RESTRUCTURING DEEP NEURAL NETWORK ACOUSTIC MODELS

- Microsoft

A Deep Neural Network (DNN) model used in an Automatic Speech Recognition (ASR) system is restructured. A restructured DNN model may include fewer parameters compared to the original DNN model. The restructured DNN model may include a monophone state output layer in addition to the senone output layer of the original DNN model. Singular value decomposition (SVD) can be applied to one or more weight matrices of the DNN model to reduce the size of the DNN Model. The output layer of the DNN model may be restructured to include monophone states in addition to the senones (tied triphone states) which are included in the original DNN model. When the monophone states are included in the restructured DNN model, the posteriors of monophone states are used to select a small part of senones to be evaluated.

Skip to: Description  ·  Claims  · Patent History  ·  Patent History
Description
BACKGROUND

There are many applications using speech recognition including applications for searching, command and control, spoken dialog systems, natural language understanding systems, and the like. Recently, Deep Neural Network (DNN) models being used in speech recognition applications have shown significant accuracy gains in many large vocabulary continuous speech recognition (LVCSR) tasks. However, DNN models use many more parameters and have a higher computation cost as compared to traditional systems (e.g. systems using a Gaussian Mixture Model (GMM)), which limits its deployment in many scenarios.

SUMMARY

This Summary is provided to introduce a selection of concepts in a simplified form that are further described below in the Detailed Description. This Summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used as an aid in determining the scope of the claimed subject matter.

A Deep Neural Network (DNN) model used in context-dependent DNN Hidden-Markov-Model (CD-DNN-HMM) framework is restructured. The restructured DNN model may include a fewer number of parameters as compared to the original DNN model. Generally, weight matrices of the DNN model are reduced in size by reducing the sparseness of the matrices within the DNN model. For example, singular value decomposition (SVD) can be applied to one or more weight matrices of the DNN model to reduce the size of the DNN Model. The restructured DNN model may be tuned (e.g. using a back-propagation method) in an attempt to improve the accuracy of the restructured model.

The output layer of the DNN model may be restructured to include monophone states in addition to the senones (tied triphone states) which are included in the original DNN model. When the monophone states are included in the restructured DNN model, we use the posteriors of monophone states to select a small part of senones to be evaluated, which reduce the number of computations performed at the output layer of the restructured DNN model.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 shows a system for restructuring a deep neural network (DNN) model used in a speech recognition system;

FIG. 2 shows a structure of a Deep Neural Network (DNN) and a restructured DNN output layer and restructured weight matrices;

FIG. 3 depicts decomposing a weight matrix of a DNN model into two matrices with smaller dimensions;

FIG. 4 illustrates a distribution of singular values for an exemplary 2048×2048 weight matrix in a 5-hidden-layer DNN;

FIG. 5 shows an output layer including monophone states with other layers of the original DNN model and with a restructured DNN model;

FIG. 6 shows a process for restructuring a DNN model;

FIG. 7 illustrates a process for using a restructured DNN model;

FIG. 8 illustrates an exemplary online system that may be used with a restructured DNN model; and

FIGS. 9, 10A, 10B, 11 and the associated descriptions provide a discussion of a variety of operating environments in which embodiments of the invention may be practiced.

DETAILED DESCRIPTION

Referring now to the drawings, in which like numerals represent like elements, various embodiment will be described elements, various embodiment will be described.

FIG. 1 shows a system for restructuring a deep neural network (DNN) model used in a speech recognition system.

As illustrated, system 100 includes model manager 26, training data 120, deep neural network (DNN) model 130, restructured DNN model 140, recognition engine 150, results 160, application 110 (e.g. a speech related application) and touch screen input device 115.

The use of deep neural network (DNN) models has shown accuracy improvements in many large vocabulary continuous speech recognition (LVCSR) tasks. DNN models that have not been restructured, however, use many more parameters than speech recognition systems that do not use a DNN model. Generally, processing and memory usage when using a DNN model within ASR are much larger as compared to using other models (e.g. a Gaussian Mixture Model (GMM)).

Model manager 26 accesses a Deep Neural Network (DNN) model (e.g. DNN model 130) for use in an Automatic Speech Recognition (ASR) system and creates restructured DNN model 140. DNN model 130 may be restructured to reduce the size of the model. For example, one or more weight matrices of DNN model 130 may be reduced in size by reducing the sparseness of the weight matrices. For example, model manager 26 may apply singular value decomposition (SVD) to weight matrices in DNN model 130 to reduce the size of DNN Model 130.

After restructuring DNN model 130, model manager 26 may tune restructured DNN model 140. For example, restructured DNN model 140 may be tuned by model manager 26 using a back-propagation method that is directed at improving the accuracy of the restructured DNN model 140.

A DNN model (e.g. DNN model 130 or restructured DNN model 140) may be restructured by model manager 26 to change the output layer of the DNN model to include monophone states in addition to the senones. The number of monophone states is much smaller (e.g. about 100) as compared to the number of senones (e.g. 6000). Generally, the largest number of parameters for the weight matrices of a DNN model is found in the output layer (e.g. more than 10 million parameters calculated between the senone output level and the last hidden layer of the DNN model).

Model manager 26 may be used in training a DNN model, such as DNN model 130 or restructured DNN model 140, to use monophone states in the output level. The number of parameters calculated between the monophone state output layer and the last hidden layer of the DNN model is much smaller (e.g. 0.2 million parameters (connections between 100 monophone states in the output layer to 2048 parameters in last hidden layer)) as compared to the number of parameters for the senone output layer. For example, there may be 12 million connections between the 6000 senones connected to the 2048 parameters in the last hidden layer of a DNN model. The monophone state output layer and senone output layer may be combined by model manager 26 to share the input layer and hidden layers of the DNN model.

Evaluation of a speech input (e.g. received from touch screen input device 115) using model manager 26 using either restructured DNN model 140 or DNN model 130 is performed in the same manner until the last hidden output layer in the DNN model is reached. When using the restructured DNN model 140 in evaluating the speech input, the posteriors of monophone states are calculated using the monophone state output layer by model manager 26 to select the monophones that are likely to be correct (e.g. one or more monophones having a high probability score). Generally, using Bayesian probability theory, the posteriors of the monophone states are calculated by determining the conditional probability of a monophone state taking prior information about the monophone states into account. Model manager 26 uses the selected monophone states to determine the related senones to be evaluated. Instead of calculating each of the parameters associated with the senone output layer, calculations are performed by model manager 26 for the senones that have the phone states that were selected by model manager 26 when using the monophone state output layer. The selected phone state is the central phone state for the related senone. For example, assume that the number of parameters is 12 million for a senone output layer and ten monophone states are selected by model manager 26. Instead of performing 12 million calculations in the senone output layer, 1.2 million calculations are performed by model manager 26.

The restructured DNN model 140 may be included in an ASR system. The restructured model may or may not include the monophone state output layer and may or may not include weight matrices reduced in size. For example, a restructured DNN model may include weight matrices reduced in size, another restructured DNN model may include the monophone state output layer, and another restructured DNN model may include both the reduced weight matrices and the monophone state output layer. Restructured DNN model 140 may be included in an ASR system (e.g. a Context-Dependent (CD)-DNN-Hidden-Markov-Model (HMM) ASR system).

In order to facilitate communication with the model manager 26, one or more callback routines, may be implemented. According to one embodiment, application 110 is a multimodal application that is configured to receive speech input (e.g. utterances) and to perform Automatic Speech Recognition (ASR) on the utterance. Application 110 may also receive input from a touch-sensitive input device 115 and/or other input devices. For example, voice input, keyboard input (e.g. a physical keyboard and/or SIP), video based input, and the like. Application program 110 may also provide multimodal output (e.g., for purposes of speech, graphics, vibrations, sounds, . . . ). Model manager 26 may provide information to/from application 110 in response to user input (e.g. speech/gesture). For example, a user may say a phrase to be recognized by application 110 (e.g. performing a search, selecting content, buying an item, identifying a product, . . . ). System 100 as illustrated comprises a touch screen input device 115 that detects when a touch input has been received (e.g. a finger touching or nearly teaching the touch screen). Any type of touch screen may be utilized that detects a user's touch input. More details are provided below.

FIG. 2 shows a structure of a Deep Neural Network (DNN) and a restructured DNN output layer and restructured weight matrices.

As illustrated, DNN model 200 includes a bottom input layer 210, hidden layers 220, and a top output layer 230. According to an embodiment, DNN model 200 includes a monophone state output layer 235 in addition to a senone output layer 230.

Generally, a DNN model is fully connected between adjacent layers (e.g. each neuron in a bottom layer is connected to each neuron in a top layer). DNNs used in ASR systems typically include 5-8 hidden layers and each layer includes thousands of neurons resulting in millions of calculations when using the DNN model for evaluating speech input. Using the same amount of training data, a DNN model typically has 2 to 10 times more parameters than traditional systems (e.g. CD-Gaussian Mixture Model (GMM)-HMMs).

As illustrated, DNN model 200 includes input layer 210 that is coupled to first hidden layer 212 using weight matrix 204 and a last hidden layer 214 that is coupled to output layer 230 using weight matrix 201. Adjacent layers between the first and last hidden layer are coupled using weight matrices (e.g. weight matrix 202, weight matrix 203, . . . ). Before DNN 200 is restructured, each weight matrix may include millions of parameters. These weight matrices are typically sparse and a large portion of the weight parameters in each weight matrix are very small (e.g. zero, or near zero in value). Generally, these very small weight parameters have a negligible effect on the output values of each layer.

According to an embodiment, singular value decomposition (SVD) is applied to decompose the weight matrices in DNN model, and then the DNN model is restructured based on the sparseness of the original format. The restructured DNN model has a similar layout as original model but replaces a single layer with two layers.

For example, original DNN model layer 221 becomes reconstructed model section 240. As can be seen, restructured model section 240 includes two extra layers as compared to the original model layer 221. After SVD restructuring, the DNN model size is reduced. The size of the reduction of the DNN model depends on a variety of factors, such as how many weight matrices are reduced and the number of Singular Values are kept after SVD decomposition. The restructured DNN model may/may not be as accurate as the original DNN model. The restructured model may be tuned (e.g. using a back-propagation method) that is directed at improving the accuracy of the restructured model.

A DNN model may also be restructured to change the output layer to include a monophone state output layer in addition to the senone output layer. Generally, the largest number of parameters for the weight matrices of a DNN model is found in the output layer (e.g. above 10 million parameters for the senone output layer to the last hidden layer). The number of parameters related to the monophone state output layer is much smaller as compared to the number of parameters for the senone output layer (e.g. 0.2 million parameters compared to 12 million parameters). The monophone state output layer and senone output layer may be combined to share the input layer and hidden layers of the DNN model.

FIG. 3 depicts decomposing a weight matrix of a DNN model into two matrices with smaller dimensions.

Weight matrix A (310) is an m×n weight matrix in a DNN model before restructuring.

Applying SVD to weight matrix A (310) produces: Am×n=Um×nΣn×nVn×nT (1) (See 320 in FIG. 3) where Σ is a diagonal matrix with A's singular values on the diagonal in the decreasing order. The m columns of U and the n columns of V are referred to as the left-singular vectors and right-singular vectors of A, respectively. Since A is a sparse matrix, a large part of A's singular values are typically very small. Setting the small values to 0 does not considerably change the values of elements in matrix A.

Element 330 of FIG. 3 shows keeping the largest K singular values of A.

Keeping the largest k singular values of A, formula (1) may be rewritten as Am×n=Um×kΣk×kVk×nT as can be seen in element 340.

This may further be represented as Am×n=Um×kWk×n where Wk×nkk×kVk×nT as can be seen in element 350 that shows matrix A decomposed into two smaller matrices U and W.

As can be seen by referring to FIG. 3, the number of parameters in changes from mn in matrix A (310) to (m+n)k. The DNN model size is reduced when k is much smaller than m and n. The value of k can be set to a pre-decided value or determined using information from the weight matrices being reduced.

FIG. 4 illustrates a distribution of singular values for an exemplary 2048×2048 weight matrix in a 5-hidden-layer DNN.

As illustrated, distribution 400 includes an x-axis showing the number of singular values, and a y-axis that shows an accumulated percentage of total singular values. Referring to FIG. 4 it can be seen that about 15% of singular values in a DNN model contribute 50% of total values, and around 40% of the singular values contribute 80% of total values. This illustration shows that setting the small values in a weight matrix to 0 does not considerably change the values of the elements in a weight matrix.

FIG. 5 shows an output layer including monophone states with other layers of the original DNN model and with a restructured DNN model.

As illustrated, DNN model 510 shows the use of an output layer including the senone output layer and the monophone state output layer where the weight matrices of the DNN model have not been reduced in size. While DNN model 510 may include more parameters then a regular DNN model, calculations are reduced at the output layer since the monophone states are used in selecting the calculations to evaluate the senones.

DNN model 520 shows the use of the monophone states in the output layer with other layers restructured in the DNN model. DNN 520 illustrates that a DNN model can be restructured by resizing one or more weight matrices of the DNN model and adding the monophone state output layer to the DNN model.

FIGS. 6-7 illustrate restructuring a DNN model and using monophone states with a DNN model. When reading the discussion of the routines presented herein, it should be appreciated that the logical operations of various embodiments are implemented (1) as a sequence of computer implemented acts or program modules running on a computing system and/or (2) as interconnected machine logic 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 and making up the embodiments described herein are referred to variously as operations, structural devices, acts or modules. These operations, structural devices, acts and modules may be implemented in software, in firmware, in special purpose digital logic, and any combination thereof. While the operations are shown in a particular order, the order of the operations may change, be performed in parallel, depending on the implementation.

FIG. 6 shows a process for restructuring a DNN model.

After a start operation, the process moves to operation 610, where a DNN model is accessed. Generally, a DNN used in and ASR systems includes an input layer, 5-8 hidden layers and a senone output layer. Each layer typically includes thousands of neurons resulting in millions of calculations when evaluating the DNN model.

Flowing to operation 620, the size of the DNN model is reduced. Generally, the size of the DNN model may be reduced depends on the sparseness of one or more weight matrices between the different layers of the DNN model.

At operation 622, the weight matrices to reduce are determined. Zero or more weight matrices in the DNN may be reduced. The determination to reduce a weight matrix may be based on the size of the weight matrix. Some weight matrices in the DNN model are larger in size as compared to other weight matrices. For example, reducing the size of the weight matrix below the output layer may reduce the size of the DNN the most since that layer is typically the largest in a DNN model. Similarly, the weight matrix immediately above the input layer may not be reduced since the number of parameters in that layer is the smallest in the DNN model. According to an embodiment, each layer in the DNN model except for the input layer is reduced.

At operation 624, singular value decomposition (SVD) is applied to each weight matrix that is determined to be reduced. As discussed herein, SVD decomposes a weight matrix into different matrices.

At operation 626, a determination is made as to what singular values to keep. Generally, setting the smaller values in the weight matrix to zero does not considerably change the values of the elements. The determination of what values to keep may be made using different methods. For example, the number of the largest singular values to keep in a weight matrix may be predetermined (e.g. ½, ¼, ⅛, and the like). The number of singular values kept may change based on the weight matrix. For example, the distribution of the singular values in the weight matrices may be used to determine how many singular values to keep (e.g. keep the same proportion in each weight matrix).

Transitioning to operation 630, the DNN model is restructured. The DNN model is restructured to include two layers for each original layer.

Moving to operation 640, the restructured model may be tuned. For example, a back-propagation method that is directed at improving the accuracy of the restructured model.

Transitioning to operation 650, a monophone state output layer may be included as part of a DNN model. The number of connections between the senone output layer and the last hidden layer used in a typical DNN model is very large. For example, there may be approximately 6000 senones connected to approximately 2000 parameters making about 12 million possible connections. The number of monophone states (about 100), however, is much smaller as compared to the number of senones (about 6000) and includes about 0.2 million possible connections. According to an embodiment, the senone output layer and the monophone state output layer share the same levels within the DNN. Recognition is initially performed using the monophone state output level to assist in determining the senones to calculate thereby reducing a number of calculations performed at the output level of the restructured DNN model.

The process then moves to an end operation and returns to processing other actions.

FIG. 7 illustrates a process for using a restructured DNN model.

After a start operation, the process moves to operation 710, where a restructured DNN model is accessed. According to an embodiment, the restructured DNN model is part of a CD-DNN-HMM framework. The restructured model may be restructured in different ways. For example, zero or more weight matrices may be reduced in size and a monophone state output layer may or may not be included in the restructured DNN model. For purposes of explanation, the restructured DNN model includes one or more weight matrices reduced in size and includes a monophone state output layer.

Transitioning to operation 720, an utterance is received. The utterance may be a live utterance and/or a recorded utterance.

Flowing to operations 730, recognition of the utterance is performed using the restructured model. Use of a restructured DNN model that does not include a monophone state output layer occurs in a same manner as a traditional DNN model used in ASR.

Moving to decision operation 740, a determination is made as to whether the restructured DNN model includes a monophone state output layer. When the restructured DNN model includes a monophone state output layer, the process flows to operation 750. When the restructured DNN model does not include a monophone state output layer, the process flows to operation 780 to calculate all the senones.

At operation 750, one or more monophones are selected based on recognition results using the monophone state output layer of the restructured DNN model. The monophones may be selected using different methods. For example, a predefined number of monophones that are most likely to be correct (e.g. the top 5, 10 monophones) may be selected. According to an embodiment, the monophones that have a recognition score within a predetermined difference of the recognition score of the top monophone are selected. The number of monophones selected may be changed dynamically based on different conditions (e.g. the presence or absence of noise in the environment, quality of language model, . . . ).

Flowing to operation 760, the selected monophones are used to determine what parameters in the senone output layer to calculate. The parameters in the senone output layer that are selected include the monophones that are selected. For example, if the top 10 monophones are selected, then instead of performing all 12 million calculations in the senone output layer, 1.2 million calculations are performed.

Moving to operation 760, the recognition results using the senone output layer are received.

The process then moves to an end operation and returns to processing other actions.

FIG. 8 illustrates an exemplary online system that may be used with a restructured DNN model. As illustrated, system 1000 includes service 1010, data store 1045, restructured DNN model 1046, touch screen input device 1050 (e.g. a slate), smart phone 1030 and display device 1080.

As illustrated, service 1010 is a cloud based and/or enterprise based service that may be configured to provide services, such as multimodal services related to various applications (e.g. searching, games, browsing, locating, productivity services (e.g. spreadsheets, documents, presentations, charts, messages, and the like)). The service may be interacted with using different types of input/output. For example, a user may use speech input, touch input, hardware based input, and the like. The service may provide speech output that combines pre-recorded speech and synthesized speech. Functionality of one or more of the services/applications provided by service 1010 may also be configured as a client/server based application.

As illustrated, service 1010 is a multi-tenant service that provides resources 1015 and services to any number of tenants (e.g. Tenants 1−N). Multi-tenant service 1010 is a cloud based service that provides resources/services 1015 to tenants subscribed to the service and maintains each tenant's data separately and protected from other tenant data.

System 1000 as illustrated comprises a touch screen input device 1050 (e.g. a slate/tablet device) and smart phone 1030 that detects when a touch input has been received (e.g. a finger touching or nearly touching the touch screen). Any type of touch screen may be utilized that detects a user's touch input. For example, the touch screen may include one or more layers of capacitive material that detects the touch input. Other sensors may be used in addition to or in place of the capacitive material. For example, Infrared (IR) sensors may be used. According to an embodiment, the touch screen is configured to detect objects that in contact with or above a touchable surface. Although the term “above” is used in this description, it should be understood that the orientation of the touch panel system is irrelevant. The term “above” is intended to be applicable to all such orientations. The touch screen may be configured to determine locations of where touch input is received (e.g. a starting point, intermediate points and an ending point). Actual contact between the touchable surface and the object may be detected by any suitable means, including, for example, by a vibration sensor or microphone coupled to the touch panel. A non-exhaustive list of examples for sensors to detect contact includes pressure-based mechanisms, micro-machined accelerometers, piezoelectric devices, capacitive sensors, resistive sensors, inductive sensors, laser vibrometers, and LED vibrometers.

According to an embodiment, smart phone 1030, touch screen input device 1050, and display device 1080 are configured with multimodal applications and each include an application (1031, 1051, 1081) that is configured to receive speech input.

As illustrated, touch screen input device 1050, smart phone 1030, and display device 1080 shows exemplary displays 1052/1032/1082 showing the use of an application using multimodal input/output. Data may be stored on a device (e.g. smart phone 1030, touch screen input device 1050 and/or at some other location (e.g. network data store 1045). Data store 1045, or some other store, may be used to store data. The applications used by the devices may be client based applications, server based applications, cloud based applications and/or some combination. According to an embodiment, display device 1080 is a device such as a MICROSOFT XBOX coupled to a display.

Model manager 26 is configured to perform operations relating to a restructuring a DNN model as described herein. For example, a DNN model may be restructured to change a sparseness of one or more weight matrices of the DNN model. A monophone state output layer may also be added to the senone output layer of a DNN model. While manager 26 is shown within service 1010, the functionality of the manager may be included in other locations (e.g. on smart phone 1030 and/or touch screen input device 1050 and/or device 1080). The restructured DNN model may be included on smart phone 1030, 1050, 1080 and used for ASR.

The embodiments and functionalities described herein may operate via a multitude of computing systems including, without limitation, desktop computer systems, wired and wireless computing systems, mobile computing systems (e.g., mobile telephones, netbooks, tablet or slate type computers, notebook computers, and laptop computers), hand-held devices, multiprocessor systems, microprocessor-based or programmable consumer electronics, minicomputers, and mainframe computers.

In addition, the embodiments and functionalities described herein may operate over distributed systems (e.g., cloud-based computing systems), where application functionality, memory, data storage and retrieval and various processing functions may be operated remotely from each other over a distributed computing network, such as the Internet or an intranet. User interfaces and information of various types may be displayed via on-board computing device displays or via remote display units associated with one or more computing devices. For example user interfaces and information of various types may be displayed and interacted with on a wall surface onto which user interfaces and information of various types are projected. Interaction with the multitude of computing systems with which embodiments of the invention may be practiced include, keystroke entry, touch screen entry, voice or other audio entry, gesture entry where an associated computing device is equipped with detection (e.g., camera) functionality for capturing and interpreting user gestures for controlling the functionality of the computing device, and the like.

FIGS. 9-11 and the associated descriptions provide a discussion of a variety of operating environments in which embodiments of the invention may be practiced. However, the devices and systems illustrated and discussed with respect to FIGS. 9-11 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 embodiments of the invention, described herein.

FIG. 9 is a block diagram illustrating physical components (i.e., hardware) of a computing device 1100 with which embodiments of the invention may be practiced. The computing device components described below may be suitable for the computing devices described above. In a basic configuration, the computing device 1100 may include at least one processing unit 1102 and a system memory 1104. Depending on the configuration and type of computing device, the system memory 1104 may comprise, but is not limited to, volatile storage (e.g., random access memory), non-volatile storage (e.g., read-only memory), flash memory, or any combination of such memories. The system memory 1104 may include an operating system 1105 and one or more program modules 1106 suitable for running software applications 1120 such as the model manager 26. The operating system 1105, for example, may be suitable for controlling the operation of the computing device 1100. Furthermore, embodiments of the invention may be practiced in conjunction with a graphics library, other operating systems, or any other application program and is not limited to any particular application or system. This basic configuration is illustrated in FIG. 9 by those components within a dashed line 1108. The computing device 1100 may have additional features or functionality. For example, the computing device 1100 may also include additional data storage devices (removable and/or non-removable) such as, for example, magnetic disks, optical disks, or tape. Such additional storage is illustrated in FIG. 9 by a removable storage device 1109 and a non-removable storage device 1110.

As stated above, a number of program modules and data files may be stored in the system memory 1104. While executing on the processing unit 1102, the program modules 1106 (e.g., the model manager 26) may perform processes including, but not limited to, one or more of the stages of the methods and processes illustrated in the figures. Other program modules that may be used in accordance with embodiments of the present invention may include electronic mail and contacts applications, word processing applications, spreadsheet applications, database applications, slide presentation applications, drawing or computer-aided application programs, etc.

Furthermore, embodiments of the invention 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, embodiments of the invention may be practiced via a system-on-a-chip (SOC) where each or many of the components illustrated in FIG. 9 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, with respect to the model manager 26 may be operated via application-specific logic integrated with other components of the computing device 1100 on the single integrated circuit (chip). Embodiments of the invention 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, embodiments of the invention may be practiced within a general purpose computer or in any other circuits or systems.

The computing device 1100 may also have one or more input device(s) 1112 such as a keyboard, a mouse, a pen, a sound input device, a touch input device, etc. The output device(s) 1114 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 1100 may include one or more communication connections 1116 allowing communications with other computing devices 1118. Examples of suitable communication connections 1116 include, but are not limited to, RF transmitter, receiver, and/or transceiver circuitry; universal serial bus (USB), parallel, and/or serial ports.

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 1104, the removable storage device 1109, and the non-removable storage device 1110 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 1100. Any such computer storage media may be part of the computing device 1100. Computer storage media does not include a carrier wave or other propagated or modulated data signal.

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.

FIGS. 10A and 10B illustrate a mobile computing device 1200, for example, a mobile telephone, a smart phone, a tablet personal computer, a laptop computer, and the like, with which embodiments of the invention may be practiced. With reference to FIG. 10A, one embodiment of a mobile computing device 1200 for implementing the embodiments is illustrated. In a basic configuration, the mobile computing device 1200 is a handheld computer having both input elements and output elements. The mobile computing device 1200 typically includes a display 1205 and one or more input buttons 1210 that allow the user to enter information into the mobile computing device 1200. The display 1205 of the mobile computing device 1200 may also function as an input device (e.g., a touch screen display). If included, an optional side input element 1215 allows further user input. The side input element 1215 may be a rotary switch, a button, or any other type of manual input element. In alternative embodiments, mobile computing device 1200 may incorporate more or less input elements. For example, the display 1205 may not be a touch screen in some embodiments. In yet another alternative embodiment, the mobile computing device 1200 is a portable phone system, such as a cellular phone. The mobile computing device 1200 may also include an optional keypad 1235. Optional keypad 1235 may be a physical keypad or a “soft” keypad generated on the touch screen display. In various embodiments, the output elements include the display 1205 for showing a graphical user interface (GUI), a visual indicator 1220 (e.g., a light emitting diode), and/or an audio transducer 1225 (e.g., a speaker). In some embodiments, the mobile computing device 1200 incorporates a vibration transducer for providing the user with tactile feedback. In yet another embodiment, the mobile computing device 1200 incorporates input and/or output ports, such as an audio input (e.g., a microphone jack), an audio output (e.g., a headphone jack), and a video output (e.g., a HDMI port) for sending signals to or receiving signals from an external device.

FIG. 10B is a block diagram illustrating the architecture of one embodiment of a mobile computing device. That is, the mobile computing device 1200 can incorporate a system 1202 (i.e., an architecture) to implement some embodiments. In one embodiment, the system 1202 is implemented as a “smart phone” capable of running one or more applications (e.g., browser, e-mail, calendaring, contact managers, messaging clients, games, and media clients/players). In some embodiments, the system 1202 is integrated as a computing device, such as an integrated personal digital assistant (PDA) and wireless phone.

One or more application programs 1266 may be loaded into the memory 1262 and run on or in association with the operating system 1264. Examples of the application programs include phone dialer programs, e-mail programs, personal information management (PIM) programs, word processing programs, spreadsheet programs, Internet browser programs, messaging programs, and so forth. The system 1202 also includes a non-volatile storage area 1268 within the memory 1262. The non-volatile storage area 1268 may be used to store persistent information that should not be lost if the system 1202 is powered down. The application programs 1266 may use and store information in the non-volatile storage area 1268, such as e-mail or other messages used by an e-mail application, and the like. A synchronization application (not shown) also resides on the system 1202 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 area 1268 synchronized with corresponding information stored at the host computer. As should be appreciated, other applications may be loaded into the memory 1262 and run on the mobile computing device 1200, including the model manager 26 as described herein.

The system 1202 has a power supply 1270, which may be implemented as one or more batteries. The power supply 1270 might further include an external power source, such as an AC adapter or a powered docking cradle that supplements or recharges the batteries.

The system 1202 may also include a radio 1272 that performs the function of transmitting and receiving radio frequency communications. The radio 1272 facilitates wireless connectivity between the system 1202 and the “outside world,” via a communications carrier or service provider. Transmissions to and from the radio 1272 are conducted under control of the operating system 1264. In other words, communications received by the radio 1272 may be disseminated to the application programs 1266 via the operating system 1264, and vice versa.

The visual indicator 1220 may be used to provide visual notifications, and/or an audio interface 1274 may be used for producing audible notifications via the audio transducer 1225. In the illustrated embodiment, the visual indicator 1220 is a light emitting diode (LED) and the audio transducer 1225 is a speaker. These devices may be directly coupled to the power supply 1270 so that when activated, they remain on for a duration dictated by the notification mechanism even though the processor 1260 and other components might shut down for conserving battery power. The LED may be programmed to remain on indefinitely until the user takes action to indicate the powered-on status of the device. The audio interface 1274 is used to provide audible signals to and receive audible signals from the user. For example, in addition to being coupled to the audio transducer 1225, the audio interface 1274 may also be coupled to a microphone to receive audible input, such as to facilitate a telephone conversation. In accordance with embodiments of the present invention, the microphone may also serve as an audio sensor to facilitate control of notifications, as will be described below. The system 1202 may further include a video interface 1276 that enables an operation of an on-board camera to record still images, video stream, and the like.

A mobile computing device 1200 implementing the system 1202 may have additional features or functionality. For example, the mobile computing device 1200 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. 10B by the non-volatile storage area 1268. Mobile computing device 1200 may also include peripheral device port 1230.

Data/information generated or captured by the mobile computing device 1200 and stored via the system 1202 may be stored locally on the mobile computing device 1200, 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 1272 or via a wired connection between the mobile computing device 1200 and a separate computing device associated with the mobile computing device 1200, 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 1200 via the radio 1272 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.

FIG. 11 illustrates an embodiment of an architecture of an exemplary system, as described above. Content developed, interacted with, or edited in association with the model manager 26 may be stored in different communication channels or other storage types. For example, various documents may be stored using a directory service 1322, a web portal 1324, a mailbox service 1326, an instant messaging store 1328, or a social networking site 1330. The model manager 26 may use any of these types of systems or the like for enabling data utilization, as described herein. A server 1320 may provide the model manager 26 to clients. As one example, the server 1320 may be a web server providing the model manager 26 over the web. The server 1320 may provide the model manager 26 over the web to clients through a network 1315. By way of example, the client computing device may be implemented as the computing device 1100 and embodied in a personal computer, a tablet computing device 1310 and/or a mobile computing device 1200 (e.g., a smart phone). Any of these embodiments of the client computing device 1100, 1310, 1200 may obtain content from the store 1316.

Embodiments of the present invention, for example, are described above with reference to block diagrams and/or operational illustrations of methods, systems, and computer program products according to embodiments of the invention. The functions/acts noted in the blocks may occur out of the order as shown in any flowchart. 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.

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. Regardless of whether shown and described in combination or separately, the various features (both structural and methodological) are intended to be selectively included or omitted to produce an embodiment with a particular set of features. Having been provided with the description and illustration of the present application, one skilled in the art may envision variations, modifications, and alternate embodiments falling within the spirit of the broader aspects of the general inventive concept embodied in this application that do not depart from the broader scope of the claimed invention.

Claims

1. A method for restructuring a Deep Neural Network (DNN) model, comprising:

accessing a DNN model that includes weight matrices and layers comprising: an input layer; hidden layers; and an output layer;
reducing a sparseness of a weight matrix in the DNN model; and
restructuring the DNN model with the weight matrix reduced in sparseness.

2. The method of claim 1, wherein reducing the sparseness of the weight matrix comprises applying a Singular Value Decomposition (SVD) to the weight matrix.

3. The method of claim 1, wherein applying the SVD to the weight matrix comprises decomposing the weight matrix into two matrices having smaller dimensions as compared to a size of the dimension of the weight matrix before applying the SVD to the weight matrix.

4. The method of claim 1, wherein restructuring the DNN model with the weight matrix reduced in sparseness comprises splitting one of the layers in the DNN model into a first layer and a second layer.

5. The method of claim 1, wherein reducing the sparseness of the weight matrix in the DNN model comprises reducing the sparseness of each weight matrix in the DNN model.

6. The method of claim 1, wherein the output layer comprises a senone output layer and a monophone state output layer.

7. The method of claim 1, further comprising training the output layer of the DNN to use a monophone state.

8. The method of claim 1, further comprising tuning the restructured model using a back-propagation method.

9. A computer-readable storage medium storing computer-executable instructions that when executed using a processor perform actions, comprising:

accessing a restructured Deep Neural Network (DNN) model that includes one or more weight matrices reduced in size as compared to the corresponding one or more weight matrices in an original DNN model and layers comprising: an input layer; hidden layers; and an output layer; and
using the restructured DNN model to recognize received utterances.

10. The computer-readable storage medium of claim 9, wherein a sparseness of the one or more weight matrices are reduced in size as compared to a sparseness of the one or more weight matrices of the original DNN.

11. The computer-readable storage medium of claim 9, wherein the output layer of the restructured DNN comprises a monophone state output layer and a senone output layer.

12. The computer-readable storage medium of claim 9, further comprising using posteriors of monophone states to select senones to be evaluated to reduce the number of calculations in the senone output layer.

13. The computer-readable storage medium of claim 9, wherein the one or more weight matrices of the restructured DNN comprises two matrices having smaller dimensions as compared to a size of the dimension of the weight matrix in the original DNN model.

14. The computer-readable storage medium of claim 9, further comprising tuning the restructured DNN model using a back-propagation method.

15. A system for restructuring a Deep Neural Network (DNN) model, comprising:

a processor and memory;
an operating environment executing using the processor; and
a model manager that is configured to perform actions comprising: accessing a DNN model that includes weight matrices and layers comprising: an input layer; hidden layers; and an output layer; reducing a sparseness of weight matrices in the DNN model by removing weight parameters that are below a threshold value; and restructuring the DNN model with the weight matrix reduced in sparseness.

16. The system of claim 15, wherein reducing the sparseness of the weight matrix comprises applying a Singular Value Decomposition (SVD) to the weight matrix.

17. The system of claim 15, wherein applying the SVD to the weight matrix comprises decomposing the weight matrix into two matrices having smaller dimensions as compared to a size of the dimension of the weight matrix before applying the SVD to the weigh matrix.

18. The system of claim 15, wherein restructuring the DNN model with the weight matrix reduced in sparseness comprises splitting one of the layers in the DNN model into a first layer and a second layer.

19. The system of claim 15, wherein the output layer comprises a senone output layer and a monophone state output layer.

20. The system of claim 15, further comprising training the output layer of the DNN to use a monophone state.

Patent History
Publication number: 20170337918
Type: Application
Filed: Aug 7, 2017
Publication Date: Nov 23, 2017
Applicant: Microsoft Technology Licensing, LLC (Redmond, WA)
Inventors: Jian XUE (Redmond, WA), Emilian STOIMENOV (Bellevue, WA), Jinyu LI (Redmond, WA), Yifan GONG (Sammamish, WA)
Application Number: 15/670,819
Classifications
International Classification: G10L 15/16 (20060101); G06N 3/04 (20060101); G06N 3/08 (20060101); G10L 15/14 (20060101);