JP3032215B2 - Sound detection device and method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Purpose of the Invention] (Industrial application field) The present invention is applied to the fields of ATM (Asynchronous Transfer Mode) communication, DSI (Digital Speech Interplation), packet communication, and speech recognition. The present invention relates to a sound detection device that accurately detects a sound section in a signal.

(Prior Art) FIG. 6 shows one configuration of a conventional sound detection device.

Feature parameters such as power, the number of zero crossings, an autocorrelation function, and a spectrum are calculated by the feature parameter calculator 101 from the audio signal input to the input terminal 100 in frame units.

The calculated feature parameters are output to the matching unit 102, and are compared with the preset sound standard pattern 103 and noise standard pattern 104 to calculate the respective distances.

If the distance between the feature parameter and the sound standard pattern 103 is smaller than the distance between the feature parameter and the noise pattern 104, the input frame is determined to belong to voice, and if the distance is opposite, it is determined to belong to noise. Output from the output terminal 105.

(Problems to be Solved by the Invention) However, even if there is a sound, the power of consonants is often lower than the power of background noise, unlike vowels. For this reason, in an environment where the background noise is large, the feature of the background noise appears largely in the feature parameter of the consonant section.

According to the above-described conventional sound detection device, the characteristic parameter affected by the background noise is used for the determination as it is. Therefore, when the background noise is large, the detection error of the consonant is increased.

This has caused deterioration in sound quality in the field of communication, and has also led to a decrease in the recognition rate in the field of speech recognition.

The present invention has been made in view of the above circumstances, and an object of the present invention is to provide a voice detection device that can improve the detection accuracy of a sound even when background noise is large.

[Structure of the Invention] (Means for Solving the Problems) In order to solve the above problems, the present invention provides a feature parameter for obtaining a feature parameter of an input audio signal input in units of frames divided for each length. Generating means, noise determining means for temporarily determining, on a frame-by-frame basis, whether or not the input speech signal is noise, based on the characteristic parameters obtained by the characteristic parameter generating means; A storage unit for storing a plurality of frames of the feature parameters of the frame temporarily determined to be obtained by the feature parameter generation unit, and using the feature parameters for the plurality of frames stored in the storage unit to generate the input voice. Conversion means for converting the characteristic parameters of the signal frame into conversion parameters; Based on the conversion parameters, the frame of the input audio signal is characterized by comprising a sound presence judgment means for judging whether belonging to or noise belongs to the voice.

(Operation) In the configuration described above, in the present invention, a conversion parameter is generated based on all the characteristic parameters obtained in frame units or the characteristic parameters of the noise section, and the conversion parameter is used to generate the sound section of the audio signal. The noise section can be determined. In particular, when a conversion parameter is generated based on a feature parameter in a noise section, it is possible to perform sound discrimination while avoiding the influence of noise.

(Embodiment) FIG. 1 is a block diagram showing a schematic configuration of a sound detection device according to the present invention. This device includes a feature parameter calculator 1, a feature parameter converter 2, and a sound determination device 3. When,
It comprises a noise detector 4, a switch 5, and a buffer 6.

In the following embodiment, an audio signal is analyzed on a frame basis, and the presence / absence and audio are determined. For example, the audio signal is sampled at 8 KHz, and 160 samples are combined to form one frame. However, the frame length does not need to be always constant.

In the feature parameter calculator 1, the Durbin
The linear prediction coefficient is calculated using the method or the like. Here, a PARCOR coefficient, an LPC cepstrum, a mel cepstrum, or the like may be calculated from the linear prediction coefficient, and may be used as a feature parameter. Also, the power, the autocorrelation function, the number of zero crossings, and the like may be calculated.

The frame for which it is currently determined that there is sound or silence is hereinafter referred to as an input frame. Further, the characteristic parameters of the input frame obtained by the characteristic parameter calculator 1 are And n is the sequential number of the frame.
The feature parameter is a p-dimensional vector and is represented by the following equation (1).

The noise detector 4 measures the average power Pow by the following equation (3) for each frame. If the sample of the audio signal in the frame is a (i) (i = 0, 1,..., S−1) and the number of samples in one frame is s, Then, a threshold value T given in advance and the average power Pow are compared with each other in order to reliably detect a section in which noise is present from the input signal.

If Pow ≧ T, it is determined that the noise is not a noise, and “0” is output to SW5.

 Otherwise, it is determined to be noise and "1" is output to SW5.

SW5 is the buffer 6 if the output of the noise detector is "1".
To store the feature parameters of the frame.

In the buffer 6, as shown in FIG. 2, in order to preserve the order of the time when the characteristic parameters are accumulated in the buffer 6, in order of the characteristic parameters to be input to the buffer, the tail from the head of the buffer is tailed. Accumulate towards. That is, the newest feature parameter (the feature parameter of the frame to be determined at present) is stored in the head of the buffer, and the oldest feature parameter is stored in the tail.

In this embodiment, only the characteristic parameters of the frame determined to be noise by the noise detector 4 are stored in the buffer. However, all the characteristic parameters may be stored in the buffer without performing the noise determination.

From among the feature parameters stored in the buffer 6, a feature parameter set Ω for N frames is extracted from S frames before the input frame (from the head of the buffer to the S frame) toward the tail of the buffer, and as shown in FIG. To And

The S frame and the N frame can take an arbitrary number of frames, but it is preferable that the number of frames is from about several frames to about 20 frames.

The feature parameter converter 2 converts feature parameters to emphasize the difference between speech and noise. Here, the converted feature parameters are referred to as And the conversion parameters Is a p-dimensional vector.

Here, the conversion parameters Is the mean vector of Ω And input frame feature parameters , The distance vector is calculated and normalized by the standard deviation of Ω. Each component is represented by the following equations (3) to (7), and FIG. Feature parameter set Ω, conversion parameter Mean vector of Ω Is shown.

Then, y _i (n) = (x _i (n) −m _i ) / σ _i (5) Here, i = 1, 2,..., P.

The sound existence determiner 3 determines a sound period based on the conversion parameter obtained from the feature parameter converter 2. As shown in FIG. 4, this sound existence judgment device 3 includes a matching device 7 and
And M standard patterns 8.

The standard pattern 8 can be defined as follows. Standard pattern 8 Mean vector μ of and Is the covariance matrix. In the following equations (8) to (10), i indicating the class of the standard pattern is omitted for simplicity.

Let L p-dimensional conversion parameters belonging to class ω be Let μ _k and Σ _kl be the elements of μ and Σ, respectively. It is expressed as

In the matching unit 7, the standard pattern ω _i and the conversion parameter If the distance is measured, is matched to the standard pattern omega _i belonging to the speech determines the speech, a silent otherwise.

First, according to the following equation, each standard pattern ω _i (i = 1,. Distance to Is measured.

this Is used to calculate the class i, Becomes by this, Belongs to ω _i of class i. If ω _i is a pattern representing voice, the frame is voiced, ω
_{If i} is a pattern representing noise, the frame is determined to be noise.

The effects of the above embodiments will be described based on specific measurement results.

Unlike vowels, the power of consonants is often lower than the power of background sounds. Therefore, in an environment where the background noise is large, the feature of the noise appears largely in the feature parameter even in the consonant section. In the conventional method, the characteristic parameter affected by the background noise is used for the determination as it is. Therefore, when the background noise is large, the detection error of the consonant is increased.

In each embodiment of the present invention, a good detection rate was obtained even in an environment with a large background noise, such as an S / N ratio of about 20 to 14 dB, in order to emphasize noise and speech characteristics. The detection result of the initial consonant when the feature parameter / feature parameter conversion method is changed is shown below. When it is determined that the frame in which the label attached to the audio data indicates a consonant is one of the consonant classes, it is determined that the frame is correctly detected.

The detection rate shown in FIG. 5 is an average value of the consonant detection rate and the noise detection rate. The consonant detection rate is defined by the following equation.

When it is determined that the frame of the noise data is any of the noise classes, it is assumed that the noise data is correctly detected. This is the noise detection rate and is defined by the following equation.

In FIG. 5, the vertical axis is the detection rate. The horizontal axis indicates the type of feature parameter, LPC is LPC cepstrum, P is average power in a frame, and P + LPC is a combination of P and LPC.

In the following, the LPC cepstrum analysis dimension is 12th order, the transformation parameter dimension is 4th order when the feature parameter is LPC,
In the case of P + LPC, the fifth order was set. The feature parameter conversion method is
The plot is shown differently.

c is a conventional method that does not perform feature parameter conversion.

n is the embodiment shown in FIG. 1 and is for noise determination.

“v” represents the embodiment shown in FIG. 1 in which the noise is not determined.

[Effects of the Invention] As described above, according to the present invention, since the influence of noise can be removed from feature parameters by feature parameter conversion, it is possible to accurately determine a sound section even in an environment where background noise is large. Can be.

[Brief description of the drawings]

1 is a block diagram showing a schematic configuration of a sound detection device according to the present invention, FIG. 2 is a configuration diagram of a buffer used in the embodiment, FIG. 3 is an explanatory diagram of conversion parameters in the embodiment, FIG. 4 is a block diagram showing a configuration example of a sound presence detector, FIG. 5 is a characteristic diagram showing a relationship between a characteristic parameter and a detection rate in each embodiment, and FIG. 6 is a configuration example of a conventional sound presence detector. FIG. DESCRIPTION OF SYMBOLS 1 ... Feature parameter calculator 2 ... Feature parameter converter 3 ... Speech presence detector 4 ... Noise detector 5 ... Switch 6 ... Buffer 7 ... Matching device 8 ... Standard pattern

──────────────────────────────────────────────────続 き Continuation of front page (56) References JP-A-2-266400 (JP, A) JP-A-60-200300 (JP, A) JP-A-1-302298 (JP, A) JP-A-4- 58297 (JP, A) JP-A-61-48898 (JP, A) JP-A-2-282798 (JP, A) JP-A-2-26640 (JP, A) JP-A-3-48900 (JP, A) Special Publication Hei 5-56512 (JP, B2) 1989 IEICE Spring National Convention Lecture Papers, Third Volume, p. 3-78 "B-372 Speech Cellization Method for ATM Communication" (March 28, 1989) Furui "Digital Speech Processing" (1985-9-25) Tokai University Press p. 44-48 Saito and Nakata, "Basics of Speech Information Processing" (56-11-30), Ohmsha p. 99-103 IEICE Technical Report [Communication] Vol. 89, No. 132, CS89-33, "Sound detection method for voice packet communication" p. 61-66 (Issued July 19, 1989) (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 11/02 G10L 15/04 H04B 14/04 JICST file (JOIS)

Claims (3)

(57) [Claims] 1. A feature parameter generating means for obtaining a feature parameter of an input audio signal input in units of a frame delimited by a certain length, based on the feature parameter obtained by the feature parameter generating means. Noise determining means for temporarily determining, for each frame, whether or not the input speech signal is noise; and a plurality of feature parameters obtained by the feature parameter generating means for the frame temporarily determined to be noise by the noise determining means. Accumulating means for accumulating frames, and converting means for converting the characteristic parameters of the frame of the input audio signal into conversion parameters using the characteristic parameters for the plurality of frames stored in the accumulating means. The frame of the input audio signal belongs to the audio based on the converted And a sound determining means for determining whether the sound belongs to noise or noise. 2. The method according to claim 1, wherein the converting unit obtains a distance vector between a feature parameter of the frame of the input voice signal and a feature parameter of the plurality of frames stored in the storage unit, thereby obtaining a frame of the frame of the input voice signal. 2. The sound detection device according to claim 1, wherein a characteristic parameter is converted into the conversion parameter. 3. A feature parameter generating step for obtaining a feature parameter of an input audio signal input in units of frames divided by a certain length, based on the feature parameter obtained in the feature parameter generating step. A noise determination step of temporarily determining whether or not the input speech signal is noise for each frame; and a plurality of feature parameters obtained by the feature parameter generation step of the frame temporarily determined to be noise by the noise determination step. An accumulating step of accumulating frames, a converting step of using the characteristic parameters of the plurality of frames accumulated in the accumulating step to convert a characteristic parameter of a frame of the input voice signal into a conversion parameter, Based on the conversion parameters A sound determination step of determining whether a frame of the input voice signal belongs to voice or noise.