JPH054355Y2 - - Google Patents

Description

[Detailed explanation of the idea] [Industrial application field]

The present invention relates to a signal processing device that processes a signal buried in noise and obtains a signal with an improved signal-to-noise ratio (S/N ratio).

[Conventional technology]

Conventionally, as this type of signal processing device, for example, a speech processing device that uses an autocorrelation function (Speech
Processing by use of Auto
Correlation function (hereinafter abbreviated as "SPAC"). The principle is as shown in Japanese Patent Publication No. 55-34954, the autocorrelation function of the signal waveform is calculated, noise components are removed using the properties of the autocorrelation function, and the signal waveform is re-edited and output. It is something. In other words, SPAC utilizes the properties of the autocorrelation function shown below. When the input signal is expressed as a time function f(t), its autocorrelation function φ(τ) is: φ(τ)=l T ₀ i → m ∽1/T ₀ ∫ ^T ₀ ^/2 _-T0/2 f(t)・f(t+τ)
dt (1), and if f(t) is f(t)=a ₀ /2+ _∞ 〓 ⁿ⁼¹ a _o (Cos nωt+θ _o ) (2), the DC component and ω are the fundamental angular frequency. When given as the sum of the alternating current components, φ(τ)=a ₀ ² /4+1/2 _∞ 〓 ⁿ⁼¹ a _o a _o ² cos(nωτ) (3). Therefore, since the autocorrelation function is composed of the same frequency components as the input signal and the phases of each harmonic match, the autocorrelation function maintains the same shape even if the origin changes. On the other hand, for broadband noise, its autocorrelation function n(τ) can be modeled as n(τ)=C ₀ exp(−ατ) (4) where C ₀ is a constant. The energy is concentrated around τ=0 and decreases exponentially, so it is considered that if τ becomes a little larger, it becomes substantially zero. That is, when the input signal is white noise, the energy of the autocorrelation function is concentrated around t=0. Therefore, if the signal is an audio signal, a short-time autocorrelation function is obtained, and the autocorrelation waveform excluding a few milliseconds after t=0 is processed by the period determining circuit described below from the point at which this waveform crosses the zero line. By cutting out the signal during the determined period and executing it one after another, a signal with reduced noise can be obtained. Further, the autocorrelation function has a property that the amplitude of each frequency component is proportional to the square of each component of the input signal, although it is not described in the above publication. Therefore, when audio contains many frequency components, a distorted waveform is reproduced. An equalizer circuit is added to remove such square distortion. FIG. 3 shows a schematic configuration of the main parts of the circuit block that performs the above signal processing. As shown in the figure, the input signal is converted into a square root component by an equalizer 1, which is a square root circuit, and is input to an autocorrelation function circuit 2. An autocorrelation function circuit 2 calculates an autocorrelation function. This output has reduced square distortion. From this autocorrelation function, the waveform extraction circuit 4, which is a waveform editor, the period determination circuit 5, and the correlation function waveform connection circuit 6 extract noise components during the pitch period of the autocorrelation function determined by the period determination circuit 5. During the period excluding the initial period, the waveform extraction circuit 4 extracts the autocorrelation function waveform, and the correlation function waveform connection circuit 6 extracts the autocorrelation function waveform.
Edit the signal waveform and output it. Note that in actual digital signal processing, data is temporally discrete values, so instead of formula (1), the autocorrelation function is φ(k)=1/N _N 〓 ^l=1 f・f _+k Take the short-time correlation function given by (1′). This is because if N is large to some extent, equations (1) and (1') can be considered to be substantially the same. In addition, when the target is an audio signal, the actual product-sum time is set to several tens of milliseconds by the waveform extraction circuit 4, and the period determination circuit 5
It has become common practice to discard the first 1 millisecond or so to remove noise. As described above, in the circuit shown in FIG. 3, square distortion is removed by passing the input signal through the equalizer 1. The equalizer 1 includes one in which one square root circuit is provided for all frequency bands, and one in which the input signal is divided into a plurality of frequency bands and a square root circuit is provided for each frequency band.

[Problem that the idea aims to solve]

In the above signal processing, if the equalizer is a single square root circuit, the frequency characteristics will not be flat and square distortion will remain. On the other hand, if a square root circuit is provided for each divided frequency band, the finer the division, the flatter the frequency characteristics will be, but the circuit will become more complicated. The present invention eliminates these drawbacks and provides a signal processing device that can completely eliminate square distortion using an extremely simple circuit.

[Means to solve the problem]

In order to solve the above problems, a signal processing device to which the present invention is applied includes a circuit that Fourier transforms an input signal and decomposes it into a spectral domain, a circuit that calculates a power spectral value from the spectral domain,
a circuit that calculates the square root of the power spectrum value;
The circuit is configured to obtain an autocorrelation function of the input signal using a circuit that performs inverse Fourier transform on the square root thereof, cuts off an initial period where noise components are concentrated from the obtained autocorrelation function, reorganizes the signal waveform, and outputs the signal waveform. The present invention is characterized in that the circuit for calculating the square root removes the square distortion of the autocorrelation function for all frequency components.

[Effect]

The Fourier transform of the autocorrelation function is equal to the power spectrum, and the inverse Fourier transform of the power spectrum gives the autocorrelation function. In other words, since the power spectrum and the autocorrelation function are in a Fourier transform pair relationship, ideally, if you perform Fourier transform on the input signal, obtain the power spectrum, take the square root of it, and perform inverse Fourier transform on this, the square distortion can be obtained. We can obtain an autocorrelation function that is not In this way, since the square root of each frequency of the power spectrum can be taken in the frequency domain of Fourier transform, the frequency characteristic of square distortion removal becomes flat.

【Example】

Hereinafter, embodiments of the present invention will be described in detail. FIG. 1 shows an example of a circuit block of a signal processing device according to the present invention. In the figure, 21 is a Fourier transform circuit that decomposes an input signal into a spectral domain. 22
calculates the power spectrum from the Fourier-transformed input signal using the power spectrum calculation circuit. 8 is a square root calculation circuit that takes the square root of the power spectrum and calculates 2
Eliminate multiplicative distortion. 23 is a Fourier inverse transform circuit,
This output can be regarded as an autocorrelation function obtained by taking the square root of the input waveform ideally. 4 is a waveform extraction circuit; 5 is a circuit for determining the pitch period of the autocorrelation function; and furthermore, determining the initial period during which noise components are concentrated. Reference numeral 6 denotes a correlation function waveform connecting circuit, which connects the waveforms smoothly. The power spectrum of the input signal and the autocorrelation function of the input signal are in a Fourier transform pair relationship. Therefore, by passing the output of the Fourier transform circuit 21 through the power spectrum calculation circuit 22 and the Fourier inverse transform circuit 23, an autocorrelation function is obtained. During this time, by passing through the square root circuit 8, square distortion occurring in the autocorrelation function is eliminated. Fourier inverse variation circuit 23
Determining the pitch period of the autocorrelation function output, cutting out the waveform, and reconnecting the waveform are performed using the waveform cutting circuit 4, the period determining circuit 5, and the correlation function waveform that constitute the waveform editor used in conventional SPACs. Connection circuit 6
It will be held in With such a circuit, noise is removed from the input signal, and a signal without square distortion is edited and output. Strictly speaking, the Fourier inverse transform circuit 23
The time domain waveform of the output of is no longer the original autocorrelation function. However, this difference can be virtually ignored, and the output of the inverse Fourier transform unit 23 can be regarded as an autocorrelation function obtained by ideally taking the square root of the input waveform. Note that in the above embodiments, no special treatment is taken with respect to the influence of residual noise in areas other than the vicinity of the time origin and the ease with which consonants are likely to be dropped. In particular, if square root processing is applied to each frequency component of the power spectrum, S/
This may lead to deterioration of the N ratio. Therefore, it is desirable to perform noise reduction processing in the frequency domain before performing square root processing. For example, by adding a noise reduction circuit 25 as shown in FIG. 2 between the power spectrum calculation circuit 22 and the square root calculation circuit 8 in FIG. 1, it is possible to increase the S/N ratio. As a device for noise reduction processing in the frequency domain, for example, there is a device that sets a certain threshold level and considers frequency components exceeding this level to be signal (speech) components. There is also a device that considers a predetermined number of frequency components having larger amplitudes as signal components. Regardless of which device is used, if the signal is speech, consonants tend to be missing if noise reduction processing is strengthened. Therefore, it is necessary to seek an appropriate compromise between the two. In reality, it is better to adaptively control the threshold level depending on the magnitude of the input. Generally speaking, for audio signals, the threshold level is set high in the vowel region and lowered in the consonant region to find a compromise between a decrease in the S/N ratio and a decrease in intelligibility.

[Effect of the idea]

As explained above, the signal processing device of the present invention is
When the input signal is white noise, the energy of the autocorrelation function concentrates around t = 0, which is the property of SPAC, but with an extremely simple circuit, it is possible to completely eliminate the square distortion for all frequency components of the signal. It is possible to obtain a high-quality output waveform with an improved signal-to-noise ratio.

[Brief explanation of drawings]

FIG. 1 is a circuit block diagram of the signal processing device of the present invention, FIG. 2 is a diagram showing a partial modification thereof, and FIG. 3 is a circuit block diagram for implementing a conventional SPAC. 1...Equalizer, 2...Autocorrelation function circuit,
4...Waveform extraction circuit, 5...Period determination circuit, 6...
... Correlation function waveform connection circuit, 8 ... Square root calculation circuit, 21 ... Fourier transform circuit, 22 ... Power spectrum calculation circuit, 23 ... Fourier inverse change circuit, 25 ... Noise reduction circuit.

Claims (1)

[Scope of utility model registration request] A circuit that performs Fourier transform on an input signal and decomposes it into a spectral domain, a circuit that computes a power spectral value from the spectral domain, a circuit that computes the square root of the power spectral value, and a circuit that inversely Fourier transforms the square root. The apparatus is configured to calculate an autocorrelation function of the input signal in advance, cut off an initial period in which noise components are concentrated from the calculated autocorrelation function, rearrange the signal waveform, and output the signal waveform, and calculate the square root. 1. A signal processing device characterized in that a circuit removes square distortion of an autocorrelation function for all frequency components.