JP3418005B2 - Voice pitch detection device - Google Patents

Abstract

PURPOSE:To surely and correctly detect a voice pitch in a voice pitch detecting device detecting the voice pitch from autocorrelations in short sections of a voice waveform. CONSTITUTION:This device is provided with an autocorrelation means 31 taking in a voice signal and calculating a correlation value by taking autocorrelations of the voice signal with respect to a preset time interval while making every voice frame a section, a correlation value smoothing means 32 giving weights being monotonically increased in order of time series of the voice frames to the correlation value calculated by the autocorrelation means 31 and calculating an evaluation value by adding these weighted correlation values at every time interval and a maximum correlation detecting means 33 detecting a time interval when the evaluation value calculated by the correlation smoothing means 32 becomes the maximum value as the voice pitch of the voice signal.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice pitch detecting apparatus for detecting a voice pitch having a maximum correlation value by taking autocorrelation in a short section of a voice signal.

[0002]

2. Description of the Related Art Generally, in a mobile communication system such as a mobile phone, a car phone or the like, there is a communication between a mobile station and a base station as shown in FIG.
Wireless communication as shown in FIG.

The voice input sampled through the sampling unit 1 is encoded by the voice encoding unit 2 and the error correction encoding unit 3 and transmitted by the modulating unit 4 toward the base station.

In such a mobile communication system, it is desirable to reduce the amount of transmission information in order to effectively use the radio frequency band. Therefore, as shown in FIG. 10, there is known a speech encoding unit 2 that highly efficiently encodes speech based on the vector sum excitation linear predictive coding method (VSELP).

In such a system, the voice pitch detection device 11 detects the basic period of voice (hereinafter referred to as "voice pitch") in a voice frame of about 5 to 20 msec. Also, the LPC analysis unit 12 and the LPC quantization unit 13
Then, the reflection coefficient that defines the waveform per basic period is detected. Further, the power quantizer 14 detects the voice power in the voice frame. In addition, the voiced / unvoiced determination unit 1
At 5, the presence / absence of voice is determined.

As described above, by extracting the characteristics such as the voice pitch, the redundancy of the voice is removed and the coding is performed with high efficiency. FIG. 11 is a diagram showing a conventional voice pitch detecting device.

In the figure, the voice input is given to the sampling unit 1, and the output of the sampling unit 1 is connected to the input of the autocorrelation unit 21. The output of the autocorrelation unit 21 is connected to the input of the peak detection unit 22.

In the voice pitch detecting apparatus having such a configuration, the sampling unit 1 samples the voice input at a frequency of 8 KHz, for example. The autocorrelation unit 21 divides the sampled voice input into voice frames of about 5 to 20 milliseconds and captures them. If the number of samples of this voice frame is shown as N and the j-th voice input in the i-th voice frame is shown as IN [i, k], the autocorrelation function B [i, k] for the time interval k is

[0009]

[Equation 1]

It is obtained by the calculation shown in. In general,
Although a voice input is an irregular fluctuating signal, a voice input in a short section such as a voice frame can be regarded as a periodic stationary signal. Therefore, the autocorrelation function B [i, k] calculated in the short section has a peak corresponding to the periodicity of the voice pitch.

The peak detector 22 searches for a peak of such an autocorrelation function B [i, k] and detects the value of the time interval k that gives the maximum value as the voice pitch.

[0012]

However, in the conventional voice pitch detection apparatus, since the autocorrelation in the short section is taken as described above, the peaks other than the voice pitch are subject to the autocorrelation function due to the instantaneous fluctuation of the voice input. Occur in large numbers. for that reason,
The voice pitch was often erroneously detected.

Further, even for a time interval which is an integral multiple of the voice pitch, the peak of the autocorrelation function is folded and generated, so that the voice pitch is often erroneously detected. Further, due to the harmonic component of the voice input, a peak often occurs in the autocorrelation function at a place that is a fraction of the voice pitch, and the voice pitch is often erroneously detected.

If the voice pitch is erroneously detected in this way, the detected value of the voice pitch fluctuates greatly in a short period of time.
Therefore, the voice reproduced at the receiving end has its pitch raised and lowered at the vowel part, resulting in a marked deterioration in the quality of the call.

It is an object of the present invention to provide a voice pitch detecting device which detects a voice pitch more reliably and more accurately than the conventional example.

[0016]

FIG. 1 shows the features of claims 1, 3 and
4 is a principle block diagram of the invention described in FIG. According to the first aspect of the present invention, an autocorrelation unit 31 that takes in an audio signal, calculates an autocorrelation of the audio signal with respect to a preset time interval for each audio frame as a section, and calculates a correlation value.
And the correlation value calculated by the auto-correlation means 31 is weighted so as to increase monotonically in time series of the audio frame, and these weighted correlation values are added at each time interval to calculate an evaluation value. The smoothing means 32 and the maximum correlation detecting means 33 for detecting the time interval at which the evaluation value calculated by the correlation value smoothing means 32 has the maximum value as the voice pitch of the voice signal are provided.

FIG. 2 is a block diagram showing the principle of the invention described in claims 2, 3, and 4. According to a second aspect of the present invention, a voice signal is taken in, and an autocorrelation of the voice signal with respect to a preset time interval is taken for each voice frame as a section,
Autocorrelation means 31 for calculating a correlation value and autocorrelation means 3
The correlation value calculated by 1 is weighted so that it decreases monotonically before and after the peak point along the time series, and these weighted correlation values are added at each time interval to calculate the evaluation value. The means 35 and the maximum correlation detecting means 33 for detecting the time interval at which the evaluation value takes the maximum value as the voice pitch of the voice frame corresponding to the above-mentioned peak point are provided.

According to a third aspect of the present invention, in the voice pitch detecting apparatus according to the first or second aspect, the autocorrelation means 31 calculates a sum of individual autocorrelation values for time intervals having different signs. It is characterized in that the correlation value is obtained.

The invention according to claim 4 is the invention as defined in claims 1, 2 and
In the voice pitch detection apparatus according to any one of 3 above, the autocorrelation unit 31 performs a process of suppressing a frequency component equal to or more than the reciprocal of the lower limit value of the time interval for the voice signal whose correlation value is to be calculated. It is characterized by including means.

[0020]

In the voice pitch detecting apparatus according to the first aspect of the invention, the autocorrelation means 31 takes the autocorrelation of the voice signal.

The correlation value smoothing means 32 adds weights to these correlation values in a time-series order of the audio frame and adds them to obtain an evaluation value. In general, the voice pitch gradually changes over several voice frames, and thus the voice pitch has a high correlation over several voice frames. Therefore, by adding the correlation values of the voice frames in the forward direction on the time series, the peaks other than the voice pitch are smoothed and the noise component is suppressed.

However, since voice pitch correlation is low for voice frames that are distant in time series, weighting is reduced for voice frames that are distant in time series to reduce the influence on the evaluation value.

The maximum correlation detecting means 33 surely and accurately detects the voice pitch by obtaining the time interval at which the smoothed evaluation value becomes maximum. In the voice pitch detecting apparatus according to the second aspect of the invention, the autocorrelation means 31 takes in a voice waveform and obtains a correlation value for each voice frame section.

The front-back smoothing means 35 adds weights that monotonically decrease around the peak point on the time series to the correlation values and adds them to obtain an evaluation value. In this way, the correlation values of the speech frames before and after the peak point are smoothed. Therefore, peaks other than the voice pitch are strongly suppressed.

The maximum correlation detecting means 33 finds the time interval at which the evaluation value thus smoothed becomes the maximum, and sets it as the voice pitch of the voice frame corresponding to the peak point. In the voice pitch detecting apparatus according to the third aspect of the present invention, the autocorrelation means 31 takes the autocorrelation in the forward direction and the backward direction in the time series, and therefore, due to the instantaneous fluctuation of the voice waveform,
The fluctuation that occurs in the correlation value becomes small. Therefore, the voice pitch is reliably and accurately detected.

In the voice pitch detecting apparatus according to the fourth aspect of the present invention, the frequency components higher than the reciprocal of the lower limit value of the time interval are reduced in the taken voice signal.
Therefore, the noise component irrelevant to the voice pitch is strongly suppressed, and the voice pitch is detected reliably and accurately.

[0027]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 3 is a diagram showing an embodiment corresponding to the invention described in claim 1.

In the figure, the voice input is given to the sampling unit 1, and the output of the sampling unit 1 is connected to the input of the autocorrelation unit 21. The output of the autocorrelation unit 21 is connected to the input of the smoothing unit 40, and the output of the smoothing unit 40 is connected to the input of the peak detection unit 22.

The smoothing section 40 is composed of a plurality of identical circuits connected to each output of the autocorrelation section 21. In such a circuit, the output corresponding to the time interval k of the autocorrelation unit 21 is connected to the first input of the adder 41k, and the adder 41k
Is connected to the delay device 42k and the peak detector 22. The output of the delay device 42k passes through the attenuator 43k,
It is connected to the second input of the adder 41k.

The correlation between the present embodiment and the invention according to claim 1 will be described. The autocorrelation unit 21 corresponds to the autocorrelation unit 31, the smoothing unit 40 corresponds to the correlation value smoothing unit 32, The peak detector 22 corresponds to the maximum correlation detector 33.

FIG. 4 is a flow chart showing the operation of this embodiment corresponding to claim 1. The operation of this embodiment will be described below. In the voice pitch detection device having such a configuration, the voice input is sampled by the sampling unit 1 at a frequency of 8 KHz, for example, and is taken into the autocorrelation unit 21.

The autocorrelation unit 21 divides the sampled voice input into voice frames of about 5 to 20 milliseconds and sets the autocorrelation function for the time interval k within the search range L _{MIN to} L _MAX of the voice pitch. B [i, k] is given by the above equation (1)
Calculation is performed by the arithmetic operation shown in (step S1).

Generally, a voice input is an irregular fluctuating signal, but a voice input in a short section such as a voice frame can be regarded as a periodic stationary signal. Therefore, the autocorrelation function B [i, k] calculated in the short section has a peak according to the periodicity of the voice pitch.

In the smoothing section 40, the autocorrelation function B [i, k] calculated for each time interval k is individually input to a recursive filter composed of an adder 41k, a delay unit 42k and an attenuator 43k. It Such a recursive filter is

[0035]

[Equation 2]

By performing the arithmetic operation shown in
A value corresponding to [i, k] is output (step S2).
In the smoothing unit 40, as shown in the above equation, the autocorrelation function in the forward direction is weighted so as to increase monotonically in the time series of the audio frame, and is sequentially added.

Generally, since the voice pitch gradually changes over several frames of voice frames, the correlation of the voice pitch is high in the autocorrelation function of several frames. Therefore, the peaks other than the voice pitch are smoothed by adding the autocorrelation functions of the voice frames in the forward direction.

However, voice frames that are distant in time series have a low voice pitch correlation, and thus are weighted smaller in the forward voice frames. The value of the attenuation ratio γ in the above equation is preferably 0.7 to 0.95 in view of the change time of the voice pitch in general voice. By setting such a value, the smoothing unit 40 can efficiently suppress the peak of the autocorrelation function other than the voice pitch.

The peak detecting section 22 has the smoothing section 4 as described above.
An output C [i, k] of 0 is fetched, a time interval k at which this output has a maximum value is searched for, and output as a voice pitch (step S3).

As described above, in this embodiment, by adding the forward autocorrelation function, the noise component of the autocorrelation function is suppressed, and the voice pitch can be detected reliably and accurately.

Further, in this embodiment, since only the forward autocorrelation function is added, there is no delay in detecting the voice pitch. Furthermore, in this embodiment, since the smoothing unit 40 is configured by the recursive filter, it is not necessary to sequentially store the forward autocorrelation function, and the configuration of the smoothing unit 40 can be simplified.

FIG. 5 is a diagram showing an embodiment corresponding to the invention described in claim 2. In FIG. In the figure, the output of the autocorrelation unit 21 is connected to the input of the front-back smoothing unit 50,
The output of 0 is connected to the input of the peak detector 22.

Further, the front-back smoothing unit 50 includes the autocorrelation unit 21.
The same circuit is connected in plural for each output.
In such a circuit, the output for the time interval k of the autocorrelation unit 21 is connected to the first input of the adder 51k, and the output of the adder 51k is added via the delay unit 52k and the attenuator 53k. Connected to the second input of.

The output of the autocorrelation unit 21 is the attenuator 5
It is connected to the first input of the adder 55k via 4k. The second input of the adder 55k is connected to the output of the delay device 52k, and the output of the adder 55k is connected to the input of the peak detector 22.

The correlation between this embodiment and the invention according to claim 2 will be described. The autocorrelation unit 21 corresponds to the autocorrelation unit 31, and the front-back smoothing unit 50 is the front-back smoothing unit 35.
The peak detector 22 corresponds to the maximum correlation detector 33.

FIG. 6 is a flow chart showing the operation of the embodiment of claim 2. The operation of this embodiment will be described below. In the speech pitch detection device having such a configuration, the autocorrelation unit 21 calculates the autocorrelation function B [i, k] (step S1).

The front-back smoothing unit 50 takes in such an autocorrelation function B [i, k] and adds it to an adder 51k and a delay unit 52.
k and an attenuator 53k.

Such a recursive filter is

[0049]

[Equation 3]

C [i
The value corresponding to [-1, k] is output (step S2). In the recursive filter, as shown in the above equation, the forward autocorrelation function is weighted so as to increase monotonically in the time series of the speech frame, and is sequentially added.

The output C [i-
1, k] is added to the i-th autocorrelation function by the adder 55k, and D [i-1, k] = B [i, k] β + C [i-1, k] is output (step S3).

As described above, the front-back smoothing portion 50 has (i-1)
The autocorrelation function is subjected to weighting that decreases monotonically in the time series, with the second speech frame as the peak, and then added. Therefore, peaks other than the voice pitch are reduced.

The peak detecting section 22 sets the time interval k for maximizing the output D [i-1, k] of the front-back smoothing section 50 to (i-
It is detected as the voice pitch of the 1) th voice frame (step S4).

As described above, in this embodiment, the noise components of the autocorrelation function are added by adding the front and rear autocorrelation functions around the predetermined voice frame, as compared with the case of adding only the forward autocorrelation function. Can be further suppressed.

Further, in this embodiment, since the front-back smoothing unit 50 is composed of the recursive filter, it is not necessary to successively store the forward autocorrelation function, and the structure of the front-back smoothing unit 50 can be simplified.

FIG. 7 is a diagram showing an embodiment corresponding to the invention described in claim 3. In FIG. In the figure, to show the feature of the configuration of this embodiment, a speech input is given to the sampling unit 1, an output of the sampling unit 1 is connected to an input of the symmetric autocorrelation unit 61,
The output of the symmetric autocorrelation unit 61 is connected to the input of the smoothing unit 40.

Further, in this figure, the same components as those of FIG. 3 are designated by the same reference numerals, and the description thereof will be omitted. The symmetric autocorrelation unit 61 corresponds to the autocorrelation means 31 in terms of the correspondence between the present embodiment and the invention according to claim 3.

In the voice pitch detecting apparatus having such a configuration, the voice input is sampled by the sampling unit 1 at a frequency of 8 KHz, for example, and is taken into the symmetric autocorrelation unit 61.

The symmetric autocorrelation unit 61 divides the sampled speech input into speech frames of about 5 to 20 milliseconds, and with respect to the time interval k of the speech pitch search range L _{MIN to} L _MAX ,

[0060]

[Equation 4]

The autocorrelation function B'by the arithmetic operation shown in
Calculate [i, k]. The autocorrelation function calculated in this way is smoothed by the smoothing unit 40 (similar to the above-described embodiment), and the pitch detecting unit 22 detects the voice pitch.

As described above, in this embodiment, even if noise or other fluctuations occur in either the forward direction or the backward direction of the time-sequential audio frames, the audio waveforms are averaged in the forward and backward directions. Therefore, the influence on the autocorrelation function is small.

In this embodiment, the smoothing section 40 is used to smooth the autocorrelation function, but the front-back smoothing section 50 described above may be used. With this configuration,
The voice pitch can be detected more reliably.

FIG. 8 is a diagram showing an embodiment of the voice pitch detecting device of the fourth aspect. In the figure, a characteristic of the configuration of the present embodiment is that a low-pass filter 70 is arranged at the input stage of the autocorrelation unit 21.

Further, in the figure, the same components as those of FIG. 3 are designated by the same reference numerals, and the description thereof will be omitted. The low pass filter 70 reduces frequency components higher than the search range of the voice pitch. The voice input from which the noise component irrelevant to the voice pitch is removed is taken into the autocorrelation unit 21 (similar to the above-described embodiment) and smoothed via the smoothing unit 40.

As described above, in this embodiment, by removing the noise component concentrated in the high frequency band of the voice input, the aliasing noise of the high frequency component generated in the short-range autocorrelation function is reduced, and the voice pitch is ensured. And it can be detected accurately.

In this embodiment, the low pass filter 70 is connected after sampling the voice input, but the low pass filter may be connected before sampling. Further, in this embodiment, the autocorrelation unit 21 and the smoothing unit 40 are applied, but a symmetric autocorrelation unit 61 may be used instead of the autocorrelation unit 21, and the front and rear parts instead of the smoothing unit 40 may be used. You may use the smooth part 50.

Further, in the above-mentioned embodiment, the smoothing portion 40
Although a recursive filter is applied to the front-back smoothing unit 50, any filter that weights and adds forward or front-side autocorrelation functions may be used, and for example, a transversal filter may be used.

Further, in the above-described embodiment, the smoothing section 40 and the front-back smoothing section 50 are constituted by digital filters, but any calculation means for weighting and adding the forward or backward autocorrelation functions may be used. , A digital signal processor (DSP) may be applied, or a microcomputer may be used.

[0070]

As described above, in the invention described in claim 1, since the correlation value smoothing means 32 adds the correlation values in the forward direction on the time series, the noise component is suppressed and the voice pitch is reduced. Reliable and accurate detection.

In the invention described in claim 2, the front-back smoothing means 35 adds the correlation values in the forward direction and the backward direction around the voice frame corresponding to the peak point. Therefore,
The noise component of the correlation value is strongly suppressed, and the voice pitch is detected reliably and accurately.

According to the third aspect of the invention, the autocorrelation means 31 takes the autocorrelation with respect to the forward and backward voice waveforms. Therefore, even if an instantaneous variation occurs in either the forward direction or the backward direction of the speech waveform, the variation in the correlation value is small and the speech pitch can be detected reliably and accurately.

According to the fourth aspect of the present invention, the high frequency component of the voice which is unnecessary for detecting the voice pitch is suppressed, so that the noise component included in the evaluation value is reduced. Therefore, the voice pitch can be detected reliably and accurately.

As described above, in the mobile communication system using the voice pitch detecting apparatus of the present invention, the voice pitch is accurately detected, so that the voice pitch is accurately encoded. Further, at the receiving end of the mobile communication system, the voice pitch does not fluctuate abruptly, so that the quality of the call can be significantly improved.

[Brief description of drawings]

FIG. 1 is a principle block diagram of the invention described in claims 1, 3, and 4.

FIG. 2 is a principle block diagram of the invention described in claims 2, 3, and 4.

FIG. 3 is a diagram showing an embodiment corresponding to the invention described in claim 1.

FIG. 4 is a flowchart showing the operation of the present embodiment corresponding to claim 1.

FIG. 5 is a diagram showing an embodiment corresponding to the invention described in claim 2.

FIG. 6 is a flowchart showing the operation of the present embodiment corresponding to claim 2;

FIG. 7 is a diagram showing an embodiment corresponding to the invention according to claim 3;

FIG. 8 is a diagram showing an embodiment corresponding to the invention described in claim 4;

FIG. 9 is a diagram showing a configuration of a mobile communication system.

FIG. 10 is a diagram showing an example of a speech encoding unit 2.

FIG. 11 is a diagram showing a conventional voice pitch detection device.

[Explanation of symbols]

1 sampling unit 21 Autocorrelation part 22 Peak detector 31 autocorrelation means 32 Correlation value smoothing means 33 maximum correlation detection means 35 Front-back smoothing means 40 smooth section 41k, 51k, 55k adder 42k, 52k delay device 43k, 53k, 54k attenuator 50 front and back smooth part 61 Symmetric autocorrelation part 70 Low pass filter

─────────────────────────────────────────────────── ─── Continuation of the front page (58) Fields surveyed (Int.Cl. ⁷ , DB name) G10L 11/04

Claims (4)

(57) [Claims] 1. An autocorrelation unit that takes in a voice signal, calculates the autocorrelation of the voice signal with respect to a preset time interval in intervals of each voice frame, and calculates a correlation value, and the autocorrelation unit calculates the correlation value. The correlation value is weighted so as to increase monotonically in the time series of the speech frame, and the correlation value smoothing means calculates the evaluation value by adding these weighted correlation values at each time interval, and by the correlation value smoothing means. A voice pitch detecting device, comprising: a maximum correlation detecting means for detecting a time interval at which the calculated evaluation value takes a maximum value as a voice pitch of the voice signal. 2. An autocorrelation unit that takes in a voice signal, calculates the autocorrelation of the voice signal with respect to a preset time interval for each voice frame as a section, and calculates a correlation value, and the autocorrelation unit calculates the correlation value. The correlation value is weighted so as to monotonically decrease back and forth around the peak point along the time series, and a front-back smoothing unit that calculates the evaluation value by adding these weighted correlation values at each time interval, and the peak. A voice pitch detecting apparatus comprising: a maximum correlation detecting means for detecting the time interval at which the evaluation value has a maximum value as a voice pitch of a voice frame corresponding to a point. 3. The voice pitch detection apparatus according to claim 1, wherein the autocorrelation means obtains a correlation value by calculating a sum of individual autocorrelation values for time intervals having different codes. Characteristic voice pitch detection device. 4. The voice pitch detecting apparatus according to claim 1, wherein the autocorrelation means has a reciprocal number of a lower limit value of a time interval or more with respect to a voice signal whose correlation value is to be calculated. A voice pitch detecting apparatus including means for performing processing for suppressing the frequency component of