KR19980051012A - A method for changing the speech pitch in the time domain by separating components - Google Patents

Abstract

The present invention relates to a method for changing the pitch of a speech in a speech synthesis technique, and more particularly to a method for separating a speech signal into a wideband component and a narrowband component, maintaining a wideband component, and scaling The present invention also relates to a method of changing a speech pitch in a time domain by separating a component obtained by adding the scaled narrowband component and the held wideband component to obtain a speech signal whose pitch is changed.

In the conventional speech synthesis technique, there are a time domain method, a frequency domain method, and a time frequency hybrid domain method in the pitch change method, but the spectrum distortion caused by the pitch change can not be overcome.

According to the present invention, there is provided a method comprising: separating a speech signal into a wideband component and a narrowband component using a variable lowpass filter; maintaining the separated wideband component; and performing time scaling on the separated narrowband component And a step of constructing a synthesized voice by adding the time-scaled narrowband component to the held wideband component.

Description

A method for changing the speech pitch in the time domain by separating components

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to speech synthesis techniques and, more particularly, to a method for changing the pitch of synthesized speech.

In a digital portable communication device, a voice encoder is implemented using various pitch change theories in order to efficiently use the bandwidth of a transmission channel and to obtain high sound quality.

Especially, in the field of voice service which has recently been diversified, high quality sound is required.

Most of the waveform coding methods and hybrid coding methods for composing synthetic speech use the same words as different data depending on the dependent words, which not only increases the size of the database but also becomes a problem in the database design. The limit is overcome by using the pitch change method.

In other words, the waveform coding method is to simply keep the shape of the waveform by reducing surplus components. In speech synthesis in this case, synthesis by analysis is mainly used for high-quality synthesis.

However, since the excitation component and the filter component are not classified in the encoding parameter in this case, it is difficult to apply the waveform encoding method to the synthesis according to the rule. In order to apply the rule-based synthesis method to the waveform encoding method, A technique of changing the pitch of the pitch is required.

The conventional pitch changing method includes a time domain method, a frequency domain method, and a time-frequency hybrid processing method depending on the processing region.

The time domain method includes a multi-pulse method and a pitch half-method.

Caspers and Atal proposed a method of inserting or deleting nulls in pulses in MPLPC, but the pulse train on MPLPC has a strong correlation with pitch and formant, which causes severe spectrum distortion.

Varga and Fallside proposed a pitch extension method using LPC coefficients. However, this method shows a lot of spectrum distortion because it only uses a method of canceling and smoothing a part of the waveform when decreasing the pitch period.

On the other hand, the pitch half-method is a pitch change method in which a period of a waveform is changed by half after making a double waveform of a pitch period to be arbitrarily changed.

However, since this method is performed only in the time domain, spectrum distortion occurs and the clarity of the synthesized sound is lowered.

McAulay and Quatieri proposed a pitch change method that preserves the phase in the frequency domain. This method extracts the amplitude and phase spectrum of the input speech and processes it separately.

That is, for the amplitude spectrum, we extract the prominent spectral peaks and then interpolate them by the pitch change rate p to change the pitch of the amplitude spectrum.

For the phase spectrum, a phase corresponding to the pitch on-set time obtained in the time domain is removed, and a new phase is formed by adding the phase of the new pitch start time when the pitch is changed.

This method has the merit that the connection between neighboring frames becomes very easy in a usual processing method of performing analysis processing on a frame basis since the shape of a waveform remains unchanged.

However, when changing the pitch, the start time of the pitch must be supplied separately from the pitch period, and the distortion of the spectrum is increased because the interpolation of the harmonics is performed mainly on the prominent peaks on the amplitude spectrum.

In the time-frequency hybrid processing method, there is a method of changing the pitch by inserting or deleting a zero value at a portion where the cepstrum value becomes almost zero using the characteristics of the cepstrum.

However, this method also has a problem that it is difficult to maintain the phase.

The time-frequency hybrid method proposed by Takagi and Miyasaka is a method of modifying the spectrum distortion caused by the pitch change in the time domain through the LPC envelope on the spectrum region.

This method has the limitation that it can not satisfy all the voices because of the spectral transfer characteristic focused on the pole of the LPC spectrum envelope.

However, if the pitch period of a voiced sound is increased to r times by a simple average interpolation when the time domain method is used, the pitch period naturally increases. However, The characteristics are changed, and the intelligibility of vocalization is lowered.

When the pitch period of the standard sentence used in the present invention is extended by the mean value interpolation method and the spectrum distortion of the original speech is measured and expressed by a percentage of 100, the spectrum distortion increases exponentially as the pitch change rate increases , And the clarity is also considerably low.

The reason for this is that not only narrowband but also broadband signals are scaled together.

The present invention firstly separates components of a speech signal into a narrowband component and a wideband component using a variable low pass filter, maintains a wideband component and performs time scaling only on narrowband components, A method of changing a pitch in a time domain by separating components by a method of constructing a speech signal whose pitch is changed by adding wideband components is provided.

That is, in order to increase the clarity of a synthetic sound, the present invention proposes a pitch changing method in which a narrow-band signal and a wide-band signal are separated and a period of a narrow-band signal is changed without changing a wide-band signal.

1 is a block diagram of a speech signal processing system for implementing the present invention;

2 is a block diagram of a speech pitch changing method according to the present invention.

3 is a block diagram of a formant detector applied to the present invention.

The configuration of a speech signal processing system for realizing the present invention is shown in Fig.

1, there is shown a microphone 1 for converting a voice into an electrical voice signal, an amplifier 2 for amplifying a voice signal input to the microphone 1, a voice amplified by the amplifier 2, A low pass filter 3 for filtering a signal, an analogue digital converter 4 for converting a voice signal output from the low pass filter 3 into a digital signal, A memory 7 for storing voice data processed by the microprocessor 6; a memory 7 for storing voice data processed by the microprocessor 6; An input / output port 8 for transmitting voice data output from the microprocessor 6 on a transmission channel; an output port 9 for voice data processed by the microprocessor 6; Voice output A low-pass filter 11 for filtering the converted analog voice signal; an amplifier 12 for amplifying the voice signal output from the low-pass filter 11; And a speaker 13 for outputting the amplified voice signal in the audio frequency band.

The voice inputted to the microphone 1 is converted into an electrical signal and amplified by the amplifier 2 to a predetermined degree of amplification.

Since the voice signal amplified by the amplifier 2 requires only a pseudo-propagation information component, it passes through the low-pass filter 3 to remove a frequency component of 4 kHz or more, is sampled at a clock of 8 kHz in the analogue digital converter 4, Is digitally converted to a 12-bit quantization level based on the telephone sound quality.

The digital voice data is input to the microprocessor 6 through the input port 5 and encoded and the processed data is stored in the memory 7 or transmitted through the input /

On the other hand, in order to confirm whether the synthesized voice signal, which has been decoded by using the voice data processed in the microprocessor 6 or the data inputted through the input / output port 8 read from the memory 7, Analog converter 10 through the output port 9. The digital-to-

The digital-to-analog converter 10 converts the digital voice data into an analog voice signal with a clock signal of 8 kHz, passes through a low-pass filter 11 to filter only a signal of a fundamental band from which a harmonic component is removed, And outputs it to the speaker 13.

FIG. 2 is a block diagram of a pitch changing method in the above-described speech data processing, which includes a strong band detecting means 201 for receiving the speech signal S (n) and separating it into a wideband component and a narrowband component, A time scaling means (203) for scaling the narrowband speech signal S _L (n) filtered by the low pass filter (202) in a time domain; a time scaling means And an adding means 204 for adding the band speech signal S _L (n ') and a wideband component.

In the band separating means 201 and 202, the narrowband signal is composed of a first formant and a second formant, and the formants therefor are broadband signal components. Particularly, since the formant frequency of the voiced sound changes, To isolate the wideband signal, the low pass filter 202 employs a cutoff frequency (f _T ) variable low pass filter.

The wideband detecting means 201 receives the audio signal S (n) and detects the high frequency zero crossing rate included in the original voice signal from the peak valley rate (PVR) and the zero crossing rate from the low- with crossing rate and variably controls the cut-off frequency f _T of the variable low-pass filter 202 to remove components of broadband signal component and a narrowband signal component.

By changing the pitch by performing time-domain scaling in the time scaling means 203 only for the separated narrowband signal component S _L (n), the scaled narrow-band signal component S _L (n ') and the wide- And added by the addition means 204 to reconstruct a signal whose pitch is finally changed.

That is, the pitch period of the narrowband signal is expanded by r times by a simple average interpolation method or compressed by r times by a decimation method. Since the present invention focuses on the preservation of a wideband signal component, The time-scaling algorithm is used for pitch change to scale the narrowband signal component, and the narrow-band and wide-band components thus changed in pitch are summed in the adding means 204 to reconstruct the pitch-changed signal.

The configuration of the formant detector as the band separating means used in the pitch changing method of the present invention is shown in Fig.

(N) of the speech signal S (n) and the output signal S '(n) of the variable low-pass filter 301 that performs low-pass filtering of the input speech signal S (n) A first calculation means 302 for estimating a high frequency zero crossing rate Z _SM included in the original voice from the peak value rate PVR of the input voice signal S (n) Second calculation means for outputting a zero crossing rate Zi of a wide band formant component of the original speech signal from the zero crossing rate (ZCR) of the error signal from the error signal e (n) output from the subtracting means 306 Frequency zero crossing rate Z _SM output from the first calculating means 302 and the zero crossing rate Z i of the error signal output from the second calculating means 303 to a point for determining the cutoff frequency of the variable low- A third calculation means 304 for calculating the second calculation means 304, Comparing the magnitude values STE and is configured to include a comparison control means 305 for controlling the cut-off frequency f _T of the variable low-pass filter 301.

The operation of the formant detector thus constructed is as follows.

The difference value between the speech signal S '(n) output from the low-pass filter 301 and the input speech signal S (n) is calculated by the subtracting means 306 to obtain an error signal e (n).

Since the error signal e (n) is an excess signal composed of high frequency components, the zero crossing rate ZCR of the error signal e (n) is equal to the zero crossing rate Zi of the wideband formant component of the original speech signal.

Therefore, the second calculation means 303 outputs the zero crossing rate Zi and inputs it to the third calculation means 304 while the high frequency zero crossing rate Z _SM included in the original voice is obtained from the peakvalue rate (PVR) of the original signal 2 calculation means 302 estimates it and inputs it to the third calculation means 304. [

Since the peak-to-peak ratio (PVR) of the voiced sound consists of about 70% of the peak-to-peak ratio of the higher-order formants and the peak-to-peak ratio (PVR) of the unvoiced sound consists of the peak- cut-off frequency of the) f _T may be determined at the point where the Zi = 0.7 × Z _SM.

Therefore, the third calculation means 304 calculates this point (point) and inputs it to the comparison control means 305. The comparison control means 305 compares Zi with 0.7Z _SM to control the cut-off frequency f _T .

That is, the cut-off frequency f _T is initially initialized at a predefined low frequency, and if Zi is less than 0.7 x Z _SM , the cut-off frequency f _T is increased and repeatedly compared until Zi becomes larger than 0.7 x Z _SM , And separates the components so that a pitch change can be made to the narrowband signal component.

In this manner, the narrowband signal component is obtained, scaled, and the result is summed with the wideband signal component to reconstruct a signal whose pitch is finally changed.

The pitch changing method according to the present invention separates a speech signal into a wideband component and a narrowband component in a time domain, performs time scaling only on a narrowband component while maintaining a wideband component, combines the wideband component with a wideband component, And constitutes a voice signal.

Therefore, it is possible to secure a synthesized voice signal with improved clarity compared to the conventional techniques.

As a result, when the pitch period is extended up to 200%, the spectrum distortion rate is 5.9%, which is about 3% higher than the conventional method.

Further, since the present invention retains the original higher-order formant component of the original voice, there is almost no deterioration in clarity. The present invention can be applied to the synthesis method using phonemes by applying to the high-quality synthesis method. .

Claims (3)

The method comprising the steps of: maintaining a wideband component by separating a speech signal into a wideband component and a narrowband component; performing scaling in the time domain with respect to the separated component of the narrowband component; And constructing a speech signal that is finally pitch-changed by adding the held wideband components to the speech signal. 2. The method of claim 1, wherein narrowband components are separated using a variable low-pass filter in the component separation step. 3. The method of claim 2, in Zi = 0.7 × Z _SM point is from the high-frequency zero crossing rate Z _SM included in the zero crossing rate Zi and the original speech in the original wideband formant components of the audio signal to set the cut-off frequency of the variable low-pass filter ≪ / RTI > wherein the temporal domain of the voice pitch is changed.