prozessor, der funktionell mit dem Speicher gekoppelt ist, wobei der Prozessor auf computerausführbare Anweisungen reagiert, die im Programmcode enthalten sind, und wobei er betriebsfähig ist zum:

Empfang (405) mindestens einer Äußerung, die eine Vielzahl von Sprachmerkmalen für einen oder mehrere Sprecher aus einer oder mehreren ASR-Aufgaben umfasst;

Bestimmen (410) eines angepassten DNN-Modells aus dem DNN-Modell, wobei das DNN-Modell eine Vielzahl von nicht angepassten Matrizen umfasst und das angepasste DNN-Modell eine Vielzahl von angepassten Matrizen umfasst;

Berechnen (415) einer Differenz zwischen der Vielzahl angepasster Matrizen und der Vielzahl nicht angepasster Matrizen, um eine Vielzahl von Delta-Matrizen zu bestimmen;

Anwenden (420) eines Zerlegungsansatzes auf jede der Vielzahl von Delta-Matrizen;

Umwandeln jeder der Vielzahl von Delta-Matrizen in eine Teilmenge von kleinen Matrizen; und

Speichern (430) der Teilmenge kleiner Matrizen, wobei die Teilmenge der kleinen Matrizen einen kleinen Prozentsatz einer Vielzahl von Parametern im DNN-Modell umfasst.


 
6. Das System nach Anspruch 5, wobei der Prozessor bei der Anwendung eines Zerlegungsansatzes auf jede der Vielzahl von Delta-Matrizen so arbeitet, dass er die Singulärwertzerlegung (SVD) auf jede der Vielzahl von Delta-Matrizen anwendet.
 
7. Das System nach Anspruch 5, bei dem der Prozessor bei der Umwandlung jeder der Vielzahl von Delta-Matrizen in eine Teilmenge von kleinen Matrizen betriebsfähig ist, das Produkt von zwei Matrizen niedrigen Ranges umzuwandeln.
 
8. Ein computerlesbares Speichermedium, das computerausführbare Befehle speichert, die, wenn sie von einem Computer ausgeführt werden, den Computer veranlassen, ein Verfahren zur Anpassung und Personalisierung eines Modells eines tiefen neuronalen Netzes (DNN) für automatische Spracherkennung (ASR) durchzuführen, wobei das Verfahren umfasst:

Empfangen (305) mindestens einer Äußerung mit einer Vielzahl von Sprachmerkmalen für einen oder mehrere Sprecher aus einer oder mehreren ASR-Aufgaben;

Anwenden (310) eines Zerlegungsansatzes auf eine Originalmatrix im DNN-Modell;

als Reaktion auf die Anwendung des Zerlegungsansatzes, Umwandeln (315) der ursprünglichen Matrix in eine Vielzahl neuer Matrizen, wobei jede der Vielzahl neuer Matrizen kleiner als die ursprüngliche Matrix ist;

Hinzufügen (320) einer quadratischen Matrix zu der Vielzahl von neuen Matrizen, wobei die quadratische Matrix sprecherspezifische Parameter speichert; und

Anpassung (325) des DNN-Modells, indem nur die Quadratmatrix aktualisiert wird, wobei das angepasste DNN-Modell eine Reduzierung einer Anzahl der Vielzahl von Parametern im DNN-Modell umfasst.


 
9. Das computerlesbare Speichermedium nach Anspruch 8, das ferner den Ersatz einer ursprünglichen Schicht im DNN-Modell durch eine Vielzahl neuer Schichten umfasst.
 
10. Das computerlesbare Speichermedium nach Anspruch 9, wobei mindestens eine der mehreren neuen Schichten eine nicht-lineare Schicht umfasst, wobei die nicht-lineare Schicht eine nicht-lineare Funktion umfasst.
 


Revendications

1. Procédé d'adaptation et de personnalisation d'un modèle de réseau neuronal profond (DNN) pour une reconnaissance automatique de la parole (ASR), comprenant les étapes ci-dessous consistant à :

recevoir (305), par le biais d'un dispositif informatique, au moins un énoncé comprenant une pluralité de caractéristiques de parole pour un ou plusieurs locuteurs à partir d'une ou plusieurs tâches de reconnaissance ASR ;

appliquer (310), par le biais du dispositif informatique, une approche de décomposition, à une matrice d'origine dans le modèle de réseau DNN ;

en réponse à l'application de l'approche de décomposition, convertir (315) la matrice d'origine en une pluralité de nouvelles matrices, chaque matrice de la pluralité de nouvelles matrices étant plus petite que la matrice d'origine ;

ajouter (320), par le biais du dispositif informatique, une matrice carrée, à la pluralité de nouvelles matrices, la matrice carrée stockant des paramètres spécifiques au locuteur ; et

adapter (325), par le biais du dispositif informatique, le modèle de réseau DNN, en mettant à jour la matrice carrée, le modèle de réseau DNN adapté comprenant une réduction d'un nombre de paramètres dans le modèle de réseau DNN.


 
2. Procédé selon la revendication 1, comprenant en outre l'étape consistant à remplacer une couche d'origine, dans le modèle de réseau DNN, par une pluralité de nouvelles couches.
 
3. Procédé selon la revendication 2, dans lequel au moins une couche de la pluralité de nouvelles couches comprend une couche non linéaire.
 
4. Procédé selon la revendication 1, dans lequel l'étape d'application, par le biais du dispositif informatique, d'une approche de décomposition à une matrice d'origine dans le modèle de réseau DNN, consiste à appliquer une décomposition en valeurs singulières (SVD) à la matrice d'origine dans le modèle de réseau DNN.
 
5. Système destiné à adapter et à personnaliser un modèle de réseau neuronal profond (DNN) pour une reconnaissance automatique de la parole (ASR), comprenant :

une mémoire destinée à stocker un code de programme exécutable ; et

un processeur, fonctionnellement couplé à la mémoire, le processeur répondant aux instructions exécutables par ordinateur contenues dans le code de programme et étant exploitable de manière à :

recevoir (405) au moins un énoncé comprenant une pluralité de caractéristiques de parole pour un ou plusieurs locuteurs en provenance d'une ou plusieurs tâches de reconnaissance ASR ;

déterminer (410) un modèle de réseau DNN adapté à partir du modèle de réseau DNN, le modèle de réseau DNN comprenant une pluralité de matrices non adaptées et le modèle de réseau DNN adapté comprenant une pluralité de matrices adaptées ;

calculer (415) une différence entre la pluralité de matrices adaptées et la pluralité de matrices non adaptées en vue de déterminer une pluralité de matrices delta ;

appliquer (420) une approche de décomposition à chaque matrice de la pluralité de matrices delta ;

convertir chaque matrice de la pluralité de matrices delta en un sous-ensemble de petites matrices ; et

stocker (430) le sous-ensemble de petites matrices, le sous-ensemble de petites matrices comprenant un faible pourcentage d'une pluralité de paramètres dans le modèle de réseau DNN.


 
6. Système selon la revendication 5, dans lequel le processeur, dans le cadre de l'application d'une approche de décomposition à chaque matrice de la pluralité de matrices delta, est exploitable de manière à appliquer une décomposition en valeurs singulières (SVD) à chaque matrice de la pluralité de matrices delta.
 
7. Système selon la revendication 5, dans lequel le processeur, dans le cadre de la conversion de chaque matrice de la pluralité de matrices delta en un sous-ensemble de petites matrices, est exploitable de manière à convertir le produit de deux matrices de rang inférieur.
 
8. Support de stockage lisible par ordinateur stockant des instructions exécutables par ordinateur qui, lorsqu'elles sont exécutées par un ordinateur, amènent l'ordinateur à mettre en oeuvre un procédé d'adaptation et de personnalisation d'un modèle de réseau neuronal profond (DNN) pour une reconnaissance automatique de la parole (ASR), le procédé comprenant les étapes ci-dessous consistant à :

recevoir (305) au moins un énoncé comprenant une pluralité de caractéristiques de parole pour un ou plusieurs locuteurs à partir d'une ou plusieurs tâches de reconnaissance ASR ;

appliquer (310) une approche de décomposition à une matrice d'origine dans le modèle de réseau DNN ;

en réponse à l'application de l'approche de décomposition, convertir (315) la matrice d'origine en une pluralité de nouvelles matrices, chaque matrice de la pluralité de nouvelles matrices étant plus petite que la matrice d'origine ;

ajouter (320) une matrice carrée, à la pluralité de nouvelles matrices, la matrice carrée stockant des paramètres spécifiques au locuteur ; et

adapter (325) le modèle de réseau DNN, uniquement en mettant à jour la matrice carrée, le modèle de réseau DNN adapté comprenant une réduction d'un nombre de paramètres dans le modèle de réseau DNN.


 
9. Support de stockage lisible par ordinateur selon la revendication 8, comprenant en outre le remplacement d'une couche d'origine dans le modèle de réseau DNN par une pluralité de nouvelles couches.
 
10. Support de stockage lisible par ordinateur selon la revendication 9, dans lequel au moins une couche de la pluralité de nouvelles couches comprend une couche non linéaire, la couche non linéaire comprenant une fonction non linéaire.
 




Drawing



























REFERENCES CITED IN THE DESCRIPTION



This list of references cited by the applicant is for the reader's convenience only. It does not form part of the European patent document. Even though great care has been taken in compiling the references, errors or omissions cannot be excluded and the EPO disclaims all liability in this regard.

Non-patent literature cited in the description