JP2005303574A - Voice recognition headset - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a voice recognition headset that does not need spherical wave approximation even when a sound source is close to a microphone array and does not have to correct a difference in an arrival time interval between microphones even about the movement of a speaker. <P>SOLUTION: Right an lift ear muffs 2 and 3 have shapes connected with a flexible support frame 1 and are put on the head of a user, wherein an arm 6 extends from one ear muff 2, and microphones 7A and 7B are arranged at the arm 6. When the user wears the headset, the user fixes the arm 6 so that a plurality of the microphones are located on the same spherical surface centered almost on the mouth 8 of the user and the respective microphones are equally distant from the mouth. A signal processing module 10 performs the microphone array processing of voice signals inputted from the plurality of microphones, suppresses noise to output the voice signal with an enhanced voice, recognizes this voice output and outputs the recognized result. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention is one of headset microphone technologies used in hands-free calling, voice recognition, and the like, and uses a plurality of microphones to emphasize a target voice signal from an input acoustic signal or detect a direction of a target voice signal. It is about technology.

  When speech recognition technology is used in a real environment, ambient noise has a large effect on the recognition rate. For example, when used in a car, there are many noises such as car engine noise, wind noise, oncoming and overtaking vehicle sounds, and car stereo sounds. Even in relatively quiet places such as offices, there are many noises that hinder voice recognition, such as footsteps and door closing sounds. Further, the present invention is applied not only to a voice input unit for voice recognition but also to a voice call in a telephone or the like in a noisy environment. These noises are mixed with the voice of the speaker and input to the voice recognition device, causing a significant reduction in the recognition rate. One method for solving such a noise problem is to use a microphone array. The microphone array performs signal processing on sound input from a plurality of microphones, and outputs a signal that emphasizes the target sound. Specifically, the target voice is emphasized by forming a sharp directivity in the direction of the target voice and lowering the sensitivity in other directions.

  For example, in the case of a delay-and-sum type microphone array (see, for example, Non-Patent Document 1), the output signal Se (t) is a signal Sn (t) (n = 1,...) Obtained by N microphones. , N) are shifted and added by a time difference τ that matches the direction of arrival of the target speech. In other words, the emphasized audio signal Se (t) is

N
Se (t) = ΣSn (t + nτ) (1)
n = 1
It is expressed. However, the microphones shall be arranged in the order of the subscript n at equal intervals. The delay-and-sum array forms directivity in the direction of the target voice by using the phase difference of the incoming signals. In other words, it is based on the principle that the target signal is superimposed and strengthened in the same phase, whereas noises coming from directions different from the target signal are weakened because the phases are shifted from each other.

There is also a method for detecting the sound source direction using an adaptive beamformer. (For example, see Patent Document 1.)
By the way, when the speaker speaks at a relatively close distance to the microphone array, the voice reaches the microphone as a spherical wave. Therefore, even if the speaker speaks in front of the microphone array, the arrival time of the sound wave is delayed in the microphone at the end as compared with the microphone in the central part constituting the microphone array. The method shown in (1) is a theory when it is assumed that the sound source is at infinity and the sound wave can be approximated to a plane wave. In this case, it is necessary to treat the sound wave as a spherical wave. When handling as a spherical wave, in addition to the disadvantage that the calculation is more complicated than a plane wave, the difference in the arrival time of sound waves between microphones changes even when the speaker moves in the depth direction. Therefore, there is a problem that the speaker's utterance position is limited.

"Acoustic systems and digital processing", Chapter 7, IEICE, 1995 Japanese Patent Laid-Open No. 11-052977

  When the sound source is close to the microphone array as described above, plane wave approximation is not established, spherical wave approximation is required, processing is complicated, and the utterance position is limited.

  The present invention has been made to solve the above-described problem. Even when the sound source is close to the microphone array, the spherical wave approximation is not required, and the difference in arrival time between the microphones is also corrected with respect to the movement of the speaker. An object of the present invention is to provide a voice recognition headset that does not require a voice recognition.

  In order to solve the above problems, the present invention synthesizes a plurality of microphones for detecting a sound and generating a sound signal, a microphone support section for arranging the plurality of microphones, and a sound signal of the plurality of microphones. Emphasized speech signal generating means for generating the enhanced speech signal, and speech recognition means for recognizing the enhanced speech signal.

  The present invention also provides a plurality of microphones that detect sound and generate sound signals, a microphone support portion that arranges the plurality of microphones on the same circumference centered on the mouth, and sound signals of the plurality of microphones. Emphasized speech signal generating means for synthesizing and generating an enhanced speech signal; and speech recognition means for recognizing the enhanced speech signal.

  In addition, a plurality of microphones that detect sound and generate a sound signal, a microphone support unit that places the plurality of microphones on the same spherical surface centered on the mouth, and a sound signal of the plurality of microphones are combined and emphasized. It is characterized by comprising enhanced speech signal generating means for generating a speech signal and speech recognition means for recognizing the enhanced speech signal.

  Further, the microphones are supported so that the distances between the plurality of microphones are kept constant.

  The present invention can easily perform signal processing and improve the recognition rate even when the sound source is close to the microphone array. Further, the speech recognition rate can be maintained without limiting the speaker position.

  Embodiments of the present invention will be described below with reference to the drawings.

  FIG. 1 and FIG. 2 show the appearance of a voice recognition headset according to Embodiment 1 of the present invention and the schematic system configuration thereof. FIG. 1A is a front view of the voice recognition headset, and FIG. 1B is a side view seen from the direction of the arrow in FIG. The voice recognition headset detects the sound produced by the support frame 1, the ear pads 2, 3, the speakers 4, 5, the arm 6, and the headset wearer (user) and generates an electrical sound signal. And a signal processing module 10 that recognizes the voice signal through digital conversion. In FIG. 1, for the sake of simplicity, two microphones are used, but it is also possible to carry out by arranging three or more microphones. Moreover, the directivity of the microphone may be arranged toward the mouth.

  This voice recognition headset (hereinafter simply referred to as “headset” in some cases) has a shape in which left and right ear pads 2 and 3 are connected by a flexible support frame 1 and is attached to a user's head. And use it. An arm 6 extends from one ear pad 2, and microphones 7 </ b> A and 7 </ b> B are disposed on the arm 6. The microphones 7A and 7B are positioned on the same circumference 9A around the user's mouth 8 when the user wears the headset, and each microphone is fixed to the arm 6 so that the distance from the mouth is equal. To do.

  A speaker 5 (left and right) and a signal processing module 16 are built in the earpiece 2. Note that the signal processing module 16 may be built in the earpiece 3, and although not shown, each element is connected by a cable as necessary.

  As shown in FIG. 2, the sound input from the first microphone 7 </ b> A undergoes analog-digital conversion by the first AD converter 11 </ b> A and is input to the microphone array signal processing unit 12. Similarly, the voice input from the second microphone 7B undergoes analog-digital conversion by the second AD converter 11B and is input to the microphone array signal processing unit 12. The microphone array signal processing unit 12 performs microphone array processing on audio signals input from a plurality of microphones, and outputs an audio signal that suppresses noise and emphasizes the audio. The voice recognition unit 13 recognizes this voice output and outputs a recognition result.

  FIG. 3 is a diagram illustrating an example of the internal structure of the microphone array signal processing unit 12. The digitally converted microphone audio signals are synthesized by the adder 121, and the audio in which the audio signals of the microphones 7A and 7B are emphasized is output. This is because the microphones 7A and 7B are arranged on the same circumference 9A with the mouth 8 as the center, so that they are equidistant from the sound source of the mouth 8, and the sound can be heard without causing a delay due to the spherical wave from the sound source. Entered. Therefore, by adding the signals having the same phase without requiring a separate delay device, the voice coming from the mouth 8 is enhanced.

  FIG. 4 is a diagram showing an example of the internal structure of another macrophone array signal processing unit 12. This is to enhance noise by suppressing noise using an adaptive beamformer. The microphone array signal processing unit 12 includes an adder 121, a beamformer processing unit 122, and a speech enhancement unit 123.

  Here, description will be made using the internal configuration of the beamformer processing unit 122 with reference to FIG. The beamformer processing unit 122 performs a filter calculation process called adaptive beamformer processing for suppressing a sound signal from a sound source on a sound signal obtained by analog-digital conversion of the microphones 7A and 7B. Various methods are known as processing methods inside the beamformer processing unit 122, such as a generalized sidelobe canceller (GSC), a frosted beamformer, and a reference signal method. This embodiment can be applied to any adaptive beamformer, but here, a two-channel GSC will be described as an example.

  FIG. 5 shows a configuration example of a general Jim-Griffith type GSC in a two-channel GSC as an example of the beamformer processing unit 122. The beamformer processing unit 122 is a GSC that includes a subtractor 1221, an adder 1222, a delay unit 1223, an adaptive filter 1224, and a subtractor 1225. Various adaptive filters such as LMS, RLS, and projective LMS can be used as the adaptive filter 1224. For example, La = 50 is used as the filter length La. The delay amount of the delay device 1223 is, for example, La / 2.

  When an LMS adaptive filter is used as the adaptive filter 1224 of the two-channel Jim-Griffith GSC shown in FIG. 5 constituting the beamformer 122, this filter update is performed by setting the coefficient of the adaptive filter 24 to W (time n). n), the i-th channel input signal is xi (n), the i-th channel input signal vector is Xi (n) = (xi (n), xi (n−1),..., xi (n−La + 1)) It is expressed by the following formula.

y (n) = x0 (n) + xl (n) (2)
X ′ (n) = X1 (n) −X0 (n) (3)
e (n) = y (n) -W (n) X '(n) (4)
W (n + 1) = W (n) 1 μX ′ (n) e (n) (5)
When a signal arrives from the direction of the target sound source, the sensitivity of the filter in the beamformer processing unit 122 is low in the direction of the target sound source, so the directivity that is the direction dependency of the sensitivity is examined from the filter coefficient of this filter. Thus, the direction of the target sound source is estimated.

  By the way, in an environment where there are too many noise sources and the noise source direction cannot be specified, the noise suppression performance by the beamformer is degraded, but the input speech has directionality, so that the beamformer with the target direction set as the noise direction. Thus, it is possible to extract only the output of noise with the signal from the target sound source suppressed. Accordingly, the output of the beamformer processing unit 122 is a noise-only signal, and the speech enhancement unit 123 enhances speech by a spectral subtraction (SS) method using the signal.

  Spectral subtraction includes 2chSS that uses two channels of a noise signal for reference and an audio signal and 1chSS that uses only an audio signal of one channel. In this embodiment, the output of the beamformer processing unit 122 is used as reference noise. Speech enhancement is performed by 2chSS using. Normally, as the 2chSS noise signal, a microphone signal separated from the target voice collecting microphone is used so that the target voice is not input, but the nature of the noise signal is different from the noise mixed in the target voice collecting microphone. As a result, there is a problem that the accuracy of the SS decreases.

  On the other hand, in this embodiment, a noise signal is extracted by a microphone array method using a plurality of microphones without using a dedicated microphone for noise collection. Can perform SS.

  2chSS has a configuration as shown in FIG. 6, for example, and the processing shown in FIG. 2chSS shown in FIG. 6 includes a first FFT 1231 for Fourier transforming a noise signal, a first band power converter 1232 for converting a frequency component obtained by the first FFT into band power, and the obtained band power as time. A noise power calculation unit 1233 that averages in the direction, a second FFT 1234 that Fourier-transforms the audio signal, and a second band power conversion unit 1235 that converts the frequency component obtained by the second FFT into band power. A voice power calculation unit 1236 that averages the band power in the time direction, a band weight calculation unit 1237 that calculates a weight for each band from the obtained noise power and voice power, and a frequency obtained by the second FFT from the voice signal A weighting unit 1238 for weighting the spectrum with a weight for each band, and performing inverse FFT on the weighted frequency spectrum for sound An inverse FFT1239 to output a.

  The block length is, for example, 256 points and is matched with the FFT score. In the case of FFT, for example, windowing is performed using a Hanning window, and the same processing is repeated while shifting by 128 points, which is half the block length. Finally, the waveform resulting from the inverse FFT is added while being overlapped by 128 points to restore the deformation due to windowing and output.

  For conversion to band power, for example, as shown in Table 1, the frequency components are divided into 16 bands, and the sum of squares of the frequency components is calculated for each band to obtain band power. The calculation of noise power and voice power is performed for each band, for example, using the first-order regression filter as follows:

pk, n = a.ppk + (1-a) .pk, n-1 (6)
vk, n = a.vvk + (1-a) .vk, n-1 (7)
Where k is the number of the band, n is the scent of the block, p is the band power of the averaged noise channel, pp is the band power of this block of the noise channel, and v is the averaged band of the voice channel Power, vv is the band power of this block of the voice channel, and a is a constant. For example, 0.5 is used as the value of a.

Next, the band weight calculation unit calculates the weight wk, n for each band, for example, by the following equation using the obtained noise and the band power of the voice.
wk, n = | vk, n−pk, n | / vk, n (8) Next, weights for each band are used, and for example, the frequency components of the voice channel are weighted by the following equation.
Yi, n = Xi, n .wk, n (9)
Here, Yi, n is a weighted frequency component, Xi, n is a frequency component obtained by the second FFT of the voice channel, i is a frequency component number, and corresponds to frequency component number i in Table 1. The weight wk, n of the band k is used.

  A processing flow of the speech enhancement unit 123 by 2chSS will be described with reference to FIG. First, initial setting is performed, for example, block length = 256, FFT point number = 256, shift point number = 128, and band number = 16 (step S101). Next, the first FFT 1231 reads noise channel data, performs windowing and FFT, and obtains a noise frequency component (step S102). Next, the second FFT 1234 reads voice channel data, performs windowing and FFT, and obtains a frequency component of the voice (step S103). Next, the first band power converter 1232 calculates the noise band power from the noise frequency component according to the correspondence in Table 1 (step S104). Next, in the second band power conversion unit 1235, the band power of the voice is calculated from the frequency components of the voice according to the correspondence in Table 1 (step S105). Next, the noise power calculation unit 1233 obtains the average noise power according to the equation (6) (step S106). Next, the sound power calculation unit 1236 obtains the average sound power according to the equation (7) (step S107). Next, the band weight calculation unit 1237 obtains the band weight according to the equation (8) (step S108). Next, the weighting unit 1238 weights the frequency component of the sound according to the equation (9) with respect to the weighting coefficient obtained in step S108 (step S109). Next, in inverse FFT 1239, the frequency component weighted in step S109 is inverse FFTed to obtain a waveform, and is superimposed on the last 128 points of the waveform obtained up to the previous block and output (step S110).

  The steps S102 to S110 are repeated until there is no input. It is convenient to perform this block processing in synchronization with the entire processing including the beamformer processing. In this case, the block length of the beamformer is made to coincide with the shift length of 128 points in the speech enhancement unit. . In this way, the speech with noise suppressed is output by the speech enhancement unit 123 and is recognized by the speech recognition unit 13.

  Here, the voice recognition unit 13 will be described with reference to FIG. FIG. 8 shows the internal configuration of the voice recognition unit 13. The sound output of the microphone array signal processing unit 12 is first input to the acoustic analysis unit 131. The acoustic analysis unit 131 converts the input voice into feature parameters. As typical feature parameters used for speech recognition, a power spectrum that can be obtained by a bandpass filter or Fourier transform, a cepstrum coefficient obtained by LPC (linear prediction) analysis, etc. are often used. The type of feature parameter does not matter. The acoustic analysis unit 131 converts input speech into feature parameters at regular time intervals. Therefore, the output is a time series of characteristic parameters (characteristic parameter series). This feature parameter series is supplied to the model matching unit 132.

  On the other hand, the recognition vocabulary storage unit 133 stores word reading information necessary to create a speech model of each word constituting the recognition vocabulary and an identifier corresponding to the recognition result when each word is recognized, for example, a command ID is stored. In the present embodiment, voice control based on word recognition is described as an example of voice recognition in the headset, but the present invention is not limited to this. The voice recognition unit 13 in the headset performs voice recognition with a small amount of calculation, memory capacity, and power consumption, such as continuous word recognition, sentence recognition, word spotting, and voice intention understanding, and outputs the result as a voice recognition result.

  The recognition model creation / storage unit 134 outputs the speech model of each word according to the recognition vocabulary stored in the recognition vocabulary storage unit 133 and the model collation unit 132 as a recognition result when each word becomes a recognition result. A word ID as an identification signal is stored in advance. Of course, when recognition other than word recognition is performed, an identification signal corresponding to the recognition is stored.

  The model matching unit 132 obtains the similarity or distance between each speech model of the word to be recognized stored in the speech model creation / storage unit 134 and the feature parameter series of the input speech, and the similarity is maximum. The word ID associated with the speech model (or the smallest distance) is output as a recognition result.

  As a matching method of the model matching unit 132, a speech model is also expressed by a feature parameter sequence, and a distance between the feature parameter sequence of the speech model and the feature parameter sequence of the input speech is obtained by DP (dynamic programming), A method of expressing a speech model using an HMM (Hidden Markov Model) and calculating the probability of each speech model when a feature parameter sequence of the input speech is input is widely used. It doesn't matter. The word ID output from the model matching unit 132 is output as the recognition result of the speech recognition unit 13 as it is.

  FIG. 9 shows the result of a recognition experiment when speech recognition is performed using the speech recognition headset of the present invention. The conventional method recognizes the sound input using only one microphone with the headset, and the method of the present invention performs the microphone array processing with the sound input from the two microphones disposed on the headset. This is the result when the emphasized voice is input to the voice recognition device. The recognition rate of speech recognition is improved so that the error is reduced by about 6% with respect to the noise of the telephone bell.

  The evaluation result when the sound source direction detection function by the microphone array signal processing unit 10 of the headset of the present invention is used to reject the input of speech or noise that is not subject to speech recognition is as follows. However, it was possible to prevent any errors due to the use of the sound source direction detection function of the headset of the present invention.

  Here, the evaluation data was collected by a person 80 cm adjacent to the recognition target speaker by generating four types of noise and superimposing them on the speech of the recognition speaker. There are four types of noise: voice, telephone bell, paper turning sound, and keyboard keystroke sound.

  As described above, since a plurality of microphones are arranged on the same circumference centered on the mouth with the above configuration, the distance from the mouth, which is a sound source, to each microphone is equal, so that the microphone array audio signal is emphasized. In this case, voice enhancement can be performed with a simple configuration without the need for a delay device, and the voice recognition rate can be further improved. Further, the speech recognition rate can be maintained without being limited to the speaker position. Although the directivity of the microphone is not particularly described here, if the directivity of each microphone is arranged toward the mouth, the voice signal for voice recognition can be further enhanced, and the recognition rate can be improved. .

  FIG. 10 shows the appearance of a voice recognition headset according to Embodiment 2 of the present invention. FIG. 10A is a front view of the voice recognition headset, and FIG. 10B is a side view seen from the direction of the arrow in FIG. The same number is attached | subjected to the same structure as Example 1 mentioned above. The voice recognition headset detects the sound produced by the support frame 1, the ear pads 2, 3, the speakers 4, 5, the arm 6, and the headset wearer (user) and generates an electrical sound signal. And a signal processing module 10 that recognizes the voice signal through digital conversion. In FIG. 10, for the sake of simplicity, two microphones are used, but it is also possible to carry out by arranging three or more microphones.

  This headset has a shape in which left and right ear pads 2, 3 are connected by a flexible support frame 1, and is used by being worn on the user's head. An arm 6 extends from one ear pad 2, and microphones 7 </ b> A and 7 </ b> B are disposed on the arm 6. When the user wears the headset, the microphones 7A and 7B are positioned on the same spherical surface 9B with the user's mouth 8 as the center, and each microphone is fixed to the arm 6 so that the distance from the mouth becomes equal. .

  A speaker 5 (left and right) and a signal processing module 16 are built in the earpiece 2. Note that the signal processing module 16 may be built in the earpiece 3, and although not shown, each element is connected by a cable as necessary.

Since the configuration of the signal processing module 10 from the audio signals detected by the microphones 7A and 7B to the audio recognition is the same as that of the first embodiment, the description thereof is omitted here.
As described above, since the plurality of microphones are arranged on the same spherical surface with the mouth as the center, the distance from the mouth to each microphone is the same as in the first embodiment. Further, it is possible to perform speech enhancement with a simple configuration without requiring a delay device, and to further improve the speech recognition rate. Further, the speech recognition rate can be maintained without being limited to the speaker position.

1 is an external view of a voice recognition headset according to a first embodiment of the present invention. The block diagram which shows the structure of a signal processing module. The block diagram which shows an example of the internal structure of a microphone array process part. Block diagram showing another example of the internal configuration of the microphone array processing unit The block diagram which shows an example of an internal structure of a beam former process part. The block diagram which shows an example of an internal structure of a speech emphasis part. The flowchart which shows the process sequence of an audio | voice emphasis part. The block diagram which shows an example of an internal structure of a speech recognition part. The figure which shows the evaluation result of recognition performance. The external view of the voice recognition headset of Example 2 of the present invention.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 ... Support frame 2, 3 ... Ear cover 4, 5 ... Speaker 6 ... Arm 7A, 7B ... Microphone 8 ... Mouth 10 ... Signal processing module 11A, 11B ... AD converter 12: Microphone array signal processing unit 13: Speech recognition unit

Claims (5)

  A plurality of microphones for detecting a sound and generating a sound signal; a microphone support section for arranging the plurality of microphones; and an enhanced sound signal generating means for synthesizing the sound signals of the plurality of microphones to generate an enhanced sound signal. And a speech recognition means for recognizing the emphasized speech signal.   A plurality of microphones that detect sound and generate a sound signal, a microphone support that places the plurality of microphones on the same circumference centered on the mouth, and a sound that is enhanced by synthesizing the sound signals of the plurality of microphones A speech recognition headset comprising enhanced speech signal generation means for generating a signal and speech recognition means for recognizing the enhanced speech signal. A plurality of microphones that detect sound and generate a sound signal, a microphone support that places the plurality of microphones on the same spherical surface centered on the mouth, and a sound signal that is enhanced by synthesizing the sound signals of the plurality of microphones A speech recognition headset, comprising: an enhanced speech signal generating unit that generates a speech signal; and a speech recognition unit that recognizes the enhanced speech signal.   4. The voice recognition headset according to claim 1, wherein the plurality of microphones are supported so that the distance between them is kept constant.   5. The voice recognition headset according to claim 2, wherein the directivity of the microphone is directed toward the mouth.

Images