JP2009216835A - Sound echo canceling device, in-vehicle device and sound echo canceling method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a sound echo canceling device canceling sound echo with high accuracy, even when an environment in which sound echo canceling for extracting an utterance signal uttered by a person is performed, is complicated. <P>SOLUTION: In the environment in which speech recognition for extracting the utterance signal uttered by a person from an observation signal for indicating a collected utterance signal is performed, the sound echo canceling device performs the steps of: measuring a transfer function for each assumed situation beforehand; holding the measured transfer function; receiving a reference signal for indicating a speech signal with a plurality of channels as it is, which is outputted in the environment with the plurality of channels; estimating the sound echo by using the transfer function which is held for each of the plurality of received channels; canceling the sound echo by using each estimated sound echo; and selecting a result in which a speech likelihood for expressing intensity of characteristic of uttered speech, from the results in which the sound echo is canceled. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to an acoustic echo removal apparatus, an in-vehicle apparatus, and an acoustic echo removal method for performing acoustic echo removal for extracting an utterance signal uttered by a person from an observation signal indicating a collected sound signal.

  Conventionally, in a car voice recognition system, when voice recognition is performed in a situation where music is being output from the speaker in the car, the voice to be recognized and a voice signal (acoustic echo) other than the voice are mixed into the voice recognition microphone. This hinders speech recognition. Therefore, various techniques for removing a sound echo by an acoustic echo canceller and ensuring a speech recognition rate are disclosed.

  For example, as shown in FIG. 7, the 2ch reference signals output from the speakers are combined into 1ch (monaural), and the reference signals combined into 1ch are input to the adaptive filter to estimate the acoustic echo in the vehicle and estimate An audio echo canceller that cancels acoustic echo from an observation signal indicating all sounds collected by a microphone and feeds back the result to an adaptive filter is disclosed (see Non-Patent Document 1).

  Specifically, the voice echo canceller applies an adaptive filter to the observation signal collected by the microphone to estimate the acoustic echo, deletes the estimated acoustic echo from the observation signal, and only the uttered speech spoken. Extract and output the extracted speech to a speech recognition engine. Then, the speech echo canceller apparatus extracts the speech speech with higher accuracy by updating the adaptive filter using the cancellation result obtained by deleting the acoustic echo from the observation signal.

  More specifically, the audio echo canceller apparatus combines the reference signal (Rch) output from the speakers (FR, RR) and the reference signal (Lch) output from the speakers (FL, RL). And input to the adaptive filter. That is, here, the audio echo canceller apparatus obtains acoustic echoes output from the four speakers as 1ch acoustic echoes. Then, the speech echo canceller apparatus estimates an acoustic echo using an adaptive filter to which the 1ch reference signal combined into one is input, deletes the estimated acoustic echo from the observed signal, and uttered speech spoken Only (speech recognition), and the extracted speech is output to the speech recognition engine, and the adaptive filter is updated using the estimated acoustic echo.

  In other words, the speech echo canceller can update the adaptive filter using the estimated acoustic echo, thereby bringing the adaptive filter characteristics closer to the actual transfer characteristics (converging), and utterance speech with higher accuracy. Can be extracted.

Abderrahman, “Speech Enhancement Using Multi-Reference Reduction in a Vehicle Environmen”, pp838-841, Interspeech 2007

  However, the above-described conventional technique has a problem that the accuracy of acoustic echo removal deteriorates when the environment for performing acoustic echo removal for extracting a speech signal uttered by a person is complicated.

  Specifically, in the prior art, after collecting the reference signals into 1ch, the acoustic echo is estimated, and the acoustic echo estimated from the observation signal collected by the microphone is deleted to perform speech recognition. If the transfer characteristic is relatively simple, such as in a room where there is little reflection of sound, the acoustic echo can be canceled sufficiently because it can be simulated by the transfer characteristic of the adaptive filter. However, for example, in the case of complex transfer characteristics in a complicated environment such as in a car, even if the adaptive filter is updated by the estimated acoustic echo, the adaptive filter transfer characteristic does not approach the transfer characteristic of the complex environment, and the acoustic echo Can not be canceled enough. As a result, the accuracy of voice recognition deteriorates.

  Therefore, the present invention has been made to solve the above-described problems of the prior art, and even when the environment for performing acoustic echo removal for extracting speech signals uttered by a person is complicated, acoustic echoes can be accurately performed. An object of the present invention is to provide an acoustic echo removing device, an in-vehicle device, and an acoustic echo removing method capable of removing noise.

  In order to solve the above-described problems and achieve the object, the present invention is assumed in an environment in which acoustic echo removal for extracting a speech signal uttered by a person from an observation signal indicating a collected speech signal is performed. The transfer characteristic holding means for measuring the transfer characteristic in advance for each possible situation and holding the measured transfer characteristic, and the reference signal indicating the audio signal output in a plurality of channels in the environment remain as the plurality of channels. The reference signal receiving means for receiving and the reference signals of the plurality of channels received by the reference signal receiving means for each of the transfer characteristics held by the transfer characteristic holding means, to indicate voice other than the speech signal Using acoustic echo estimation means for estimating acoustic echo, and each acoustic echo estimated by the acoustic echo estimation means, Acoustic echo removal means for performing echo echo removal, and acoustic echo removal result selection for selecting the result having the highest speech likelihood representing the strength of the feature of the speech signal from the result of acoustic echo removal by the acoustic echo removal means Means.

  ADVANTAGE OF THE INVENTION According to this invention, even when the environment which performs the acoustic echo removal which extracts the speech signal uttered by the person is complicated, an acoustic echo can be removed with high precision.

  Exemplary embodiments of an acoustic echo removal device, an in-vehicle device, and an acoustic echo removal method according to the present invention will be described below in detail with reference to the accompanying drawings. In the following, the outline and features of the acoustic echo removal apparatus according to the present embodiment, the configuration of the acoustic echo removal apparatus and the flow of processing will be described in order, and finally various modifications to the present embodiment will be described.

[Outline and features of acoustic echo canceller]
First, the outline and characteristics of the acoustic echo removing apparatus according to the embodiment will be described with reference to FIGS. 1 and 2. 1 and 2 are diagrams for explaining the outline and features of the acoustic echo removal apparatus according to the first embodiment.

  The acoustic echo removal apparatus (acoustic echo canceller) according to the first embodiment is connected to a speech recognition engine that executes various processes using the result of acoustic echo removal from which a speech signal uttered by a person is extracted. It is intended to extract an utterance signal from an observation signal indicating an audio signal collected in an environment such as in a car, and in particular, an environment for performing acoustic echo removal for extracting an utterance signal uttered by a person. The main feature is that acoustic echo can be removed with high accuracy even in a complicated case.

  This main feature will be specifically described using an example in which the acoustic echo removing apparatus according to the first embodiment is applied to a vehicle interior. The acoustic echo removal apparatus is connected to a voice recognition engine, and the voice recognition engine performs a voice operation such as a car navigation operation or a speaker volume based on the voice recognition result received from the acoustic echo removal apparatus. .

  In such a state, the acoustic echo removal apparatus according to the first embodiment can be assumed in an acoustic echo removal environment for extracting a speech signal uttered by a person from an observation signal indicating a collected speech signal. The transfer characteristics are measured in advance for each situation, and the measured transfer characteristics are held.

  Specifically, as shown in FIG. 1, the acoustic echo removal apparatus can be assumed in a vehicle equipped with four speakers (FR, RR, FL, RL) that perform voice recognition, as shown in FIG. The impulse response is measured with the occupant arrangement of the pattern, the transfer characteristic is measured in advance based on the measured impulse response, and the measured transfer characteristic is held. In addition, as the situation that can be assumed in the vehicle, that is, the situation where the impulse response will be different, in addition to the above-mentioned occupant arrangement, the type of car and the arrangement of objects can be considered. A situation where a difference occurs in the impulse response will be described by taking an occupant arrangement as an example. In addition, the passenger arrangements other than the passenger arrangement shown in FIG. 1 include a situation in which there is no driver seat and a situation in which two people are biased to one side of the rear seat. Because it is difficult to think, it is excluded here.

  The acoustic echo removal apparatus receives a reference signal indicating an audio signal output in a plurality of channels in the environment as it is in a plurality of channels. Specifically, in the above example, as shown in FIG. 2, the acoustic echo canceller is configured such that the reference signal (1ch) output from the speaker (FR, RR) provided on the right side of the car and the left side of the car. The reference signal consisting of 2ch with the reference signal (1ch) output from the speaker (FL, RL) provided in is received as a 2ch stereo signal without being integrated into 1ch as a monaural signal.

  Thereafter, the acoustic echo removal apparatus estimates an acoustic echo indicating a sound other than the speech sound by using all of the held transfer characteristics for each of the received reference signals of the plurality of channels. Specifically, in the above example, the acoustic echo canceller recognizes each of the transfer characteristics of the 12 patterns of occupant arrangement held for each of the 2ch reference signals output from the received left and right speakers. Twelve patterns of acoustic echoes indicating voices other than the spoken voice are estimated by inputting to the filter to be performed. In FIG. 2, for the sake of explanation, for the 12 patterns of transfer characteristics shown in FIG. 1, a filter corresponding to FR, a filter corresponding to RR, a filter corresponding to FL, and a filter corresponding to FR However, the present invention is not limited to this, and the transfer characteristics of 12 patterns in which all speakers are input and held in one filter are illustrated. Can be read out one by one to estimate the acoustic echo.

  Then, the acoustic echo removal apparatus performs acoustic echo removal using each estimated acoustic echo. More specifically, in the above example, the acoustic echo removal apparatus includes reference signals output from four speakers (reference signals output from the speakers at the same timing as the reference signals used when estimating acoustic echoes), and A filter that applies estimated 12 patterns of acoustic echoes to an observation signal that is a signal collected by a microphone together with an utterance signal uttered by an occupant when the reference signal is output. Then, acoustic echo removal is performed, and 12 patterns of acoustic echo removal results (speech signal extraction results) are extracted.

  Subsequently, the acoustic echo removal apparatus selects a result having the highest speech likelihood that represents the strength of the feature of the speech that remains after the acoustic echo is removed from the result of acoustic echo removal. Specifically, in the above example, the acoustic echo removal apparatus represents acoustic feature quantities using a function such as MFCC (Mel Frequency Cepstrum Coefficient) for each of the extraction results of 12 patterns subjected to acoustic echo removal. The speech likelihood is calculated, and the acoustic echo removal result having the highest calculated speech likelihood is selected. Then, the acoustic echo removal apparatus outputs the selected result with the highest speech likelihood to the speech recognition engine, and the speech recognition engine executes various operations based on the received speech recognition result.

  Further, the acoustic echo removal apparatus may execute the above-described processing every time it is spoken, or may be executed at a predetermined timing, for example, every 10 minutes.

  As described above, the acoustic echo canceller according to the first embodiment calculates in advance the transfer characteristics that can be assumed in the in-car multi-speaker environment, receives the 2ch reference signal independently, and calculates the received 2ch reference signal in advance. As a result of estimating acoustic echoes from transfer characteristics and canceling acoustic echoes from observed signals, as in the above main features, when the environment for performing acoustic echo removal that extracts speech signals uttered by people is complex However, it is possible to remove acoustic echoes with high accuracy.

[Configuration of acoustic echo canceller]
Next, the configuration of the acoustic echo removal apparatus shown in FIGS. 1 and 2 will be described with reference to FIG. FIG. 3 is a block diagram illustrating the configuration of the acoustic echo removal apparatus according to the first embodiment. As shown in FIG. 3, the acoustic echo removal apparatus 10 includes a microphone 11, a transfer characteristic generation unit 12, a transfer characteristic DB 13, a reference signal reception unit 20, an acoustic echo estimation unit 21, and an acoustic echo removal unit 22. And an acoustic echo removal result output unit 23.

The microphone 11 collects various signals output in the vehicle. Specifically, in the above example, the microphone 11 is a reference output from four speakers (2ch) at the same timing as the reference signal used when the acoustic echo is estimated by the acoustic echo estimation unit 21 described later. The signal and an utterance signal uttered by a person at the same timing as the reference signal are collected as an observation signal, and the collected observation signal is output to an acoustic echo removal unit 22 described later. For example, reference signals of 2ch are respectively x _L and x _R , and transmission characteristics from each speaker to the microphone 11 in the vehicle interior i are h ⁽ⁱ⁾ _FL , h ⁽ⁱ⁾ _FR , h ⁽ⁱ⁾ _RL , h ⁽ⁱ⁾ _Assuming that _{RR is} the voice of the driver and s, the observation signal y ⁽ⁱ⁾ collected by the microphone 11 can be defined as equation (1). Here, N ⁽ⁱ⁾ is an estimated acoustic echo and can be defined as Equation (2).

  The transfer characteristic generation unit 12 transfers transfer characteristics for each possible situation in an environment where acoustic echo removal is performed to extract an utterance signal uttered by a person from an observation signal indicating an audio signal collected by the microphone 11. Is measured in advance, and the measured transfer characteristics are retained. Specifically, in the above example, the transfer characteristic generation unit 12 generates an impulse with an occupant arrangement of 12 patterns that can be assumed in a vehicle equipped with four speakers (FR, RR, FL, RL) that perform voice recognition. The response is measured, and the transfer characteristic is measured in advance based on the measured impulse response (reference: Sato, Journal of the Acoustical Society of Japan, Vol.58, No.58, pp669-676, 2002). To store. The transfer characteristic generation unit 12 may be executed every time an utterance is made, or may be executed at a predetermined timing, for example, every ten minutes.

  The transfer characteristic DB 13 stores the transfer characteristic generated by the transfer characteristic generation unit 12. More specifically, the transfer characteristic DB 13 stores 12 patterns of transfer characteristics that can be assumed and are generated and stored by the transfer characteristic generation unit 12.

  The reference signal receiving unit 20 receives a reference signal indicating an audio signal output in a plurality of channels in the environment as it is. Specifically, in the above example, the reference signal receiving unit 20 includes a reference signal (1ch) output from a speaker (FR, RR) provided on the right side of the car and a speaker ( The reference signal consisting of 2 channels with the reference signal (1 channel) output from (FL, RL) is accepted as a 2ch stereo signal without consolidating it into 1 channel. Then, the reference signal reception unit 20 outputs the received 2ch reference signal to the acoustic echo estimation unit 21 described later.

  The acoustic echo estimation unit 21 estimates an acoustic echo indicating speech other than the spoken speech by using all the transfer characteristics held for each of the reference signals of a plurality of channels received by the reference signal reception unit 20. Specifically, the acoustic echo estimation unit 21 is held in the transfer characteristic DB 13 for each of the 2ch reference signals output from the left and right speakers received by the reference signal receiving unit 20. Each of the transfer characteristics of the occupant arrangement of the pattern is acquired, and the acquired transfer characteristics are input to an acoustic echo removing unit 22 to be described later, and 12 patterns of acoustic echo indicating speech other than the spoken speech are estimated.

For example, when N ^{′ (i)} is the acoustic echo to be estimated, h ^{′ (i)} _L is the transfer characteristic corresponding to the FL, RL speaker, and h ^{′ (i)} _R is the transfer characteristic corresponding to the FR, RR speaker. The acoustic echo estimation unit 21 can estimate the acoustic echo as Expression (3) (i = 1 to 12:12 pattern).

  The acoustic echo removal unit 22 is a filter that performs speech recognition using each acoustic echo estimated by the acoustic echo estimation unit 21. Specifically, in the above example, the acoustic echo removing unit 22 utters the reference signals output from the four speakers received by the reference signal receiving unit 20 and the occupant when the reference signals are output. The acoustic echo is removed by applying 12 patterns of acoustic echo estimated by the acoustic echo estimation unit 21 to the observation signal that is a signal collected by the microphone 11 together with the uttered speech signal. The acoustic echo removal result (speech signal extraction result) is extracted. Then, the acoustic echo removal unit 22 outputs the extracted 12 patterns of acoustic echo removal results to the acoustic echo removal result output unit 23 described later. For example, when the observation signal equation (1) is collected by the microphone 11 and the equation (2) is estimated as the acoustic echo by the acoustic echo estimation unit 21, the acoustic echo removal unit 22 obtains the equation (4) as the acoustic echo removal result. It can be calculated by the formula defined in

  The acoustic echo removal result output unit 23 selects a result having the highest speech likelihood representing the strength of the feature of the uttered speech from the result of the acoustic echo removal performed by the acoustic echo removal unit 22. Specifically, in the above example, the acoustic echo removal result output unit 23 uses a function such as MFCC for each of the 12 patterns extracted by the acoustic echo removal unit 22 to obtain acoustic features. A speech likelihood representing the amount is calculated, and an acoustic echo removal result having the highest calculated speech likelihood is selected. Then, the acoustic echo removal result output unit 23 outputs the selected result with the highest speech likelihood to the speech recognition engine, and the speech recognition engine executes various operations based on the received acoustic echo removal result.

For example, the acoustic echo removal result output unit 23 calculates the MFCC for each of the 12 patterns extracted from the acoustic echo removed by the acoustic echo removal unit 22 shown on the left side of Equation (4). Then, the acoustic echo removal result output unit 23 uses the MFCC, which is the calculated acoustic feature amount, to calculate speech likelihood compared with a speech GMM (Gaussian Mixture Model) prepared in advance. If the MFCC feature of the speaker is “o”, the speech likelihood P (0) is obtained as the sum of W weighted normal distributions as shown in Equation (5). Of the W normal distributions, the “ω” -th average “μ _ω ” and the variance are “μ _σ ”. Also, λ _ω is given by equation (6).

More specifically, the acoustic echo removal unit 22 performs 12 patterns of sound estimated by the acoustic echo estimation unit 21 with respect to the observation signal y ⁽ⁱ⁾ represented by the equation (7) observed in the in-vehicle environment i. Using the echo, 12 patterns of cancellation results are calculated by canceling the acoustic echo from the observation signal. Then, the acoustic echo removal result output unit 23 uses the equations (5) and (6) to calculate the MFCC feature amount from the 12 pattern cancellation results. Here, when each speaker's acoustic model is λ _s , the acoustic echo removal result output unit 23 calculates the value shown on the left side of Equation (8). The acoustic echo removal result output unit 23 selects “i” at this time (for example, the fourth pattern if i = 4) as the result with the highest speech likelihood, and the result with the highest selected speech likelihood (for example, If i = 4, the acoustic echo removal result when the fourth pattern is used is output to the speech recognition engine.

[Processing by acoustic echo canceller]
Next, processing performed by the acoustic echo removal apparatus will be described with reference to FIG. FIG. 4 is a flowchart illustrating a flow of acoustic echo removal processing in the acoustic echo removal apparatus according to the first embodiment.

  As shown in FIG. 4, the acoustic echo removal apparatus 10 applies a transfer characteristic measured in advance to a filter under a situation that can be assumed in an actual vehicle environment (step S101). As a specific example, the acoustic echo removal apparatus 10 applies 12 patterns of transfer characteristics measured with 12 patterns of occupant arrangement that can be assumed in an actual vehicle environment to a filter.

  Then, the acoustic echo removal apparatus 10 inputs the received 2ch reference signal as it is to the filter and estimates acoustic echo (steps S102 and S103). As a specific example, the acoustic echo removal apparatus 10 performs speech recognition on each of the 12-pattern occupant placement transfer characteristics to be held for each of the 2ch reference signals output from the received left and right speakers. Twelve patterns of acoustic echoes indicating voices other than the spoken voice are input to the filter and estimated.

  Subsequently, the acoustic echo removal apparatus 10 deletes the estimated acoustic echo from the observation signal (step S104). To give a specific example, the acoustic echo removal apparatus 10 includes reference signals output from four speakers (reference signals output from the speakers at the same timing as the reference signals used when estimating acoustic echoes), and Using an estimated 12-pattern acoustic echo filter applied to an observation signal that is a signal collected by a microphone together with an utterance signal uttered by an occupant when the reference signal is output Acoustic echo removal is performed, and 12 patterns of acoustic echo removal results (speech signal extraction results) are extracted.

  And the acoustic echo removal apparatus 10 calculates the speech likelihood showing the strength of the characteristic of speech sound from the result of acoustic echo removal (step S105). As a specific example, the acoustic echo removal apparatus 10 calculates a speech likelihood representing an acoustic feature amount using a function such as MFCC for each of the 12 patterns extracted from the acoustic echo.

  Thereafter, the acoustic echo removal apparatus 10 selects and outputs the result having the highest calculated speech likelihood (step S106). As a specific example, the acoustic echo removal apparatus 10 calculates a speech likelihood representing an acoustic feature, selects an acoustic echo removal result with the highest calculated speech likelihood, and outputs the result to the speech recognition engine.

[Effects of Example 1]
As described above, according to the first embodiment, in an in-vehicle multi-speaker environment, an assumed transfer characteristic is calculated in advance, a 2ch reference signal is received independently, and the received 2ch reference signal and the previously calculated transfer characteristic are used. As a result of estimating the acoustic echo and canceling the acoustic echo from the observed signal, the acoustic echo can be removed with high accuracy even in a complicated environment for removing the acoustic echo that extracts the speech signal uttered by the person. Is possible.

  Although the embodiments of the present invention have been described so far, the present invention may be implemented in various different forms other than the embodiments described above. Therefore, as shown below, (1) selection of transfer characteristics by weight, (2) speech likelihood selection method, (3) system configuration, etc. To do.

(1) Selection of transfer characteristics based on weight For example, in the first embodiment, a case has been described in which acoustic echoes are estimated and speech recognition is performed using transfer characteristics of all 12 patterns measured in advance. However, the present invention is not limited thereto, and it is also possible to estimate the occupant arrangement based on the weight detected by the load sensor provided on the seat in the vehicle, and perform voice recognition using the estimated occupant arrangement transfer characteristics.

  Specifically, as shown in FIG. 5, the acoustic echo canceller detects that the driver's seat and the rear seat have one or more weights by means of load sensors provided on the respective seats in the vehicle. Then, the acoustic echo removal apparatus specifies two patterns of “situation 8” and “situation 11” as the current occupant arrangement from the 12 occupant arrangements. Then, the acoustic echo removal apparatus performs speech recognition by estimating the two patterns of acoustic echoes using the transfer characteristics of only the two specified patterns. By doing in this way, it is possible to speed up a process compared with the case where voice recognition is performed using the situation of all 12 patterns. FIG. 5 is a diagram illustrating an example of estimating transfer characteristics using a load sensor.

  In addition, the acoustic echo removal device can use, as parameters for estimating acoustic echo, parameters such as air temperature, running noise, air conditioner sound, seat position, etc. in addition to occupant placement and load. As a result of accurately measuring the transfer characteristics, it is possible to perform acoustic echo removal with higher accuracy even when the environment for speech recognition is complex.

(2) Speech likelihood selection method In the first embodiment, 12 patterns of acoustic echoes are estimated by using all 12 patterns of transfer characteristics measured in advance, acoustic echo removal is performed, and 12 patterns of acoustic echo removal are performed. Although the case where the result having the highest speech likelihood is selected from the results has been described, the present invention is not limited to this.

  For example, as shown in FIG. 6, the acoustic echo removal apparatus first performs acoustic recognition by estimating acoustic echoes using the transfer characteristics of the minimum occupant (1 person) and the maximum occupant (5 persons). To calculate the speech likelihood. Then, the acoustic echo canceling device has a higher likelihood of speech in the case of one occupant than in the case of five occupants, and then the transfer characteristics of each of the occupants (2 persons) and the occupants (3 persons). Is used to estimate acoustic echo and perform speech recognition to calculate speech likelihood. Then, the acoustic echo canceller selects the largest speech likelihood for each of one occupant, two occupants, and three occupants. FIG. 6 is a diagram illustrating an example of a speech likelihood selection method.

  On the other hand, when the acoustic likelihood is larger in the case of five occupants than in the case of one occupant, the acoustic echo canceller next transfers the respective transfer characteristics of the occupants (3 persons) and the occupants (4 persons). Is used to estimate acoustic echo and perform speech recognition to calculate speech likelihood. Then, the acoustic echo canceller selects the largest speech likelihood for each of three passengers, four passengers, and five passengers.

  In this way, by narrowing down from the situation where there is a difference in transfer characteristics, estimating acoustic echoes in a situation close to the actual occupant arrangement, performing speech recognition, and calculating the speech likelihood, all 12 patterns are combined. Compared with processing, it is possible to reduce the processing load, and it is possible to perform acoustic echo removal with higher accuracy even when the environment in which acoustic echo removal is performed is complicated.

  Note that the method illustrated in the first embodiment and FIG. 6 is merely an example, and the processing can be performed in the same manner as the above-described method even when the maximum number of passengers is eight or ten. In that case, the occupant arrangement is 12 patterns or more, and acoustic echoes are estimated using all patterns, and acoustic likelihood is calculated by performing acoustic echo removal.

(3) System configuration etc. In addition, among the processes described in the present embodiment, all or a part of the processes (for example, transfer characteristic measurement process) described as being automatically performed may be manually performed. it can. In addition, the processing procedure, control procedure, specific name, and information including various data and parameters shown in the above-described document and drawings can be arbitrarily changed unless otherwise specified.

  Further, each component of each illustrated apparatus is functionally conceptual, and does not necessarily need to be physically configured as illustrated. In other words, the specific form of distribution / integration of each device is not limited to that shown in the figure, and all or a part thereof may be functionally or physically distributed or arbitrarily distributed in arbitrary units according to various loads or usage conditions. It can be configured by integrating (for example, integrating an acoustic echo removal unit and an acoustic echo removal result output unit). Further, all or any part of each processing function performed in each device may be realized by a CPU and a program analyzed and executed by the CPU, or may be realized as hardware by wired logic.

(4) Program The acoustic echo deletion method described in the present embodiment can be realized by executing a prepared program on a computer such as a personal computer or a workstation. This program can be distributed via a network such as the Internet. The program can also be executed by being recorded on a computer-readable recording medium such as a hard disk, a flexible disk (FD), a CD-ROM, an MO, and a DVD and being read from the recording medium by the computer.

  As described above, the acoustic echo removing device, the vehicle-mounted device, and the acoustic echo removing method according to the present invention perform voice recognition that extracts a speech signal uttered by a person from an observation signal indicating a collected speech signal. This is useful, and is particularly suitable for removing acoustic echoes with high accuracy even when the environment for removing acoustic echoes for extracting speech signals uttered by a person is complex.

It is a figure for demonstrating the outline | summary and the characteristic of the acoustic echo removal apparatus which concern on Example 1. FIG. It is a figure for demonstrating the outline | summary and the characteristic of the acoustic echo removal apparatus which concern on Example 1. FIG. 1 is a block diagram illustrating a configuration of an acoustic echo removal apparatus according to Embodiment 1. FIG. 3 is a flowchart showing a flow of acoustic echo removal processing in the acoustic echo removal apparatus according to Embodiment 1. It is a figure which shows the example in the case of estimating a transfer characteristic using a load sensor. It is a figure which shows the example of the speech likelihood selection method. It is a figure for demonstrating a prior art.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 Acoustic echo removal apparatus 11 Microphone 12 Transfer characteristic production | generation part 13 Transfer characteristic DB
DESCRIPTION OF SYMBOLS 20 Reference signal reception part 21 Acoustic echo estimation part 22 Acoustic echo removal part 23 Acoustic echo removal result output part

Claims (6)

In the environment where acoustic echo removal is performed to extract the speech signal uttered by a person from the observed signal indicating the collected speech signal, the transfer characteristics are measured in advance for each possible situation and measured transmission. Transfer characteristic holding means for holding the characteristics;
A reference signal receiving means for receiving a reference signal indicating an audio signal output in a plurality of channels in the environment as it is in the plurality of channels;
An acoustic echo for estimating an acoustic echo indicating speech other than the speech signal, using each of the transfer characteristics held by the transfer characteristic holding means for each of the reference signals of the plurality of channels received by the reference signal receiving means. An estimation means;
Using each acoustic echo estimated by the acoustic echo estimation means, acoustic echo removal means for performing the acoustic echo removal,
From the result of acoustic echo removal by the acoustic echo removal means, acoustic echo removal result selection means for selecting the result with the highest speech likelihood representing the strength of the feature of the speech signal;
An acoustic echo removal apparatus comprising:   The acoustic echo deletion result selection means is connected to the acoustic echo removal apparatus and outputs the result having the highest selected speech likelihood to a speech recognition engine that executes various processes using the acoustic echo removal result. The acoustic echo removing apparatus according to claim 1. In the vehicle that performs acoustic echo removal to extract the utterance signal uttered by the driver's seat from the observation signal indicating the collected voice signal, the transfer characteristic is measured in advance for each possible occupant placement situation. A transfer characteristic holding means for holding the measured transfer characteristic;
A reference signal receiving means for receiving a reference signal indicating an audio signal output in two channels in the vehicle as it is in the two channels;
An acoustic echo for estimating an acoustic echo indicating speech other than the speech signal, using each of the transfer characteristics held by the transfer characteristic holding means for each of the two-channel reference signals received by the reference signal receiving means. An estimation means;
Using each acoustic echo estimated by the acoustic echo estimation means, acoustic echo removal means for performing the acoustic echo removal,
From the result of acoustic echo removal by the acoustic echo removal means, acoustic echo removal result selection means for selecting the result with the highest speech likelihood representing the strength of the feature of the speech signal;
A vehicle-mounted device provided with an acoustic echo removing device.   In addition to the occupant arrangement, the transfer characteristic holding means measures the transfer characteristic in advance for each situation including temperature, running noise, air conditioner sound, and seat position, and holds the measured transfer characteristic. The in-vehicle device according to claim 3.   The acoustic echo estimation means estimates the occupant arrangement based on the weight detected by the load sensor provided on the seat in the vehicle for each of the two-channel reference signals received by the reference signal reception means, The acoustic echo indicating the voice other than the speech signal is estimated using the estimated transfer characteristics of the occupant arrangement among all the transfer characteristics held by the transfer characteristic holding means. In-vehicle device. In the vehicle that performs acoustic echo removal to extract the utterance signal uttered by the driver's seat from the observation signal indicating the collected voice signal, the transfer characteristic is measured in advance for each possible occupant placement situation. A transfer characteristic holding step for holding the measured transfer characteristic;
A reference signal receiving step for receiving a reference signal indicating an audio signal output in two channels in the environment as it is in the two channels;
An acoustic echo that estimates an acoustic echo indicating a sound other than the speech signal by using each of the transfer characteristics held by the transfer characteristic holding step for each of the two-channel reference signals received by the reference signal receiving step. An estimation process;
Using each acoustic echo estimated by the acoustic echo estimation step, an acoustic echo removal step for performing the acoustic echo removal,
From the result of acoustic echo removal by the acoustic echo removal step, an acoustic echo removal result selection step for selecting a result with the highest speech likelihood representing the strength of the feature of the speech signal;
A method for removing acoustic echo in a vehicle-mounted device, comprising:

Images