JPH04230798A - Noise predicting device - Google Patents

Abstract

PURPOSE:To improve a noise suppression ratio and to improve the quality of output voices by predicting the noise of signal parts with good accuracy. CONSTITUTION:Signals are inputted to a band dividing means 1 and the presence or absence of voices is discriminated by a signal discriminating means 3. The noise is predicated by a noise predicting means 21. The noise predicted by this noise predicting device 21 is subtracted by a canceling means 4 by which the noise components are cut. The signals are outputted after band synthesis with a band synthesizing means 5. The threshold component of the noise components to be canceled at the time when the canceling means 4 cancels the noise is controlled by an attenuator 22 and an attenuation coefft. setting means 23, by which the excess noise suppression is eliminated and the distinctness of the voice signals is improved.

Description

[Detailed description of the invention]

0001

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a noise prediction device for predicting noise in a noise mixed signal such as a voice signal mixed with noise.

0002

2. Description of the Related Art Conventionally, techniques have been developed for predicting noise in audio signals, removing the noise, and obtaining audio signals of excellent quality. In such a technique, an important point is a method of predicting noise contained in a signal portion. For example, a method is known in which a noise-containing audio signal is Fourier transformed and noise in that signal portion is predicted based on past noise information. In this method, in order to predict the noise contained in a signal part, the noise data of the part where only noise exists before the part where the signal exists is retained as is, and based on the retained data, the noise data contained in the signal part is predicted. It predicts the noise that will occur.

0003

[Problem to be solved by the invention] However, with this prediction method, noise data from a long time ago is used without change, so the prediction of noise in the signal part tends to be rough and inaccurate. be.

[0004] It is an object of the present invention to solve the problems of such conventional noise prediction devices.

[0005]

[Means for Solving the Problems] A noise prediction device according to the present invention collects past noise data for a predetermined period of time when inputting a mixed signal and predicting noise according to a predetermined prediction method.
The present invention is characterized by comprising a noise prediction means for calculating a calculated noise prediction value by taking into account the amount of change in past noise data.

Furthermore, the noise prediction device according to the present invention has a signal discrimination means for discriminating between a portion where a signal exists and a portion where only noise exists, and an output noise prediction value of the signal portion obtained by the discrimination. and a noise prediction means for attenuating the noise.

[0007]

[Operation] Since the present invention takes into account the amount of change in past noise data, it is possible to obtain a calculated noise prediction value that is closer to the actual noise. Further, the signal discriminating means inputs a mixed signal including noise, detects the presence of a signal in the mixed signal, and calculates the predicted output noise value of the signal portion determined by the signal discriminating means into the predetermined prediction. The calculated noise prediction value obtained by the method is assumed to be smaller than the predicted value of the calculated noise.

[0008]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Examples of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram of a signal processing device that uses an embodiment of the noise prediction device of the present invention to obtain a signal of good quality. The configuration and operation will be explained below.

In the figure, band dividing means 1 inputs a signal, which is a mixture of voice and noise, as an example of a signal, and
This is a means of performing Fourier transformation after D transformation.

The signal determining means 3 receives a mixed signal, for example, a mixed signal of a voice signal and noise, from the band dividing means 1, and detects the presence or absence of a signal portion, in this example a voice signal portion, in the mixed signal. It is a means. The presence or absence of such a signal portion can be determined by, for example, the cepstrum (CEPSTRUM) described later.
This is done using analysis.

The noise prediction means 2 receives the mixed signal from the band division means 1, collects past noise data for a predetermined period, and calculates a noise prediction value by taking into account the amount of change in the past noise data. be. Furthermore, in a preferred embodiment, the noise prediction means 2 includes attenuation means for attenuating the noise prediction value during the period in which the signal portion is determined to be present. The noise prediction means 2 equipped with attenuation means will be explained below.

As shown in FIG. 2, this noise prediction means 2 is a noise predictor 2 which receives the mixed signal from the band division means 1 and predicts the noise in the signal portion according to a predetermined prediction method.
1, an attenuation coefficient setting means 23 which inputs the output of the signal discriminating means 3 and sets an attenuation coefficient based on the output; The attenuator 22 attenuates the calculated noise prediction value predicted by the noise predictor 21 according to the attenuation coefficient set by the setting means 23.

The noise predictor 21 inputs a Fourier-transformed signal mixed with noise, collects past noise data for a predetermined period, and predicts the noise in the signal portion based on, for example, the average value of the noise data. Figure 7 shows how the data was collected. The x-axis shows frequency, the y-axis shows noise level, and the z-axis shows time. Past noise data p1 for a predetermined period at frequency f1,
Take p2, . . . , pi and predict the next pj. As an example of a prediction method, the average of the noise parts p1 to pi is taken as pj. FIG. 6(a) shows a calculated noise prediction value predicted and calculated by such a noise prediction method. The period between t1 and t2 is a portion where a voice signal and noise coexist, and the rest is a portion where only noise exists. The calculated noise prediction value of the portion where the audio signal exists is a value roughly predicted based on the noise prediction value data and actual noise data of the portion where only noise exists before t1.

The attenuation coefficient setting means 23 is means for setting an attenuation coefficient during the period (t1 to t2) in which the signal discriminating means 3 determines that the signal portion is present. This damping coefficient is
As shown in FIG. 8(a), the coefficient value gradually increases as time passes, corresponding to the period (t1 to t2) in which the signal portion is determined to be present.
It is desirable to change according to the curve. but,
The shape is not limited to this, and may change like a hyperbola (b), a downward arc (c), a stepped line (d), or the like.

The attenuator 22 has this attenuation coefficient setting means 23
Using the attenuation coefficient set by the noise predictor 21, the period (t1
~t2) is a means for attenuating calculation prediction noise. That is, the calculated predicted noise value at time t1 is multiplied by the attenuation coefficient at time t1. Even after t1, the corresponding damping coefficient is multiplied in the same way. Therefore, when an exponential curve-shaped attenuation coefficient is used, the output noise prediction value is almost the same as the calculated noise prediction value, as shown in FIG. 6(b), and then the output noise prediction value gradually increases 6(a) represents the input of the attenuator 22, while the solid line in FIG. 6(b) represents the output of the attenuator 22.

[0017] As a result, the predicted noise value of the signal portion becomes sufficiently small. Therefore, even if the calculated noise prediction value is rough, even if the noise is canceled using the attenuated calculated noise prediction value as described later, it will have little effect on the speech signal, so the intelligibility of the speech signal will be improved. No need to worry about losing. Moreover, since the calculated noise prediction value uses noise data of only the noise before the signal part, the inaccuracy increases as time passes, so it is necessary to increase the amount of attenuation as time passes. is desirable, but it can be appropriately realized in the case of an exponential damping coefficient. Note that at the end time t2 of the signal portion,
Although the calculated noise prediction value is considerably attenuated, after that, t
1, as before, the noise prediction method calculates t2+ based on the actual noise and the predicted output noise value at t2.
As the noise at time α is predicted, it gradually approaches the calculated noise prediction value in the case of no attenuation. Note that, as shown in the figure, the predicted value is attenuated in the same manner not only when the calculated noise predicted value increases but also when it decreases. Similarly, when other attenuation coefficients shown in FIG. 8 are used, the calculated noise prediction value is attenuated by a predetermined amount.

The canceling means 4 is a means for subtracting the predicted noise value predicted by the noise predicting means 2 from the Fourier-transformed signal mixed with noise. For example, it is configured with a subtracter.

The band synthesizing means 5 is a means for synthesizing and outputting the signal whose prediction noise has been canceled by the canceling means 4 in this way. For example, inverse Fourier transform is performed and then D/A signal conversion is performed.

In FIG. 1, (a) is the waveform of the input signal mixed with noise (see FIG. 9 (a)), and (b) is the waveform obtained by Fourier transforming that signal (see FIG. 9 (ro)). ), (c) is the predicted noise signal waveform, and (
D) is the waveform of the signal with noise canceled (Figure 9)
(d)) and (e) are waveforms obtained by inverse Fourier transform of this canceled signal (see FIG. 9(e)). In the present invention, the noise prediction value is attenuated to an appropriate value during the period in which the signal portion is determined to be present, so that noise is not inaccurately canceled in the signal portion, and a high-quality signal can be obtained.

[0021] Note that it is also possible to discriminate the signal as it is as an analog signal without performing band division, to predict and attenuate noise, and to cancel the noise using the attenuated output noise prediction value.

In this embodiment, as shown in FIG. 1, a noise-containing signal is band-divided by a band division means 1, and a signal discrimination means 3 discriminates whether or not a signal portion exists in the signal. Then, regarding the band-divided noise-containing signal, the noise prediction means 2 attenuates the calculated noise prediction value during the period in which the signal discrimination means 3 determines that the signal portion is present, and outputs noise prediction. Used as a value. The canceling means 4 obtains a signal from which noise has been removed by subtracting the output noise prediction value predicted by this noise prediction value from the noise-containing signal. In addition, during the period where there is only noise other than the signal part, the attenuation is eliminated and the calculated noise is simply inverted.
You can just add it. The signals from which noise has been removed by the canceling means 4 are combined by the band combining means 5 and output as a target signal.

FIG. 3 is a block diagram showing another embodiment of the present invention. In this embodiment, a noise-containing signal is band-divided, and an attenuation coefficient setting means 23 and an attenuator 22 are provided for each of the divided frequency bands (hereinafter referred to as channels). Signal channel determination means 6 are provided for detecting the appearing channel. For channels where the audio signal appears strongly, the attenuation coefficient setting means 23 is set to the driven state, while for channels where the sound signal does not appear strongly, the attenuation coefficient setting means 23 is set to the non-driven state or close to it. state. That is, the audio channel determining means 6 is a means for determining the level of the audio signal for each of a plurality of channels, and the attenuation coefficient setting means 2 provided corresponding to each channel.
3 changes the attenuation coefficient according to the audio signal level of that channel.

In the above embodiment, although the attenuation coefficient changed over time, the change was unrelated to the change in the level of each audio frequency channel. but,
In this embodiment, the attenuation coefficient is optimally changed also in response to changes in the level of each audio frequency channel. For example, for a channel with a low audio level, the attenuation coefficient is decreased and the output noise prediction value is increased to sufficiently cancel the noise from the signal, while for a channel with a high audio level, the attenuation coefficient is increased and the output noise Reduce the predicted value to avoid canceling too much noise from the signal. Other means are similar to the previous embodiment.

FIG. 4 is a block diagram showing another embodiment of the present invention, which is different from the embodiment of FIG. 3 in that the level of the audio channel is determined directly from the noise-containing signal before band division. There is a difference. The methods include autocorrelation method, L
These include PC analysis method and PACOR analysis method. The audio channel determining means 7 is a means for inputting a signal mixed with noise and determining the level for each audio channel.

[0026] According to PACOR analysis or the like, it is possible to extract the spectral envelope of the input signal and the periodicity of the sound source. This is based on Durbin's solution, a lattice circuit,
It is determined by a known method such as the deformed lattice method or the Le Roux method. Then, using this spectral envelope and periodicity information of the sound source, the level of each audio frequency channel may be determined according to the number m of band divisions of the system. P.A.
COR analysis, LPC analysis, and autocorrelation methods also allow frequency information to be obtained from calculations on the time axis, so channel classification can be performed arbitrarily according to the system.

Furthermore, as a modification, the input to the signal channel determining means 6 of FIG. 3 is not from the band dividing means 1,
The signal may be received from the signal determining means 3.

Next, the signal discrimination means 3 will be explained in detail. In the example shown in FIG. 5, using the cepstral analysis method,
The presence or absence of an audio signal portion is determined, and the channel in which the audio signal appears strongly is determined. In FIG. 5, the cepstrum analysis means 8 is a means for performing cepstrum analysis on the signal that has been Fourier transformed by the band division means 1. The cepstrum is the inverse Fourier transform of the logarithm of the short-time amplitude spectrum of a waveform, as shown in FIG. FIG. 10(a) is a short-time spectrum, and FIG. 10(b) is its cepstrum.

The peak detection means 9 detects the cepstrum peak (P) determined by the cepstrum analysis means 8.
) to distinguish between signals and noise. Where the peak exists, it is determined that an audio signal portion exists. The peak can be detected, for example, by determining a threshold value of a predetermined size in advance and comparing it with the threshold value. Further, the pitch frequency detecting means 10 is a means for obtaining a pitch frequency by calculating the reciprocal of the quefrency value (obtained from FIG. 10(b)) of the peak detected by the peak detecting means 9. For example, this quefrency value can be obtained by Fourier transformation, and the obtained pitch frequency indicates which channels have a strong audio signal and which channels do not. The audio channel calculation means 11 detects the signal level of each channel and determines which channels have strong audio signals and which channels do not. The cepstrum analysis means 8, the peak detection means 9, the pitch frequency detection means 10, and the audio channel calculation means 11 constitute the audio channel determination means 6. In addition, cepstrum analysis means 8
and the peak detection means 9 constitute the voice discrimination means 3.

FIG. 11 is a block diagram of a signal processing apparatus showing the signal discriminating means 3 in more detail, and shows an embodiment in which no attenuating means is included. In FIG. 11, the signal discrimination means 3
is a cepstrum analysis means 10 for performing cepstrum analysis.
2. Peak detection means 103 for detecting the peak of the cepstrum distribution, means 104 for calculating the average value of the cepstral distribution,
It has vowel/consonant determining means 105 for determining vowels and consonants, signal determining means 106, and noise section setting means 108. That is, the FFT means in the band division means 1 fast Fourier transforms the audio signal input, and the cepstrum analysis means 102
supply to Cepstrum analysis means 102 obtains the cepstrum of the spectrum signal and supplies it to peak detection means 103 and average value calculation means 104. This is shown in FIGS. 12(a) and 12(b).

The peak detection means 103 determines the peak of the cepstrum obtained by the cepstrum analysis means 102 and supplies it to the vowel/consonant determination means 105 .

On the other hand, the average value calculation means 104 calculates the average value of the cepstrum obtained by the cepstrum analysis means 102, and supplies it to the vowel/consonant determination means 105. vowel/
The consonant determining means 105 determines vowels and consonants of the audio signal input using the peak of the cepstrum supplied from the peak detecting means 103 and the average value of the cepstrum supplied from the average value calculating means 104, and outputs the determination result. That is.

The speech signal discriminating means 106 is a means for discriminating, based on the result of the vowel/consonant discriminating means 105, that the vowel part and the consonant part are speech intervals.

The noise section setting means 108 is a means for inverting the output of the speech discriminating means 106 to find and set a section containing only noise.

The operation of the embodiment shown in FIG. 11 described above will now be described. The input audio signal mixed with noise is fast Fourier transformed by FFT means 1, and then subjected to cepstrum analysis means 10.
2, the cepstrum is obtained, and the peak detection means 10
3, the cepstrum peak is determined. Further, the average value of the cepstrum is calculated by the average value calculating means 104. When the vowel/consonant determining means 105 receives a signal indicating that a peak has been detected from the peak detecting means 103, it determines that the audio signal input is in a vowel section. Regarding the determination of consonants, for example, if the cepstrum average value inputted from the average value calculation means 4 is larger than a predetermined value, or if the amount of increase (differential coefficient) of the cepstrum average value is larger than a predetermined value, If it is larger than the value, it is determined that the audio signal input is in a consonant section. As a result, a signal indicating a vowel/consonant or a signal indicating a speech section including a vowel and a consonant is output. The speech discrimination means 106 uses the consonant and vowel speech section signals to
Determine the voice section. The noise section setting means 108 sets a section other than the voice section as a noise section. Noise prediction means 2
predicts the noise in the noise interval set by the noise interval setting means 108 for the band-divided noise-mixed signal. Then, in the canceling means 4, the noise is canceled.

[0036] Generally, as an example of a cancellation method,
Cancellation on the time axis, as shown in FIG. 13, subtracts the predicted noise waveform (b) from the noise-containing audio signal (a). As a result, only the signal is extracted (c). In addition, in this embodiment, as shown in FIG. 14, the noise-containing speech signal (A) is Fourier transformed (B) using frequency-based cancellation, and then the predicted noise spectrum (C) is subtracted (N). ), it is inversely Fourier transformed to obtain a noise-free audio signal (e). The band synthesizing means 5 is a means for obtaining a high quality audio output by performing inverse Fourier transform on the m-channel signals supplied from the canceling means 4 as described above. In this way, the noise section and the speech section are clearly distinguished, noise prediction is performed, and highly accurate prediction noise is removed from the signal, so that high-quality speech can be obtained.

The vowel/consonant determining means 105 in the embodiment shown in FIG. 11 has a configuration consisting of means 151 to 154. The first comparator 152 includes the peak detection means 10
This circuit compares the peak information obtained in step 3 with a predetermined threshold set by the first threshold setting means 151 and outputs the result. Further, the first threshold setting means 151 is means for setting a threshold value according to the average value obtained by the average value calculation means 104.

Further, the second comparator 153 compares a predetermined threshold value set by the second threshold setting means 154 and the average value obtained by the average value calculation means 104,
This is a circuit that outputs the results.

Further, the vowel/consonant determination circuit 155
The comparison result obtained by the comparator 152 and the second comparator 153
This circuit determines whether the input audio signal is a vowel or a consonant based on the comparison results obtained in the above.

Next, the operation of this vowel/consonant determining means 105 will be explained. The first threshold setting means 151 sets a threshold value as a reference for determining whether the peak obtained by the peak detection means 103 is a peak sufficient to be determined to be a vowel. At that time, the threshold value is determined by referring to the average value obtained by the average value calculation means 104. For example, if the average value is large, the threshold is set high to ensure that a peak indicating a vowel can be selected.

The first comparator 152 compares the threshold set by its threshold setting means 151 with the peak detected by the peak detecting means 103, and outputs the comparison result.

On the other hand, the second threshold setting means 15
4 sets a predetermined threshold value. This may be a threshold value for the average value itself, or a threshold value for a differential coefficient that indicates an increasing tendency of the average value. Then, the second comparator 153 calculates the average value obtained by the average value calculation means 104 and the second threshold setting means 15.
It compares it with the threshold value set in step 4 and outputs it. That is,
The calculated average value and the threshold average value are compared, or the increased value of the calculated average value and the threshold differential coefficient value are compared.

The vowel/consonant determination circuit 155 determines vowels and consonants based on the comparison results of the first comparator 152 and the second comparator 153. If a peak is definitely detected in the comparison result of the first comparator 152, that area is determined to be a vowel. Further, in the comparison result of the second comparator 153, if the average value exceeds the average value of the threshold value, that area is determined to be a consonant. Alternatively, the increase in the average value is compared with the differential coefficient of the threshold value, and if it exceeds the threshold value, it is determined that it is a consonant.

[0044] The vowel/consonant determining means 105 takes into account the nature of the vowel and consonant sections of speech, for example, the property that speech is composed of consonants + vowels, and determines when the consonant section and vowel section are aligned. The determination threshold as a consonant may be issued starting from the first consonant interval. That is, in order to more reliably distinguish between noise and consonants, even if a consonant is determined based on the average value, it is determined to be noise if there is no subsequent interval.

FIG. 15 shows an application example of the present invention, in which FIG.
An example will be described in which speech recognition is performed using high quality speech obtained by the embodiment of the present invention shown in 1. That is,
After the band synthesis means 5, each word or "A", "I"
, "U", etc., is connected to a voice extraction means 111 that extracts each syllable, and furthermore, each phoneme, and then a feature extraction means 11 that extracts features of the extracted syllable, etc.
2 are connected, and then the feature comparison means 114 compares the extracted features with standard features such as standard syllables stored in advance in the storage means 113 and recognizes the type of the syllable. It is connected. In this way, in this embodiment of speech recognition, speech recognition is performed on speech from which noise has been accurately predicted and removed, so that the speech recognition rate is particularly high.

In the embodiments described above, many of the means such as the signal discrimination means, the noise prediction means, and the cancellation means can be realized in software using a computer, but it is also possible to use dedicated hardware circuits having the respective functions. It is also possible.

[0047] In the present invention, noise means a signal other than the signal of interest, so even speech may be treated as noise.

[0048]

[Effects of the Invention] As is clear from the above explanation, the signal processing device according to the present invention detects the vowel/consonant area of a speech signal mixed with noise, and based on that detects the noise area. Since the noise is predicted based on the settings, the predicted noise is highly accurate.

Furthermore, since the noise is removed using the highly accurate predicted noise, a signal of good quality can be obtained.

[0050] Furthermore, since speech recognition is performed using the high quality signal, the speech recognition rate is increased.

Furthermore, according to the present invention, for the signal part, a noise value smaller than the noise prediction value calculated by a predetermined noise prediction method is set as the noise value, so that subsequent processing, for example, the voice part In this case, there is no concern that noise will be largely canceled, and there is no concern that the clarity will decrease significantly due to noise removal.

[Brief explanation of the drawing]

FIG. 1 A block diagram showing an embodiment of a noise prediction device according to the present invention.

[Figure 2] Block diagram showing more specific details of the embodiment

[Figure 3] Block diagram showing another embodiment of the present invention

[
FIG. 4 Block diagram showing another embodiment of the present invention

[Figure 5
] Block diagram showing another embodiment of the present invention

[Figure 6]
Graph showing calculated noise prediction values and output noise prediction values according to a predetermined method in an embodiment of the present invention

[Figure 7] Graph for explaining general noise prediction method

[Fig. 8] Graph showing the damping coefficient in one embodiment of the present invention

[Fig. 9] Graph for explaining the operation in one embodiment of the present invention

[Figure 10] Graph for explaining general cepstral analysis

[Fig. 11] Block diagram showing details of signal discrimination means 3

[Figure 12] Graph showing cepstrum peaks in the present invention

[Figure 13] Waveform diagram for explaining the cancellation method of the present invention

[Figure 14] Waveform diagram for explaining the cancellation method of the present invention

[Figure 15] Block diagram showing an example of application of the present invention

[Explanation of symbols]

1 Band division means 2 Noise prediction means 3 Signal discrimination means 4 Cancellation means 5 Band synthesis means 6 Signal channel judgment means 7 Signal channel judgment means 8 Cepstrum analysis means 21 Noise predictor 22 Attenuator 23 Attenuation coefficient setting means 103 Peak detection means 104 Average value calculation means 105 Vowel/consonant determination means 106 Signal discrimination means 108 Average value calculation means 113 Storage means 114 Feature comparison means 151 First threshold setting means 152 First
Comparator 153 Second comparator 154 Second threshold setting means 155 Vowel/consonant determination circuit

Claims (10)

[Claims] 1. Signal discrimination means for inputting a mixed signal containing noise and detecting the presence of a signal in the mixed signal; 1. A noise prediction device comprising: noise prediction means for reducing an output noise prediction value of a signal portion determined by the discrimination by the signal discrimination means, compared to a calculated noise prediction value obtained by the prediction method. 2. The noise prediction according to claim 1, wherein the output noise prediction value decreases with respect to the calculated noise prediction value as it approaches the end of the signal portion, rather than the initial value of the signal portion. Device. 3. The degree of reduction of the output noise predicted value with respect to the calculated noise predicted value is determined for each frequency channel in accordance with the level of the noise-mixed signal for each frequency channel. The noise prediction device according to claim 1. 4. The noise prediction device according to claim 3, wherein the level for each frequency channel is determined based on a signal obtained by dividing a noise-containing signal into bands. 5. The noise prediction device according to claim 3, wherein the level for each frequency channel is determined directly from the noise-mixed signal. 6. Band division means for dividing a noise-containing audio input signal into bands; cepstrum analysis means for performing cepstrum analysis on the output of the band division means; and a peak for detecting a cepstrum peak in the cepstrum analysis output of the cepstrum analysis means. a detection means; an average value calculation means for calculating an average level in the cepstrum analysis output of the cepstrum analysis means; a vowel/consonant determining means for determining vowels and consonants by determining vowels based on the level of the average value information, and a vowel/consonant determining means for determining vowels and consonants based on the level of the average value information; a noise section setting section for setting a noise section using the result of the speech discriminating section; and a noise section setting section for setting a noise section using the result of the speech discriminating section; 1. A signal processing device comprising: noise prediction means for predicting noise in a voice section using noise data in the noise section obtained by 7. A method comprising a canceling means for subtracting the noise predicted by the noise predicting means from the band-divided signal by the band dividing means, and a band synthesizing means for band-synthesizing the output of the canceling means. 7. The signal processing device according to claim 6. 8. The signal according to claim 7, wherein speech recognition is performed by cutting out speech from the signal band-synthesized by the band synthesis means, extracting features, and comparing with reference features. Processing equipment. 9. The vowel/consonant determining means includes at least a first comparator that compares the peak detected by the peak detecting means and a threshold value set by the threshold setting means; a second comparator for comparing the calculated average value and a predetermined threshold value set by the threshold setting means;
7. The signal processing device according to claim 6, further comprising a vowel/consonant determination circuit that determines vowels and consonants based on the comparison results of the comparator and outputs the results. 10. The vowel/consonant determining means includes at least a first comparator for comparing the peak detected by the peak detecting means and the threshold value set by the threshold setting means, and the average value calculating means. a second comparator for comparing the average value calculated by the threshold value and a predetermined threshold value set by the threshold setting means;
8. The signal processing device according to claim 7, further comprising a vowel/consonant determination circuit that determines whether a vowel or a consonant is determined based on the comparison result of the second comparator and outputs the result.