Apparatus and method for recognizing speech

Patent metadata

Field	Value
Publication number	US-2017206894-A1
Application number	US-201615187581-A
Country	US
Kind code	A1
Filing date	Jun 20, 2016
Priority date	Jan 18, 2016
Publication date	Jul 20, 2017
Grant date	—

Abstract

A speech recognition apparatus based on a deep-neural-network (DNN) sound model includes a memory and a processor. As the processor executes a program stored in the memory, the processor generates sound-model state sets corresponding to a plurality of pieces of set training speech data included in multi-set training speech data, generates a multi-set state cluster from the sound-model state sets, and sets the multi-set training speech data as an input node and the multi-set state cluster as output nodes so as to learn a DNN structured parameter.

First claim

What is claimed is:

1 . A speech recognition apparatus based on a deep-neural-network (DNN) sound model, comprising:
a memory; and a processor configured to execute a program stored in the memory, wherein, as the program is executed, the processor generates sound-model state sets corresponding to a plurality of pieces of set training speech data included in multi-set training speech data, generates a multi-set state cluster from the sound-model state sets, and sets the multi-set training speech data as an input node and the multi-set state cluster as an output node so as to learn a DNN structured parameter, and when a user's speech and characteristic information thereof are received via a user interface, the processor recognizes the user's speech on the basis of the learned DNN structured parameter by setting a sound-model state set corresponding to the characteristic information of the user's speech as an output node.

2 . The speech recognition apparatus of claim 1 , wherein the processor generates multi-state sets by collecting the sound-model state sets, and generates the multi-set state cluster by clustering the multi-state sets.

3 . The speech recognition apparatus of claim 2 , wherein the processor calculates state log likelihoods of each state of the sound-model state sets, and generates the multi-set state cluster by merging similar state clusters on the basis of the state log likelihoods and state tying information of the sound-model state sets.

4 . The speech recognition apparatus of claim 3 , wherein the processor calculates a state log likelihood corresponding to a result of merging states of two random sound-model state sets included in the multi-state sets, and merges the two random sound-model state sets when a difference between a sum of the state log likelihoods of the two random sound-model state sets and the state log likelihood corresponding to the result of merging the two random sound-model state sets is equal to or less than a predetermined threshold.

5 . The speech recognition apparatus of claim 3 , wherein the processor merges two random sound-model state sets included in the multi-state sets when logical tri-phone sets corresponding to the two random sound-model state sets are the same.

6 . The speech recognition apparatus of claim 3 , wherein the processor merges two random sound-model state sets included in the multi-state sets when logical tri-phone sets of the two random sound-model state sets are mutually inclusive and no logical tri-phone set has a relation including another sound-model state set.

7 . The speech recognition apparatus of claim 4 , wherein each of the sound-model state sets configures an independent state space on the multi-set state cluster, and the result of merging the two random sound-model state sets shares the independent state space.

8 . The speech recognition apparatus of claim 5 , wherein each of the sound-model state sets configures an independent state space on the multi-set state cluster, and the result of merging the two random sound-model state sets shares the independent state space.

9 . The speech recognition apparatus of claim 1 , wherein the processor generates state-level alignment information regarding each of the sound-model state sets, and sets the multi-set training speech data including the state-level alignment information as an input node.

10 . The speech recognition apparatus of claim 1 , wherein the processor sets the plurality of pieces of set training speech data included in the multi-set training speech data as input nodes, and sets the sound-model state sets included in the multi-set state cluster and corresponding to the plurality of pieces of set training speech data as output nodes.

11 . The speech recognition apparatus of claim 1 , wherein the plurality of pieces of set training speech data comprise different acoustic-statistical characteristics.

12 . The speech recognition apparatus of claim 11 , wherein the different acoustic-statistical characteristics comprise acoustic-statistical characteristics corresponding to speakers of different native languages.

13 . A speech recognition method based on a deep-neural-network (DNN) sound model, comprising:
generating sound-model state sets corresponding to a plurality of pieces of set training speech data included in multi-set training speech data; generating a multi-set state cluster from the sound-model state sets; learning a DNN structured parameter by setting the multi-set training speech data as an input node and the multi-set state cluster as an output node; receiving a user's speech and characteristic information thereof via a user interface; and recognizing the user's speech on the basis of the learned DNN structured parameter by setting a sound-model state set corresponding to the characteristic information of the user's speech as an output node.

14 . The speech recognition method of claim 13 , wherein the generating of the multi-set state cluster comprises:
generating multi-state sets by collecting the sound-model state sets; and generating the multi-set state cluster by clustering the multi-state sets.

15 . The speech recognition method of claim 14 , wherein the generating of the multi-set state cluster by clustering the multi-state sets comprises:
calculating state log likelihoods of each state of the sound-model state sets; and merging similar state clusters on the basis of the state log likelihoods and state tying information of the sound-model state sets.

16 . The speech recognition method of claim 15 , wherein the merging of the similar state clusters comprises:
calculating a state log likelihood corresponding to a result of merging states of two random sound-model state sets included in the multi-state sets; and merging the two random sound-model state sets when a difference between a sum of the state log likelihoods of the two random sound-model state sets and the state log likelihood corresponding to the result of merging the two random sound-model state sets is equal to or less than a predetermined threshold.

17 . The speech recognition method of claim 15 , wherein the merging of the similar state clusters comprises merging two random sound-model state sets included in the multi-state sets when logical tri-phone sets corresponding to the two random sound-model state sets are the same.

18 . The speech recognition method of claim 15 , wherein the merging of the similar state clusters comprises merging two random sound-model state sets included in the multi-state sets when logical tri-phone sets of the two random sound-model state sets are mutually inclusive and no logical tri-phone set has a relation including another sound-model state set.

19 . The speech recognition method of claim 13 , further comprising generating state-level alignment information regarding each of the sound-model state sets, and
wherein the learning of the DNN structured parameter comprises setting the multi-set training speech data including the state-level alignment information as an input node.

20 . The speech recognition method of claim 13 , wherein the learning of the DNN structured parameter comprises setting the plurality of pieces of set training speech data included in the multi-set training speech data as input nodes, and setting the sound-model state sets included in the multi-set state cluster and corresponding to the plurality of pieces of set training speech data as output nodes.

Assignees

• Electronics & Telecommunications Res Inst

Inventors

• Kang Byung Ok
• Park Jeon Gue
• Song Hwa Jeon
• Lee Yun Keun
• Chung Eui Sok

Classifications

• G10L15/063

  Training · CPC title

• G10L2015/022

  Demisyllables, biphones or triphones being the recognition units · CPC title

• G10L15/07

  to the speaker · CPC title

• G10L2015/0636

  Threshold criteria for the updating · CPC title

• G10L15/16 Primary

  using artificial neural networks · CPC title