Speech recognition device and the operation method thereof

• US 9,514,751 B2
• Filed: 03/25/2014
• Issued: 12/06/2016
• Est. Priority Date: 09/17/2013
• Status: Active Grant

• Current Assignee: Electronics and Telecommunications Research Institute
• Original Assignee: Electronics and Telecommunications Research Institute
• Inventors: Kim, Dong-Hyun
• Primary Examiner(s): Desir, Pierre-Louis
• Assistant Examiner(s): KIM, JONATHAN C
• Application Number:
  US14/224,427
• Publication Number:
  US 20150081288A1
• Time in Patent Office: 987 Days
• Field of Search: None
• US Class Current: 1/1
• CPC Class Codes: G10L 15/183 using context dependencies, e.g. language models
• CPC Class Codes: G10L 15/187 Phonemic context, e.g. pronunciation rules, phonotactical constraints or phoneme n-grams
• CPC Class Codes: G10L 15/30 Distributed recognition, e.g. in client-server systems, for mobile phones or network applications
• CPC Class Codes: G10L 2015/228 of application context

Abstract

Described herein is a speech recognition device comprising: a communication module receiving speech data corresponding to speech input from a speech recognition terminal and multi-sensor data corresponding to input environment of the speech; a model selection module selecting a language and acoustic model corresponding to the multi-sensor data among a plurality of language and acoustic models classified according to the speech input environment on the basis of previous multi-sensor data; and a speech recognition module controlling the communication module to apply a feature vector extracted from the speech data to the language and acoustic model and transmit speech recognition result for the speech data to the speech recognition terminal.

Claims

1. A speech recognition device comprising:

at least one hardware processor configured to;

receive, from a speech recognition terminal, speech data corresponding to a speech input by a speaking person and multi-sensor data corresponding to an environment in which the speech is input by the speaking person, the multi-sensor data being useable as additional information to the speech input for performing speech recognition and the multi-sensor data including an image of the speaking person and estimated location and position of the speech recognition terminal to the speaking person while the speech is input;

select a language model from a plurality of language models for the speech input, the language model being selected as representing a correspondence between a plurality of data among the multi-sensor data including the image of the speaking person of the speech input, the environment in which the speech is input by the speaking person, and the estimated location and position of the speech recognition terminal to the speaking person and previous multi-sensor data including a plurality of data among previous images of speaking persons and corresponding environments in which previous speeches are input;

select an acoustic model from among a plurality of acoustic models for the speech input, the acoustic model being selected as representing a correspondence between a plurality of data among the multi-sensor data including the image of the speaking person of the speech input, the environment in which the speech is input by the speaking person, the estimated location and position of the speech recognition terminal to the speaking person, and an estimated signal to noise ratio (SNR) for the speech data and the previous multi-sensor data including the plurality of data among previous images of speaking persons and the corresponding environments in which previous speeches are input; and

control the speech recognition of the speech input to be performed according to the selected language model and the selected acoustic model which varies in consideration of the plurality of data among the multi-sensor data obtained while the speech is input through application of a feature vector extracted from the speech data to the selected language model and the selected acoustic model, and transmit a result of the speech recognition of the speech data to the speech recognition terminal,

wherein the estimated SNR for the speech varies according to a relationship determined between the speech input and proximity of a distance between the speech recognition terminal and the speaking person obtained through the estimated location and position of the speech recognition terminal to the speaking person while the speech is being input.

2. The speech recognition device of claim 1, wherein

the estimated location and position of the speech recognition terminal to the speaking person is based on a basis of location data through one or any combination of a network-based method, a GPS method, a satellite signal-based method, and a WiFi signal-based method;

and

the at least one hardware processor is further configured to;

estimate an age of the speaking person based on the image of the speaking person for the selecting of the language model and the selecting of the acoustic model;

wherein proximity data for the distance is generated by a proximity sensor and the estimated SNR for the speech data is estimated based on the proximity data generated by the proximity sensor.

3. The speech recognition device of claim 1, further comprising:

a database storing the plurality of language models and the plurality of acoustic models classified according to the previous multi-sensor data including the plurality of data among the previous images of speaking persons and the corresponding environments in which the previous speeches are input.

4. The speech recognition device of claim 1, wherein

the feature vector is extracted from the speech data during a preprocessing operation of the speech recognition for application to the selected language model and the selected acoustic model.

5. The speech recognition device of claim 4, wherein the extracted feature vector is based on classifying of the speech data into frame units and eliminating frame-basis noise components.

6. A method of operating a speech recognition device via at least one hardware processor, the method comprising:

receiving, from a speech recognition terminal, speech data corresponding to a speech input by a speaking person and multi-sensor data corresponding to an environment in which the speech is input by the speaking person, the multi-sensor data being useable as additional information to the speech input for performing speech recognition and the multi-sensor data including an image of the speaking person of the speech and estimated location and position of the speech recognition terminal to the speaking person while the speech is input;

selecting a language model from a plurality of language models for the speech input, the language model being selected as representing a correspondence between a plurality of data among the multi-sensor data including the image of the speaking person of the speech input, the environment in which the speech is input by the speaking person and the estimated location and position of the speech recognition terminal to the speaking person and previous multi-sensor data including a plurality of data among previous images of speaking persons and corresponding environments in which previous speeches are input;

selecting an acoustic model from among a plurality of acoustic models for the speech input, the acoustic model being selected as representing a correspondence between a plurality of data among the multi-sensor data including the image of the speaking person of the speech input, the environment in which the speech is input by the speaking person and the estimated location and position of the speech recognition terminal to the speaking person, and an estimated signal to noise ratio (SNR) for the speech data, the previous multi-sensor data including the plurality of data among previous images of speaking persons and the corresponding environments in which previous speeches are input;

controlling the speech recognition of the speech input to be performed according to the selected language model and the selected acoustic model which varies in consideration of the plurality of data among the multi-sensor data obtained while the speech is input through application of a feature vector extracted from the speech data to the selected language model and the selected acoustic model; and

transmitting a result of the speech recognition of the speech data to the speech recognition terminal,

wherein the estimated SNR for the speech varies according to a relationship determined between the speech input and proximity of a distance between the speech recognition terminal and the speaking person obtained through the estimated location and position of the speech recognition terminal to the speaking person while the speech is being input.

7. The method of operating the speech recognition device of claim 6, wherein

the estimated location and position of the speech recognition terminal to the speaking person is based on location data through one or any combination of a network-based method, a satellite signal-based method, a GPS method, and a WiFi signal-based method;

the selecting of the language model and the selecting of the acoustic model further comprise estimating an age of the speaking person based on the image of the speaking person; and

wherein proximity data for the distance is generated by a proximity sensor and the estimated SNR for the speech data is estimated based on the proximity data generated by the proximity sensor.

8. The method of operating the speech recognition device of claim 6, wherein

the feature vector is extracted from the speech data during a preprocessing operation of the speech recognition for application to the selected language model and the selected acoustic model.

9. The method of operating the speech recognition device of claim 6, wherein the extracted vector is based on classifying of the speech data into frame units and estimating frame-basis noise components.