JP2011118124A - Speech recognition system and recognition method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a speech recognition system and a recognition method which is robust to noise, even when a speaker moves. <P>SOLUTION: A noise spectrum by position is stored, an acoustic model including the noise by position is stored, and a position is discriminated. A spectrum of an acoustic signal from a microphone in a non-speech section is compared with the noise spectrum by position, and when difference of a threshold or more is detected, noise of at least a part of a section exceeding the threshold is removed. Then, speech recognition is performed by the acoustic model by position. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to speech recognition, and more particularly to improving noise robustness.

  Patent Document 1: Japanese Patent Laid-Open No. 2005-208483 discloses changing a language model according to an organization to which a speaker belongs for speech recognition at a call center or the like. In this technique, a plurality of small vocabulary language models are prepared and changed according to the speaker, thereby reducing misrecognition. Patent Document 2: Japanese Patent No. 4003566 discloses that a nonlinear spectral subtraction is applied on the low frequency side and a Wiener filter or a Kalman filter is applied on the high frequency side in order to suppress noise.

  When the switching of the language model in Patent Document 1 is applied to the switching of the noise removal parameter, the noise parameter is learned over a certain period of time. However, when the moving speed of the speaker is fast, learning of noise parameters cannot follow.

JP2005-208483 Patent 4003566

An object of the present invention is to provide a speech recognition system and a recognition method that are robust against noise even when a speaker moves.
Another object of the present invention is to switch the language model appropriately in speech recognition for picking, guidance, navigation, and the like.

The speech recognition apparatus according to the present invention is a system that recognizes speech by a speech recognition unit that processes noise in an acoustic signal from a microphone by a noise removal unit and stores an acoustic model,
A noise spectrum storage unit for storing a noise spectrum for each position;
An acoustic model storage unit that stores an acoustic model for each position and including noise at the position;
Position discrimination means;
A noise comparison unit that compares the spectrum of the acoustic signal from the microphone in the non-speech interval with the noise spectrum stored in the noise spectrum storage unit for the position determined by the determination unit,
When the noise comparison unit detects a difference greater than or equal to the threshold value, at least a part of the portion exceeding the threshold value is removed by the noise removal unit with respect to the acoustic signal from the microphone in the utterance interval,
Voice recognition is performed by the voice recognition unit using an acoustic model for the position determined by the determination means.

The speech recognition method of the present invention is a method of recognizing speech in a speech recognition unit that stores noise in an acoustic signal by processing noise in a sound signal from a microphone and storing an acoustic model,
The noise spectrum for each position is stored in the storage means,
Storing an acoustic model including noise at each position in the storage means;
The position is determined by the position determination means,
The comparison means compares the spectrum of the acoustic signal from the microphone in the non-speech section with the noise spectrum stored in the noise spectrum storage unit for the position determined by the determination means, and detects a difference greater than the threshold. At the time, for the acoustic signal from the microphone in the utterance section, at least a part of the portion exceeding the threshold value is denoised by the noise removing unit,
Voice recognition is performed by the voice recognition unit using an acoustic model for the position determined by the determination means.

In the present invention, the following effects can be obtained.
(1) The speaker's position is determined through comparison of noise spectra or by a position sensor such as GPS. An acoustic model including noise is used for each position. As a result, when the change from the stored noise spectrum is small, the influence of noise can be made extremely small. Even with the same speaker, the voice changes depending on the surrounding environment. This is already incorporated into the acoustic model for each location.
(2) Instead of learning noise in response to the movement of the speaker, the speaker's position is specified and an acoustic model corresponding to the position is used. Therefore, it is possible to quickly follow the movement of the speaker.
(3) If the noise spectrum at each location differs from the current noise spectrum by more than a threshold, noise removal is performed. Since the noise removal may be performed on a small part exceeding the threshold or a part thereof, the audio signal is hardly distorted.
In this specification, the description related to the speech recognition system is also applied to the speech recognition method, and the description related to the speech recognition method is also applied to the speech recognition system.

  Preferably, the speech recognition system includes a server and a plurality of mobile terminals, and the microphone and the determination unit are provided in the terminal, a noise spectrum storage unit, an acoustic model storage unit, and a noise comparison unit are provided in the server, and a noise removal unit And a voice recognition unit are provided in either the server or the terminal.

  Preferably, the determination means is a position sensor such as a GPS. When the position of the speaker is determined by the position sensor, the switching of the acoustic model is not delayed by the movement of the speaker. Note that the position of the speaker can also be determined by comparing the noise spectrum with the actual noise, but the determination of the position may be delayed due to the movement of the speaker due to the acquisition and comparison of noise. The position sensor only needs to be able to determine the position for switching the acoustic model, and does not need to detect the position with high accuracy.

Preferably, a plurality of language models composed of a dictionary for speech recognition are stored, and the language model is switched according to the position determined by the determining means. By switching the language model according to the position of the speaker, more accurate speech recognition can be performed.

Block diagram of voice recognition system of embodiment The figure which shows the switching algorithm of the acoustic model and language model in an Example The figure which shows the example where the stored value of the noise spectrum and the actual value almost coincide Figure showing an example of noise removal when the difference between the stored value of the noise spectrum and the actual value is greater than or equal to a threshold value The figure which shows the example which switches a language model between the object for books and the object for food by the position of a speaker Diagram showing an example of switching the language model between picking and assorting depending on the speaker's position

  In the following, an optimum embodiment for carrying out the present invention will be shown. The scope of the present invention should be determined in accordance with the understanding of those skilled in the art based on the description of the scope of claims, taking into consideration the specification and well-known techniques in this field.

  1 to 6 show a speech recognition system according to an embodiment. In FIG. 1, reference numeral 2 denotes a server of a voice recognition system, which is connected to a plurality of terminals 4 by communication units 18 and 38. A noise comparison unit 6 compares the acoustic data or noise spectrum of noise received from the terminal 4 with the data of the noise spectrum storage unit 8 that stores the noise spectrum for each position, for example, for n positions. When the noise comparison unit 6 obtains a position that gives the closest noise spectrum to the actual noise, the noise comparing unit 6 determines whether or not the difference between the actual noise and the noise spectrum at that position is equal to or greater than a threshold value. When a position signal is sent from the position information acquisition device 22 of the terminal 4, it is not necessary to determine which position's noise spectrum is closest to the actual noise. When the difference between the actual noise and the stored noise spectrum is equal to or greater than the threshold, the noise removal unit 10 performs spectral subtraction or Kalman filter on the portion exceeding the threshold or a part thereof before processing by the speech recognition unit 16. Noise removal is performed using a Wiener filter.

  The language model storage unit 12 stores, for example, n language models in accordance with the movement range of the speaker wearing the terminal 4, and its action is a dictionary that converts a phoneme sequence obtained by speech recognition into a language. . By switching the language model according to the position, accurate and high-speed speech recognition is performed with a small dictionary having a small vocabulary according to the position. In the embodiment, the position, the language model, and the noise spectrum correspond to each other 1: 1: 1. However, a plurality of positions and a plurality of noise spectra may be associated with one language model. The acoustic model storage unit 14 stores an acoustic model for each position and for each speaker. However, for the sake of simplicity, an acoustic model corresponding to only the position may be used instead of each speaker and position. The acoustic model here is created using a small number of samples, and the pronunciation differs depending on the speaker. Therefore, the acoustic model is preferably set for each speaker. Even if the same speaker has different environmental noise, the same phoneme is pronounced differently, so an acoustic model corresponding to the position is used. When there is one terminal 4, it is not necessary to provide an acoustic model for each speaker, and the server 2 may be abolished by providing the terminal 4 with the function of the server 2.

  The voice recognition unit 16 recognizes the voice signal transmitted from the terminal 4 and the acoustic model used corresponds to the speaker wearing the terminal 4 and includes standard noise at the speaker position. Yes. The language model used also corresponds to the position of the speaker. If there is a difference equal to or greater than a threshold value between the stored noise spectrum and the actual noise spectrum, the noise is removed by the noise removing unit 10 and then speech recognition is performed.

  In the embodiment, the server 2 instructs the plurality of terminals 4 to perform work such as picking from the speaker 20, and the worker wearing the terminal 4 reports the work status and the like from the microphone 23. Speech recognition. Therefore, the task status storage unit 17 stores the task status for each terminal 4. The task status is, for example, a picking command and a report for the command. In addition to this, the embodiment can also be used for voice recognition when a speaker speaks while moving between a plurality of positions, for example, for maintenance or work. It can also be used for speech recognition by pilots, surgeons, dentists, etc.

  The configuration on the terminal 4 side will be described. The speaker 20 gives an instruction such as picking to the worker, and the position information acquisition device 22 is a GPS, for example, and obtains the current position of the terminal 4. The microphone 23 picks up the operator's voice, and the feature extraction unit 24 converts it into a feature vector. The feature amount vector is obtained by dividing the speech section into appropriate time widths and extracting the feature amount of the frequency spectrum of the voice data for each time width as a vector of 10 to 20 dimensions or the like. The utterance section detection unit 25 discriminates between the utterance section and the non-utterance section, for example, the number of times that the signal from the microphone 23 (bypassing the feature extraction section) changes from outside a predetermined threshold value to the inside and crosses zero. Count. A case where this number of times per time is equal to or greater than a predetermined value is set as an utterance interval, and if not, it is set as a non-utterance interval. This means that an utterance period is set when the frequency per time when a signal having a strength equal to or greater than the threshold is zero crossing is high. In addition, the detection method itself of an utterance area itself is arbitrary. The noise storage unit 26 stores ambient noise in a section other than the speech section (no speech section) via the microphone 23. Noise may be stored as acoustic data or as its spectrum.

  In the embodiment, the speech recognition is performed on the server 2 side, and the noise removal unit 10 to the speech recognition unit 16 are provided in the server 2. However, these processes may be performed by the terminal 4, in which case the server 2 compares the actual noise on the terminal 4 side with the noise spectrum stored in the storage unit 8, and if the difference is greater than or equal to the threshold value, the threshold value is set. The terminal 4 side is requested to remove the excess noise. It also commands which acoustic model to use. Note that GPS is an expensive device, and if the terminal 4 does not include the device 22, the server 2 compares the noise from the terminal 4 with all the noise spectra stored in the storage unit 8, and determines the position of the terminal 4 as noise. Judged from the similarity of.

  When speech recognition is performed on the terminal 4 side, the noise removal unit 28, the language model storage unit 30, the acoustic model storage unit 32, and the speech recognition unit 34 are provided in the terminal 4. The language model is a language model for each position, and is for changing the vocabulary according to the work position. The acoustic model storage unit 32 stores the acoustic model of the worker wearing the terminal 4 with noise at different positions. The task state storage unit 36 stores the state of tasks such as picking related to the terminal 4, the communication unit 38 communicates with the server 2, and the power source 40 such as a battery supplies power to the terminal 4.

  Here, the meaning of creating and storing an acoustic model for each position will be described. Even if the same speaker has different ambient noise, the voice changes. This is recognized on a daily basis when comparing a conversation in a quiet place with a conversation in a noisy place. The stored acoustic model includes not only the audio signal but also ambient noise. Therefore, by using an acoustic model for each location, more precisely, an acoustic model for each location and person, it is possible to store an acoustic model that includes noise at that location and is close to actual speech.

  FIG. 2 shows a speech recognition algorithm. On the terminal side, noise is recorded outside the speech section, and noise data (acoustic data) or its spectrum is obtained and stored. If a GPS is provided, the current position is always obtained (step 1). When communicating with the server 2 due to the start of speech or the like, the current position and noise data are transmitted from the terminal side to the server side. The noise data may be transmitted as a noise spectrum. In addition to the noise data, the voice data is converted into, for example, a feature vector by the feature extraction unit 24 and transmitted to the server 2 (step 2). When voice recognition is performed by the terminal 4, transmission of voice data itself is not necessary.

  On the server side, if the terminal does not have a GPS or the like, the current position is unknown and the noise received in step 3 is compared with the stored noise spectrum. It is assumed that the terminal exists at the nearest noise spectrum position. In step 4, the stored value of the noise spectrum corresponding to the current position is compared with the actual noise spectrum. If the difference is equal to or smaller than the threshold value, the acoustic model at that position is used as it is. If the difference is equal to or greater than the threshold, noise removal is performed so as to eliminate the error exceeding the threshold. Spectral subtraction or the like is used for noise removal.

  In this way, the acoustic model used is based on the speaker's voice at the position where the noise exists. Accordingly, ambient noise has already been incorporated into the speech model, and when the actual noise deviates from the stored noise spectrum by a threshold or more, noise removal is executed so as to eliminate these differences. As a result, small noise removal may be performed as noise removal, and the audio signal is not distorted as in the case of large-scale noise removal. The server selects a language model according to the position (step 5), and executes speech recognition in step 6. In the case of work such as picking, the progress of the work is grasped by voice recognition, and the next work is instructed together with the current position of the worker.

  3 and 4 show examples of spectral subtraction. The dark lines in FIGS. 3 and 4 indicate the stored noise spectrum, and the light lines indicate the actual noise spectrum. In FIG. 3, these are almost the same. In FIG. 4, the stored noise spectrum 42 and the actual noise spectrum 44 have a frequency band that is shifted by a threshold value or more (not shown). For example, a portion exceeding the threshold value is a region 46 between the spectra 42 and 44. Then, noise in the region 46 or a part thereof is removed, and for example, a portion (region 46) exceeding the threshold from the spectrum of the acoustic signal in the utterance interval, or 2/3 thereof is removed by spectrum subtraction. The noise of 42 minutes in the spectrum has already been incorporated into the acoustic model and does not hinder speech recognition. And since the small noise removal of the area | region 46 is performed, it is small to distort an audio | voice signal.

  5 and 6 show applications in picking. In FIG. 5, 50 and 52 are two workers, for example, the worker 50 moves from the book shelf 53 to the food shelf 54 by a solid line route for picking, and the worker 52 moves on the food shelf 54 by a dotted line route. It is assumed that the picking is performed first and the book shelf 53 is next picked. The positions of the workers 50 and 52 are obtained from GPS or noise spectrum comparison, and the language model is switched accordingly. Therefore, the ambiguous ending in the vicinity of the food shelf 54 should be “individual” instead of “book”, and the ambiguous ending in the vicinity of the book shelf 53 is recognized as “book” instead of “individual”. By narrowing down the vocabulary, high-speed and accurate speech recognition is possible.

  In FIG. 6, it is assumed that the worker moves between the picking zone 60 and the assortment zone 62 (zone where assortment is performed). In the picking zone 60, the voice signal of the operator is interpreted by a language model containing vocabulary corresponding to picking, and in the assort zone 62, it is interpreted by a language model containing vocabulary related to assortment.

In the embodiment, the following effects can be obtained.
(1) The speaker's position is determined through comparison of noise spectra or by a position sensor such as GPS. An acoustic model including noise is used for each position. As a result, when the change from the stored noise spectrum is small, the influence of noise can be made extremely small. Even with the same speaker, the voice changes depending on the surrounding environment. This is already incorporated into the acoustic model for each location.
(2) Instead of learning noise in response to the movement of the speaker, the speaker's position is specified and an acoustic model corresponding to the position is used. Therefore, it is possible to quickly follow the movement of the speaker.
(3) Especially when the position of the speaker is determined by the position sensor, the switching of the acoustic model is not delayed by the movement of the speaker. Note that the position of the speaker can also be determined by comparing the noise spectrum with the actual noise, but the determination of the position may be delayed due to the movement of the speaker due to the acquisition and comparison of noise.
(4) If the noise spectrum at each location differs from the current noise spectrum by more than a threshold, noise removal is performed. Since the noise removal may be performed on a small part exceeding the threshold or a part thereof, the audio signal is hardly distorted.
(5) More accurate speech recognition is possible by switching the language model according to the position of the speaker.
(6) If the speaker does not move but the noise spectrum changes between various patterns, the noise spectrum for each pattern is memorized and it is determined which spectrum the current noise is closest to. . If the acoustic model for each noise spectrum is switched according to the discrimination result, speech recognition can be performed efficiently.

2 Server 4 Terminal 6 Noise comparison unit 8 Noise spectrum storage unit 10, 28 Noise removal unit 12, 30 Language model storage unit 14, 32 Acoustic model storage unit 16, 34 Speech recognition unit 17 Task state storage unit 18, 38 Communication unit 20 Speaker 22 Position information acquisition device 23 Microphone 24 Feature extraction unit 25 Speech section detection unit 26 Noise storage unit 36 Task state storage unit 40 Power supply 50, 52 Worker 53 Book shelf 54 Food shelf 60 Picking zone 62 Assortment zone

Claims (5)

A system for recognizing a noise in a sound signal from a microphone by processing a noise removing unit and storing a sound model by a voice recognition unit,
A noise spectrum storage unit for storing a noise spectrum for each position;
An acoustic model storage unit that stores an acoustic model for each position and including noise at the position;
Position discrimination means;
A noise comparison unit that compares the spectrum of the acoustic signal from the microphone in the non-speech interval with the noise spectrum stored in the noise spectrum storage unit for the position determined by the determination unit,
When the noise comparison unit detects a difference greater than or equal to the threshold value, at least a part of the portion exceeding the threshold value is removed by the noise removal unit with respect to the acoustic signal from the microphone in the utterance interval,
A speech recognition system, wherein speech recognition is performed by the speech recognition unit based on an acoustic model for a position determined by a determination means. A server and a plurality of mobile terminals, and the microphone and the discrimination means are provided in the terminal, and a noise spectrum storage unit, an acoustic model storage unit, and a noise comparison unit are provided in the server,
The speech recognition system according to claim 1, wherein the noise removing unit and the speech recognition unit are provided in either the server or the terminal. The voice recognition system according to claim 1 or 2, wherein the discrimination means is a position sensor. The speech recognition system according to any one of claims 1 to 3, wherein a plurality of language models composed of a dictionary for speech recognition are stored, and the language model is switched according to the position determined by the determining means. A method of recognizing a noise in an acoustic signal from a microphone by a noise recognizing unit and recognizing a sound by a speech recognizing unit storing an acoustic model,
The noise spectrum for each position is stored in the storage means,
Storing an acoustic model including noise at each position in the storage means;
The position is determined by the position determination means,
The comparison means compares the spectrum of the acoustic signal from the microphone in the non-speech section with the noise spectrum stored in the noise spectrum storage unit for the position determined by the determination means, and detects a difference greater than the threshold. At the time, for the acoustic signal from the microphone in the utterance section, at least a part of the portion exceeding the threshold value is denoised by the noise removing unit,
A speech recognition method, wherein speech recognition is performed by the speech recognition unit using an acoustic model for a position determined by a determination means.