JP2002366189A - System for identifying and detecting music and voice - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide signal processing technique for automatically improving clarity in sound transmission in an electronic circuit and a digital signal processing field in the industry of equipment such as acoustic equipment, broadcasting unit, a receiver or guide broadcasting equipment. SOLUTION: A signal is generated which has a specified delay time within the range of the several hundreds of milliseconds of an original acoustic signal, energy is calculated between the signal and the original signal, a low-pass filter is made to work and, then, a kind of relative strength is calculated in a time difference. The correlation is statistically evaluated so as to obtain a music and sound identifying function.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0101]

BACKGROUND OF THE INVENTION Signal processing for automatically improving voice transmission clarity in the field of electronic circuits and digital signal processing in the device industry such as [acoustic equipment, broadcasting equipment, receiving equipment, and guide broadcasting equipment] Technology.

[0092]

In recent years, audio equipment,
The sound quality of AV equipment tends to emphasize the bass range by modifying the transfer characteristics of the electronic circuit. In many cases, such a method is designed mainly with an emphasis on the sound quality of music. In such a system, it is often difficult to ensure clarity in a signal such as news or weather forecast that emphasizes clarity of voice. Generally, pass characteristics that are incompatible with [characteristics necessary for expressing richness of music] and [characteristics required for transmitting audio content] are required. Since there is no filter that satisfies both needs, a solution to this problem requires the ability to distinguish between music and speech.

[0093]

In general, it has been confirmed by experiments that the intensity of the autocorrelation function of a voice signal tends to attenuate [as the time difference increases], as compared with a music signal. If it is voice, it is influenced by [individual difference] and [the content of the voice such as news, weather forecast, conversation, dialogue of movies, etc., and the presence or absence of BGM]. Although there are variations due to the effects of music tempo, instrument type, and playing method, on average, speech and music differ greatly in this respect. The present invention pays attention to this point, and solves this problem by a small-scale arithmetic processing step without performing a large amount of arithmetic processing [Fourier analysis and pattern analysis of the result thereof].

[0093]

2. Description of the Related Art If it ’s a consumer product,
Normally, the user adjusts the sound quality to his or her preference, and in the field of recording and PA, a specialized adjuster adjusts the sound quality on a case-by-case basis.

[0056]

[Definition of terms] [Speech] is a signal of a voice spoken by a person. [Music signal] is a signal such as an instrumental sound or a singing voice that constitutes music. [Addition] generally includes subtraction. [Addition, subtraction, multiplication, division, logarithm] is not strictly mathematical, but is a function that can tolerate an error that does not hinder a practical device. The “effective component” is assumed to be a “positive or negative DC component” or a “long-period AC component” having a strong relationship with the correlation strength included in the product of two acoustic signals. It is assumed that the "ineffective component" is a short-period AC component that is not included in the correlation strength and included in the product of the two acoustic signals. [DSP] is an arithmetic processing integrated circuit specialized in signal processing.

FIG. 1 is a block diagram showing one embodiment of the present invention. In order to make the description simple and practically sufficient, an embodiment in which there are two delay functions will be described. Since this is an embodiment of practical signal processing for obtaining actual functions, it includes many functions not related to the essence of the present invention, but does not limit the generality of the present invention. In general, the number of delay functions is determined depending on [performance and cost] required for an application device. Since the present invention is applied to a consumer product, it is an embodiment satisfying necessary and sufficient functions in light of general human senses.

F (t) is the input audio signal, DLY_1, DL
Y_2 is a delay function having specific delay times D1 and D2, respectively.
This is the second delay function. MPY_0, MPY_1, MPY
_2 is a multiplication function that multiplies two inputs by four quadrants, and their names are a zeroth multiplication function, a first multiplication function, and a second multiplication function. LPF_ of block LPF
0, LPF_1 and LPF_2 respectively have a cutoff frequency F
R is a low-pass filter function having the same characteristics and has the same name as the 0th low-pass filter function, the first low-pass filter function, and the second low-pass filter function. RMS_0, RMS
_1 and RMS_2 are functions for obtaining a short-time average value or an execution value of an input signal or a value similar thereto, respectively, and their names are a zeroth averaging function, a first averaging function, and a second averaging function, respectively. . Let the outputs of the averaging function be P0 (t), P1 (t) and P2 (t).

SMPL_0, SMPL of block SMPL
_1 and SMPL_2 output the output of the averaging function during the period TS.
This is a sampling function that integrates and samples the result. The outputs are PS0 (t), PS1
(T) and PS2 (t). SGT is a selection function for selecting the larger of the outputs PS1 (t) and PS2 (t) of the sampling function. Its output is PS12 (t).

LOG_0 and LOG_12 are PS
It is a logarithmic operation function that performs logarithmic conversion of 0 (t) and P12 (t). PL0 (t) and PL12 (t) are LO
Outputs of G_0 and LOG_12. NRM is PL12
This is a normalization function that outputs the difference between (t) and PL0 (t), and the output is G (t). DIFF is a difference function for calculating the difference between G (t) and G (t-TS) for each cycle TS, and its output is H (t). AVG is period T
This is an integration function of integrating N times of S and sampling and outputting the result. The output is J (t). DTC
T further averages J (t) to distinguish between voice and music,
This is a detection function for converting into a signal necessary for controlling the acoustic filter FLT. DTCT is a parameter for averaging, dead zone Zded, threshold Lt.
hd, an averaging attack time constant TAavg and a release time constant TRavg.

CTRL is determined by the output M (t) of the detection function.
This is a control function of determining the correlation strength of the input audio signal or generating an acoustic characteristic control signal [necessary for controlling the acoustic characteristic] corresponding to the correlation intensity. FLT is a variable constant filter function that changes acoustic characteristics.

In the following description, the number n is 0, 1, or 2, and the same number belongs to the same block. [MPY_0,
LPF_0, RMS_0] outputs the short-time average intensity P0 (t) of the input audio signal. The two inputs of the multiplication function MPY_0 are both input sound signals f (t).
It is. Therefore, the output C0 (t) of MPY_0 is f
Since it is the square of (t), all components are effective components and always have positive values. Output P of averaging function RMS_0
0 (t) is the short-time average intensity of C0 (t). P0
(T) is used for normalizing the correlation strength between f (t) and f (t-Dn). The dimension of P0 (t) is the square of the acoustic signal.

The block consisting of [MPY_n, LPF_n, RMS_n] is a short-time average correlation strength Pn (t) between the input acoustic signal f (t) and the delay signal f (t_Dn) of the input signal.
Is output. One of two inputs of the multiplication function MPY_n is an input sound signal f (t), and the other is a delay signal f (t-Dn) delayed by a time Dn of f (t). In this embodiment, it is confirmed by experiment that the delay time is several tens msec to several hundred msec. The output Cn (t) of MPY_n is the product of f (t) and f (t-Dn). C
0 (t) is the perfect square of f (t), but Cn (t)
Is the product of two signals whose phase and frequency are not necessarily the same in all frequency bands of the acoustic signal, and therefore contains a frequency component that is close to twice the original frequency and changes to positive or negative. Inactive ingredients]. As the period of f (t) becomes more stable, the number of [positive or negative] effective components contained in Cn (t) increases.

Generally, a signal of [music or voice] is generated by [a natural vibration of a string, a membrane, or a structure], so that the signal has an autocorrelation strength. Therefore, Cn (t) includes an [effective component, that is, a low-frequency component], but the effective component increases as the vibration of the sound source is stably maintained, and the effective component decreases as the change in the vibration increases. . From music to audio, there is no clear line between them, but in general,
Voice tends to have a large change in [sound quality and pitch] in a complicated manner, and music tends to have stable [sound quality and pitch]. Therefore, the magnitudes of P1 (t) and P2 (t) are P0
It is generally smaller than (t). And, by numerical evaluation of the degree of the magnitude and the degree of time change,
This is used as a material for determining whether the sound or the music.

The low-pass filter function LFP_n is Cn (t)
The inactive component is removed from, and the active component CLn (t) is extracted. The low pass filter function LPF_n is usually a simple [1
Second-order low-pass filter or second-order low-pass filter].
In this embodiment, the filter is a secondary low-pass filter having a cutoff frequency of 1 Hz to 20 Hz. Since the characteristics of the LPF_n are not the essence of the present invention, detailed description is omitted.

The averaging function RMSn extracts the short-time average intensity Pn (t) of CLn (t). Since the dimension of CLn (t) is a multiplication of two acoustic signals, P0
Since the RMSn function must be matched with the dimension of (t), the short-time average of the absolute value is a simple and effective function of the RMSn. RMSn (t) determines how C
Since obtaining the short-time average intensity of Ln (t) is not the essence of the present invention, the detailed description is omitted.

In most cases, a digital signal processor is used for a series of signal processing. Since the cost allowed for this type of signal processing is about several tens of yen (several tens of cents) due to the nature of applied products, the number of calculation steps and the number of memories used must be minimized. On the other hand, it is not necessary to control the sound quality of the acoustic signal at a high speed in light of human sensitivity. In this embodiment, until the SMPL, every sampling cycle of the acoustic data, the subsequent processing is 50 ms.
The determination is made by calculation based on sampling every c to 100 msec. In SMPL_n, during time TS, P
n (t) is integrated, and the result is output as a sampling value. The SGT selects the larger value of PS1 (t) and PS2 (t), and obtains the output PS12 (t). Since obtaining the correlation strength accurately in an arbitrary frequency range requires a large amount of calculation, in the present embodiment, two correlation strengths are calculated, and by selecting the larger one, the practicality can be reduced with a small calculation amount. Is secured. The value of n is not the essence of the present invention.

Since the magnitude of PL12 (t) depends on the magnitude of the input signal, a ratio with PL0 (t) is required for normalization. Rather than calculating the value of PS12 (t) / PS0 (t), {Log (PL12 (t))} −
Since {Log (PL0 (t))} is convenient for signal processing, PS0 (t) and PS12 (t) are logarithmically converted by logarithmic calculation functions LOG_0 and LOG_12. NRM is a mere subtraction process whose output G
(T) is PL12 (t) -PL0 (t), the value of which is Log {P12 (t) / P0 (t)}, the correlation strength normalized by the root mean square value of the original signal. Is the short-term average of Therefore, G (t) is not affected by the strength of the input signal.

The output H (t) of the difference function DIFF is G (t)
−G (t−Ts). Generally, in the case of a music signal, the change of the magnitude of G (t) with respect to time including vocals is small, and in the case of an audio signal such as a news or weather forecast, the change of the magnitude of G (t) with time is large. . In this embodiment, [G
(The degree of temporal change in the magnitude of (t)) is used for the determination of music and voice. About the judgment of music and voice, G
The nature of (t) to use is not the essence of the present invention. SMPL_N further smoothes H (t). In this embodiment, J (t) is H (t) + H (t−
TS) + H (t−2TS) + H (t−3TS), but the presence or absence or method of this smoothing function is not the essence of the present invention.

The DTCT receives an averaged output J (t) as an input,
This function is to further smooth the signal to make it easier to control the sound quality. The statistical properties of the sound signal always fluctuate greatly, and if a value that fluctuates greatly is used as it is as the sound quality control signal, the controlled sound signal gives a sense of incongruity to human hearing.
For this reason, various smoothing methods are used. In this embodiment, a center value Lthd and a dead zone Zded for determination are provided, and further, the sound quality control signal M is smoothed by a time constant function having an attack time TAavg and a release time TRavg. (T) is generated. The existence or method of the smoothing function of the DTCT is not the essence of the present invention.

Generally, in order to enhance the clarity of speech, it is desirable to remove pitch components, which are unnecessary frequency components for speech recognition. Since this pitch component exists on the lowest tone side in the voice spectrum distribution, the clarity is improved by suppressing the low tone range. The pitch component of voice varies greatly from person to person, and is greatly influenced by the use of microphones, the type of microphones, and the sound effects at the time of broadcasting.In any case, suppression of the low-frequency range improves clarity. It is well known that it is effective. Therefore, in this embodiment, if it is determined that the input signal is a voice, the CTR operates so that the acoustic filter FLT has characteristics of suppressing the low frequency range.

How to select the value of the delay time Dn is not the essence of the present invention. However, since the value of Dn affects the performance of discriminating between music and voice, the embodiment of FIG. Dn will be described in detail. As a result of the experiment, it is appropriate that the average value of Dn is approximately several tens to several hundreds of milliseconds. The average time of this Dn is [the active ingredient generally increases as the length decreases] and [the active ingredient decreases as the length increases]. [Somewhat faster sound such as news]
And [somewhat slower sound such as commentary] have different active component sizes. Also, individual differences are large.

[0222]

Description of Table 1 For simplicity of explanation, as an example, in the embodiment based on the configuration of the embodiment of FIG. 1 where the delay time D1 is 100 msec and D2 is 82 msec, the signal is a pure sine wave. Table 1 shows the calculated expected detection rates that can be detected as music.

To simplify the description in Table 1, A
The names of the columns in columns F to F and the names of the columns in rows 1 to 85 in the vertical direction are given. Row 1 shows that the numerical value shown is the detection rate (unit%). Row 2 shows the number of samples having a strong correlation with a correlation strength of 0.5 or more. Row 3 shows the sample parameter.
8 to 85 in column A indicate the frequencies of the input signals on the 12-average scale. The delay function D1 is 82 ms for 8 to 85 in column B
ec is the correlation strength with the original signal. 8 in column C
Nos. To 85 are marked with "1" for those having a correlation intensity of row B of 0.5 or more. 8 to 85 in the D column indicate the correlation strength with the original signal when the delay time is 100 msec. In columns 8 to 85 in column E, "1" marks are entered for columns D having a correlation strength of 0.5 or more. The F column indicates “1” for one having a correlation strength of 0.5 or more when the delay time is 82 msec and 100 msec.
Is marked. As shown in the table, when the delay time is 82 msec, the number of detected music signals is 50 and the detection rate is 70.4 (%) with respect to the sample parameter 71.
It is. When the delay time is 100 msec, the number of detections as a music signal is 50 with respect to the sample parameter 71, and the detection rate is 70.4 (%). 82 msec delay time
In the system adopting the two parameters of 100 and 100 msec, the number of detections is 66 and the detection rate is 93.0 (%) with respect to the population parameter 71. The above is the case of a pure sine wave.

[0243]

Description of Table 2 With respect to the embodiment of FIG. 1, the performance of an actual signal processing program for actual music signals and audio signals was confirmed. The description of the dimension of each numerical value is omitted. No. Column indicates signal numbers. The column of M / S indicates whether the source signal is a music signal or an audio signal. S
The column of source indicates the type of the source signal. Numbering_
E indicates that the speech is in English, and Numbering_JP indicates that the speech is in Japanese. The column of G (t) shows the output G (t) of the block LOG_12 in the embodiment of FIG. 1 for each source. The column of H (t) shows the output H (t) of the block DIFF in the embodiment of FIG. 1 for each source. The column of M (t) shows the output M (t) of the block DTCT in the embodiment of FIG. 1 for each source.

The numbers 1 to 9 in Table 2 show the values of G (t), H (t) and M (t) when the signal source is music. Lines 10 to 20 in Table 2 show the values of G (t), H (t), and M (t) when the signal source is audio, respectively. The value of G (t) is 79.8-1 for music.
Although it is in the range of 15 and in the case of voice, it is in the range of 165 to 213, and it can be seen that it is clearly discriminated. H
The value of (t) is -17 to -26.8 in the case of music.
However, in the case of voice, it is in the range of 1.0 to 13.1, and it can be seen that it is clearly discriminated. M (t)
Is in the range of -11.4 to -32.7 in the case of music, but is in the range of 7.0 to 37.8 in the case of audio, which indicates that the value is clearly discriminated. The distribution of these calculation results in a certain range is a manifestation of the characteristics of each signal, and the boundaries are not clear in practice. At least, clear voice is clearly determined to be voice, and clear music is clearly determined to be music. This does not impair the essence of the present invention.

[0262]

BEST MODE FOR CARRYING OUT THE INVENTION An apparatus for transmitting music and voice as in the following example. 1) Computer and DSP programs 2) DSPP chips 3) AV equipment, stereo equipment, televisions, radios, PA systems, etc.

[0279]

As described above, the present invention is a function of automatically controlling the clarity of voice by combining known techniques. The clarity of voice, which is often found in devices mainly made of music, is improved. Automatically correct missing sound quality. When it ’s important to hear about news, weather, stocks, etc.
It is convenient. In particular, it is extremely effective for [speech related to numerical information and the like] and [speech that conveys many contents in a short time].

[0285]

[Brief description of the drawings]

FIG. 1 shows an embodiment of the present invention.

[Explanation of symbols]

 In the following description, the number n is 0, 1, or 2.

[INPUT] Input signal terminal

[F (t)] input signal

[Output] Output terminal

[F (t)] output signal

[DLY_n] delay function

[Dn] Delay time

[F (t-Dn)] delay signal

[MPY_n] Multiplier

[Cn (t)] Multiplier output signal

[LPF] Low pass filter

[LPF_n] Low-pass filter

[FR] Cutoff frequency of low-pass filter

[CLn (t)] Output of low-pass filter

[RMS_n] Short-time average intensity calculation function

[Pn (t)] Output of short-time average intensity function

[SMPL] Integration and sampling function

[SMPL_n] Integration and sampling function

[TS] Integration time or sampling cycle

[PSn (t)] sampling signal

[SGT] Larger signal selection function

[PS12 (t)] Larger signal

[LOG_0] Logarithmic calculation function

[LOG_12] Logarithmic operation function

[PL0 (t)] reference correlation strength

[PL12 (t)] detected correlation strength

[NRM] Normalization function

[G (t)] normalized detected correlation intensity

[DIFF] Difference calculation function

[H (t)] Difference output

[SMPL_N] Integration and sampling function

[N] Integration time of multiple of TS

[J (t)] N ^* TS Detected value of sampling period

[DTCT] Smoothing function

[TAavg] Attack time constant

[TRavg] Release time constant

[Zded] Detection dead zone

[Lthd] Detection level

[M (t)] Output of smoothing function

[CTR] Acoustic filter control function

[FLT] Variable constant filter under control to obtain optimal voice / music characteristics

[0290]

[Brief explanation of the table]

Table 1 Response of the detection unit to the sine wave input of the embodiment of FIG.

Table 2 Response of the detection unit to the actual signal input in the embodiment of FIG.

[0302]

[Explanation of symbols in table]

[Signal frequency] The frequency of the signal when the input signal is a sine wave on a 12-average scale

[Delay time] Delay time of D1 or D2 in FIG.

[Correlation strength] Pure autocorrelation coefficient

[Correlation strength is 0.5 or more] The autocorrelation coefficient that maximizes 1 is 0.5 or more

[The correlation strength is 0.5 or more as a system] Delay time D
1, whichever is greater, D2

[Parameter] Number of samples

[Number with correlation strength of 0.5 or more] Among the sample numbers, the number of strong samples with correlation strength of 0.5 or more

[Detection rate%] Percentage of samples with correlation strength of 0.5 or more

[No. ] Sample number column

[S] Field for distinction between music and voice

[Source] Column indicating signal type

[G (t)] Output of the normalization function NRM of FIG.

[H (t)] Output of the difference function DIFF of FIG.

[M (t)] Output of the smoothing function DTCT of FIG.

Claims (2)

[Claims] An arbitrary sound signal is set as a first sound signal, and
One or a plurality of delay functions having different delay times with an audio signal as an input are referred to as a first delay function group, and respective output signals of the first delay function group are referred to as a first delay signal group. ] And [individual first delay signal group] are referred to as a first multiplication function group, and a function of applying an [integration or low-pass filter] to each output signal of the first multiplication function group is referred to as a first multiplication function group. The first averaging function group is a function of obtaining an [average value or an execution value] by an arbitrary method for each output signal of the first low-pass filter function group. Each output of the group is referred to as a first correlation strength group, and a function of changing a filter characteristic, which receives a first acoustic signal as an input, is referred to as a first variable filter function. A first feature is that it has at least a [first delay function group, a first multiplication function group, and a first averaging function group].
A second feature is that a signal group is used as a signal for [discriminating between music and voice], and the signal group is dependent on the correlation strength group or the first correlation strength group. 2. A first feature of the present invention is that it has at least a [first delay function group, a first multiplication function group, and a first averaging function group]. The second characteristic is that the signal has a structure for controlling the filter characteristic of the first variable filter function.