JPH10124084A - Voice processer - Google Patents

Abstract

PROBLEM TO BE SOLVED: To prevent the occurrence of the deterioration in the precision of voice recognition even though the average value of an estimated noise spectrum fluctuates by providing the pattern recognition means which collates the estimated voice spectrum against the reference parameters held internally to identify inputted voices. SOLUTION: A pattern recognition section 104 compares the reference parameters, which are the spectrum feature parameter group of various voice elements stored internally and the spectrum feature parameter group made by background noise only, and the estimated voice spectrum group outputted from an estimated voice spectrum computing section 111 and outputs the name of the category of the voice standard pattern, that gives a highest degree of similarity, to an external device as an identification result. Then, the collating errors of the background noise segment portion are reduced in the comparison computations of the estimated voice spectrum group and the voice standard pattern in the section 104 and the collating precision is increased. As a result, the voice recognition performance is improved.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an audio processing device having a function of removing a background noise signal included in an audio signal.

[0002]

2. Description of the Related Art An audio signal input to an audio processing device includes:
A background noise signal is superimposed, and a spectral subtraction method (hereinafter, referred to as an SS method) is used as a simple and effective method for removing the background noise signal.

In the SS method, a speech spectrum is extracted by subtracting (subtracting) a spectrum of a noise signal (hereinafter, referred to as a noise spectrum) from a spectrum of a speech signal on which a noise signal is superimposed (hereinafter, referred to as a speech spectrum). Is the way.

Normally, the noise spectrum is estimated from a portion where no voice is uttered (hereinafter referred to as an estimated noise spectrum). Presumed.

[0005]

However, in estimating the noise spectrum, if the average value of the noise spectrum fluctuates, the error between the estimated noise spectrum and the true noise spectrum increases. In particular, when the estimated noise spectrum value exceeds the true noise spectrum value, the speech spectrum value obtained by the SS method may be negative. Such a phenomenon is called overestimation and causes a decrease in accuracy of speech recognition.

Therefore, a simple and effective means for avoiding a decrease in the accuracy of speech recognition caused by overestimation has been desired.

[0007]

SUMMARY OF THE INVENTION In order to solve this problem, a voice processing apparatus according to the present invention analyzes the spectrum of an input voice signal including a section where voice is not input, obtains a spectrum feature parameter, and obtains an input voice signal. In a sound processing apparatus for processing a signal, (1) sound analysis means for analyzing a spectrum of an input sound signal for each frame and converting it into a spectrum characteristic parameter to calculate sound power; and (2) a section in which sound is input. Voice detection means for detecting from the sound power the voice start frame which is the first frame of
(3) Estimated noise spectrum calculating means for estimating a noise spectrum from a spectrum feature parameter of a section from a frame immediately before the speech start frame to a frame which has been retroactive for a predetermined time and outputting an estimated noise spectrum; Estimated noise spectrum holding means for holding the estimated noise spectrum output from the arithmetic means, (5) reference pattern noise spectrum holding means for holding a spectrum feature parameter of only background noise as a reference pattern noise spectrum, and (6) spectrum feature parameter An estimated voice spectrum is calculated based on a difference between the estimated voice spectrum and the estimated noise spectrum, and when the estimated voice spectrum value is equal to or smaller than the reference pattern noise spectrum value, an estimated voice spectrum calculating means that uses the reference pattern noise spectrum as the estimated voice spectrum; ) Pattern recognition that has a voice spectral characteristic parameter and a spectral feature parameter of only background noise as its reference parameters inside itself, and compares an estimated voice spectrum with the above-mentioned reference parameters contained therein to identify input voice. Means.

[0008]

BEST MODE FOR CARRYING OUT THE INVENTION

(A) Embodiment Hereinafter, an embodiment of an audio processing device according to the present invention will be described in detail with reference to the drawings. In this embodiment, an audio processing device that recognizes the audio type of an input audio signal will be described. Also, in this embodiment, various speech elements (phoneme fragments and words) required to recognize the speech type of the input speech signal
, An operation mode (hereinafter, referred to as a learning mode) for learning a voice standard pattern (reference spectral feature parameter) and learning a spectrum of only background noise at that time (hereinafter, referred to as a learning noise spectrum). It has two operation modes, an operation mode for recognizing a voice signal (hereinafter, referred to as a recognition mode).

(A-1) Configuration of Embodiment FIG. 1 is a functional configuration diagram of a voice processing apparatus of this embodiment. In FIG. 1, the audio processing device of this embodiment
Acoustic analysis unit 101, voice detection unit 102, SS processing unit 103, pattern recognition unit 104, estimated noise spectrum storage memory 105, learning noise spectrum storage unit 1
06.

The SS processing section 103 can be divided into an estimated noise spectrum operation section 110 and an estimated speech spectrum operation section 111 as shown in FIG.

As described above, the speech processing apparatus of this embodiment has the learning mode and the recognition mode. The mode selection signal for selecting these modes is the noise spectrum storage unit 106 for learning, the estimated speech spectrum calculation Part 111;
It is input to the pattern recognition unit 104. Each of the above units has a different function depending on the operation mode. Therefore, first, when the speech processing apparatus of the present embodiment is in the recognition mode by the mode selection signal, the functions of each unit constituting the present embodiment will be described in detail below.

The sound analyzer 101 converts an input speech signal from an analog signal to a digital signal, performs a spectrum analysis process for each frame, and calculates a spectrum feature parameter and a sound power. In the spectrum analysis processing, for example, 16 band-pass filters (hereinafter, referred to as BPs) having different center frequencies are used as spectral feature parameters.
A value Xi (t) obtained by averaging the absolute value of a signal passing through F) for each frame is used. Where t is the frame number,
i represents a BPF number (i = 1 to 16).

The sound analysis unit 101 has a sound power P
As (t), the sum of the spectral feature parameters Xi (t) is calculated as shown in equation (1). In addition, sum 総
Is for i = 1-16.

P (t) = ΣXi (t) (1) The sound detection unit 102 outputs a section in which a sound is input from the sound power P (t) output from the sound analysis unit 101 (hereinafter referred to as a sound section). Call).

Here, the first frame of the voice section is called a voice start frame ts, and the last frame of the voice section is called a voice end frame te. Regarding the detection of the voice section, the sound power of the voice section and the background noise section (the section where no voice is input and only the background noise is input)
The sound section may be detected by utilizing the fact that the sound power is different from the sound power. Further, a detection method based on audio feature amounts as disclosed in JP-A-62-211698 may be used.

The estimated noise spectrum calculation unit 110 calculates a frame (ts-NT) which is a predetermined time (NT-1) from the frame (ts-1) immediately before the speech start frame ts.
The average value of the spectral characteristic parameters up to this point (hereinafter referred to as an estimated noise spectrum Ai) is calculated and output to the estimated noise spectrum storage memory 105. The equation for calculating the estimated noise spectrum Ai is shown in equation (2). Note that the sum Σ is for t where ts−NT to ts−1.

Ai = (1 / NT) ΣXi (t) (2) The estimated noise spectrum storage memory 105 stores the estimated noise spectrum Ai output from the estimated noise spectrum calculation unit 110, and 111.

The learning noise spectrum storage section 106 supplies the learning noise spectrum Bi stored therein to the estimated speech spectrum calculation section 111 in a learning mode described later.

The estimated speech spectrum calculator 111 calculates the estimated noise spectrum component k · from the spectrum feature parameter Xi (t) of the speech section output from the acoustic analyzer 101.
The estimated speech spectrum sequence XS (t) is calculated by subtracting Ai. However, after the estimated noise spectrum component subtraction processing, the estimated speech spectrum series XS
If the value of XSi (t), which is an element of (t), is equal to or less than the value of the learning noise spectrum Bi output from the learning noise spectrum storage unit 106, the learning noise spectrum Bi
Is substituted into the estimated voice spectrum XSi (t). The equation for calculating the estimated voice spectrum XSi (t) is shown in equation (3). Here, k is a coefficient indicating a preset degree of subtraction, and usually a value of about 0.5 to 2.0 is selected.

XSi (t) = Xi (t) −k · Ai When Xi (t) −k · Ai> Bi = Bi Xi (t) −k · Ai ≦ Bi (3) Pattern Recognition Unit 104 Is a spectral characteristic parameter sequence of various speech elements stored in itself (called a speech standard pattern; not shown) and a spectral characteristic parameter sequence Xi (t) of only background noise (noise spectrum at learning). The reference parameter, which is the same as the learning noise spectrum Bi stored in the storage memory, is compared with the estimated speech spectrum series XS (t) output from the estimated speech spectrum calculation unit 111, and is the highest. The category name of the audio standard pattern giving the similarity is output to an external device (not shown) or the like as a recognition result.

As a method of calculating the similarity, for example,
A DP (Dynamic Programming) method may be applied.

Next, the function of each section constituting the embodiment when the embodiment is in the learning mode by the mode selection signal will be described in detail. Note that the functions of the sound analysis unit 101 and the sound detection unit 102 are the same as those of the present embodiment in the recognition mode, and a description thereof will be omitted. Further, among the units constituting this embodiment, the functions of the SS processing unit 103 (the estimated noise spectrum calculation unit 110 and the estimated speech spectrum calculation unit 111) and the estimated noise spectrum storage memory 105 are only available in the recognition mode. It is necessary.

The learning noise spectrum storage unit 106 stores the spectrum feature parameters Xi (t) of the background noise section from the speech start frame ts and the speech end frame te among the spectrum feature parameters Xi (t) output from the acoustic analysis unit 101. t) is discriminated and stored as a learning noise spectrum Bi.

The pattern recognizing unit 104 includes the sound analyzing unit 10
1 is the spectral characteristic parameter Xi (t) output from
, A speech standard pattern and a spectrum characteristic parameter sequence of only background noise at that time are stored in the self as reference parameters. Specifically, the pattern recognizing unit 104 calculates the spectral characteristic parameter Xi (t)
Is determined from the speech start frame ts and the speech end frame te as to whether this is a spectrum feature parameter in the speech section or a spectrum feature parameter in the background noise section. Then, a speech standard pattern is learned for various speech elements from the speech section, and a spectrum sequence of only the background noise at that time (the learning noise spectrum B stored in the learning noise spectrum storage unit 106) is learned from the background noise section.
(i.e., the same spectral information as i) is learned, and the above-described spectral characteristic parameter sequence is stored in the self.

The learning noise spectrum storage unit 106
In the pattern recognition unit 104, the noise spectrum at learning Bi and the reference parameter are updated each time the embodiment is newly learned. Further, the learning noise spectrum storage unit 106 and the pattern recognition unit 10
In 4, the spectral characteristic parameter of the speech section and the spectral characteristic parameter of the background noise section are discriminated by the speech start frame ts and the speech end frame te, but this method is not limited to this method.

(A-2) Operation of the Embodiment The operation of the audio processing apparatus of the embodiment having the above configuration will be described below. As described above, this embodiment includes the recognition mode and the learning mode, and the operation of each unit configuring the embodiment differs depending on the operation mode. Therefore, first, the operation of this embodiment when the voice processing device of this embodiment is in the recognition mode by the mode selection signal will be described.

The input speech signal input to the speech processing apparatus of this embodiment is subjected to spectrum analysis processing for each frame by the sound analysis unit 101, and the sound power P (t) and the spectrum characteristic parameter Xi (t) are calculated. Is done.

The sound power P (t) is calculated by
, And the spectral feature parameter Xi (t) is input to the estimated noise spectrum calculator 110 and the estimated speech spectrum calculator 111.

The sound power P input to the voice detector 102
From (t), a voice start frame ts and a voice end frame te are detected.

The speech start frame ts input to the estimated noise spectrum calculator 110 and the spectrum feature parameter X
From i (t), the estimated noise spectrum Ai is calculated by the above (2)
Estimated noise spectrum storage memory 1 calculated by the equation
05 is output.

The learning noise spectrum storage section 106 supplies the learning noise spectrum Bi stored therein to the estimated speech spectrum calculating section 111.

The estimated speech spectrum calculation unit 111 receives a spectrum feature parameter Xi (t), an estimated noise spectrum Ai, and a learning noise spectrum Bi.
Then, the estimated speech spectrum sequence XS (t) is calculated by the above equation (3) and output to the pattern recognition unit 104.

The pattern recognition unit 104 calculates the similarity between the estimated speech spectrum sequence XS (t) output from the estimated speech spectrum calculation unit 111 and a reference parameter stored in itself, and determines the most similar speech. The category name of the standard pattern (spectral feature parameter of the voice element) is output as a recognition result.

Next, the operation of this embodiment when the embodiment is in the learning mode will be described. Note that the operations of the sound analysis unit 101 and the voice detection unit 102 are the same as the operations in this embodiment in the recognition mode, and a description thereof will be omitted.

The spectrum feature parameter Xi (t) output from the acoustic analysis unit 101 is input to the learning noise spectrum storage unit 106 and the pattern recognition unit 104.

The learning noise spectrum storage unit 106 stores only the spectral feature parameter Xi (t) of only the background noise section among the spectral feature parameters Xi (t) as the learning noise spectrum Bi.

On the other hand, the pattern recognition unit 104 discriminates the speech standard pattern and the noise spectrum at the time of learning from the spectrum characteristic parameter Xi (t), and stores it inside itself.

(A-3) Effects of the Embodiment According to the speech processing apparatus of this embodiment, the estimated speech spectrum calculator 111 subtracts the estimated noise spectrum component k · Ai from the spectrum feature parameter Xi (t). When XSi (t) is equal to or smaller than the noise spectrum Bi at the time of learning, processing is performed to substitute the noise spectrum Bi at the time of learning into the estimated speech spectrum XSi (t).

The spectrum Bi of the background noise section when learning the spectrum series of the background noise section in the estimated speech spectrum series XS (t) (the section in which no speech is uttered, for example, the pause section such as a segment break). Can be matched. Therefore, in the comparison calculation between the estimated speech spectrum sequence XS (t) and the speech standard pattern in the pattern recognition unit 104, the matching error in the background noise section is eliminated, and the matching accuracy is increased, and as a result, the speech recognition performance is improved. .

FIG. 2 shows the results of evaluation experiments on the speech processing apparatus of this embodiment and a speech processing apparatus using the conventional SS method (hereinafter, referred to as a conventional type). In the experiment, a signal in which 25 types of noise were superimposed on a voice uttering 100 words was generated, and the speech recognition rates were compared. In the conventional type, the estimated speech spectrum sequence XS (t) was estimated by replacing the noise spectrum Bi at learning with k1 · A1 (k1 is a predetermined constant) in the above equation (3) ((4) ) Expression).

XSi = Xi−k · Ai When Xi−k · Ai> k1 · Ai = k1 · Ai Xi−k · Ai ≦ k1 · Ai (4) Regarding the recognition rate, the voice power and the noise power And the ratio (S
/ N ratio) is 83.2% in the conventional type when it is 20 dB, 87.1% in this embodiment, and 65.7% in the conventional type when it is 15 dB.
1.0%, and at 10 dB, 47.0 in the conventional type.
% In this embodiment, 49.7%, and 5 dB
Is 31.5% in the conventional type, and 33.1% in this embodiment.

In any case, the recognition rate of the speech processing apparatus of this embodiment is higher than the recognition rate of the conventional speech processing apparatus, especially when the S / N ratio is 15 dB. As a result, the recognition rate exceeds 5% or more, and the effectiveness of the voice processing device of this embodiment can be evaluated.

Further, the estimated speech spectrum sequence XS
The amount of calculation for calculating (t) is the same for the voice processing device of this embodiment and the conventional voice processing device.

(B) Other Embodiments In the above embodiment, the BPF is used for the spectrum analysis processing in the acoustic analysis unit 101. However, the present invention is now mainly used for the spectrum analysis processing. L
For the characteristic parameters of the PC cepstrum system, for example, the discrete cosine transform (DCT) is used.
This can be easily applied by transforming the data into the frequency spectrum domain using, for example, a sform.

The present invention is characterized by a method for estimating an estimated speech spectrum, and any method may be used for pattern recognition. For example, a pattern recognition method using a spectrum feature parameter sequence, a Hidden Markov Model (HMM), or the like may be applied as the voice standard pattern.

Further, in the above-described embodiment, the case where the speech processing apparatus of the present invention newly updates the learning noise spectrum for each learning has been described. However, the average value (simple average) of the learning noise spectrum obtained for each learning is shown. Or a weighted average) may be used as the noise spectrum at the time of learning, or the noise spectrum at the time of learning may be determined by forgetting learning.

Further, in the above embodiment, the speech processing apparatus for recognizing the type of the speech of the input speech signal has been described. However, the present invention can be applied to a speech processing apparatus for discriminating the speech of a person.

In the above-described embodiment, the pattern recognition unit 104 recognizes a speech standard pattern necessary for recognizing the speech type of an input speech signal and a spectrum of only background noise.
In addition, there is a learning mode in which the noise spectrum storage unit 106 at the time of spectrum learning of only background noise is learned, but these learnings may be performed by another device. That is, after performing the learning on the pattern recognition unit 104 and the learning noise spectrum storage unit 106 by another device,
A method incorporated in this embodiment may be used.

[0049]

As described above, according to the speech processing apparatus of the present invention, when the difference between the spectrum feature parameter and the estimated noise spectrum is equal to or smaller than the reference pattern noise spectrum value, Since the process of substituting the reference pattern noise spectrum into the spectrum feature parameter is performed, the following effects are produced.

Even when the estimated noise spectrum value exceeds the true noise spectrum value, the estimated speech spectrum value can be prevented from becoming negative, and the speech recognition performance of the pattern recognition means can be improved.

Further, the spectrum of the background noise section of the estimated speech spectrum can be made to match the spectrum of the background noise at the time of learning. Then, in the comparison calculation between the estimated speech spectrum and the speech standard pattern, the matching error in the background noise section is eliminated, and as a result, the speech recognition performance is improved.

[Brief description of the drawings]

FIG. 1 is a functional block diagram of an embodiment.

FIG. 2 is a table for explaining effects of the embodiment.

[Explanation of symbols]

101: Acoustic analysis unit, 102: Voice detection unit, 104: Pattern recognition unit, 105: Estimated noise spectrum storage memory, 106: Learning noise spectrum storage unit, 110: Estimated noise spectrum calculation unit, 111: Estimated voice spectrum calculation unit .

Claims (2)

[Claims] 1. An audio processing apparatus for analyzing a spectrum of an input audio signal including a section in which no audio is input, obtaining a spectrum characteristic parameter, and processing the input audio signal. Sound analysis means for converting to spectral feature parameters and calculating sound power; sound detection means for detecting a sound start frame, which is the first frame of a section in which sound is being input, from sound power; Estimated noise spectrum calculating means for estimating a noise spectrum from a spectral feature parameter of a section from a frame to a frame which is retroactive for a predetermined time and outputting an estimated noise spectrum; and estimating holding the estimated noise spectrum output from the estimated noise spectrum calculating means. Noise spectrum holding means and background noise only A reference pattern noise spectrum holding means for holding a spectrum feature parameter as a reference pattern noise spectrum; and calculating an estimated speech spectrum by a difference between the spectrum feature parameter and the estimated noise spectrum, wherein the estimated speech spectrum value is equal to or less than the reference pattern noise spectrum value. In the case of, an estimated speech spectrum calculation means for using the reference pattern noise spectrum as an estimated speech spectrum, and a speech characteristic parameter and a spectrum feature parameter of only background noise are internally provided as reference parameters, and the estimated speech spectrum and A speech recognition device comprising: pattern recognition means for identifying an input speech by comparing the reference parameter with the internal reference parameter. 2. The reference pattern noise spectrum holding means holds a spectrum characteristic parameter of only background noise output from the acoustic analysis means as a reference pattern noise spectrum when learning reference parameters necessary for performing speech processing. The above-mentioned pattern recognition means, at the time of learning the reference parameters necessary for performing the voice processing, the voice standard pattern which is a spectrum characteristic parameter of the voice, and the spectral feature parameter of only the background noise output from the above-mentioned acoustic analysis means. 2. The input speech is identified by storing the reference speech parameter as the reference parameter in the self, and comparing the estimated speech spectrum with the reference parameter stored in the self during input speech processing. An audio processing device according to claim 1.