KR19980035868A - Speech data encoding / decoding device and method - Google Patents

Abstract

According to the present invention, in the speech signal processing of a digital communication system, the amount of data required is reduced by encoding and decoding a speech signal using a pitch synchronous interpolation method according to a pitch change, thereby simplifying a structure and reducing a calculation amount. The present invention relates to a voice signal processing system and a method of controlling the same, which facilitates real-time processing and improves sound naturalness and clarity. The matrixing process constituting the matrix, and the set sequence between the first pitch period set sequence and the last pitch period set sequence among the pitch period set sequence constructed in the matrix process is removed, and only the first pitch period set sequence and the last pitch period set sequence The decimation process to be transmitted and the first to be transmitted by the decimation process. It is composed of an interpolation process of reconstructing the sets between the two pitch period set sequences by the difference signal between the second pitch period set sequence and the last pitch period set sequence and the number of decimated pitch period set sequences to reproduce the original audio signal. Voice can be applied to input / output methods such as mouse, monitor, and keyboard used for communication between machines, and it is effective to synthesize even a small amount of data in text-to-speech field.

Description

Speech data encoding / decoding device and method

1 (a) to (c) are diagrams showing an example waveform according to a general pitch synchronous interpolation scheme.

2 is a block diagram of an encoder according to the present invention.

3 is a block diagram of a decoder in accordance with the present invention.

4 is a diagram showing an interpolation process according to the present invention.

5 is a waveform diagram of audio signals in one frame, and (b) is diagram of audio signal waveforms according to the present invention.

* Description of the symbols for the main parts of the drawings *

10: voice / silent sound separation unit 20: pitch detection unit

30: matrix part 40: decimation part

50: Gaussian noise generator 60: Decoder

70: encoder 80: first pitch period set sequence storage unit

90: last pitch period set sequence storage unit 100: adder

110: pitch set difference signal compensation unit 120: interpolation unit

According to the present invention, an amount of data required by encoding and decoding voice data by using pitch synchronous interpolation (hereinafter, referred to as 'PSI') method in accordance with a pitch change in a digital data processing system. The present invention relates to a speech data encoding / decoding apparatus and a method for simplifying a structure and reducing a calculation amount in real time, and for improving the naturalness and clarity of sound.

In general, a lot of information pouring in a rapidly changing industrial society is transmitted through video or audio, and in particular, the information transmission method using the voice is a communication method that has been used for a long time as a communication means or a conversation means.

Recently, in the voice signal processing of digital communication that digitally stores, transmits and receives a voice signal when the communication method is digitalized, many voice synthesis techniques, encoding, and decoding techniques for maintaining high sound quality with a small number of bits have emerged. It is becoming.

In the speech encoding / decoding technique, a method of encoding and storing a speech includes a waveform coding method and a speech production process of removing and storing a repetitive and unnecessary surplus present in a speech signal and storing and encoding the speech. model is considered as a filter and is synthesized with the above-mentioned waveform coding method among the source coding method and the signal source coding method for encoding as an excitation filter of an excitation source and a filter of a vocal tract. There is hybrid coding.

The waveform coding method removes only the excess part in the time domain and encodes the audio so that its naturalness and intelligibility are very high. However, the waveform coding method has a disadvantage in that it is not efficient in memory due to the large amount of data required for transmission. Methods include pulse code modulation (PCM), delta modulation (DM), adaptive modulation (ADM), adaptive pulse code modulation (ADPCM), and the like.

In addition, based on the speech generation process, a speech is generated by a sound source and a vocal filter, and a signal source encoding method that artificially encodes the model by converting it from the time domain to the frequency domain is used to generate excitating and formant components. (formant) is treated separately.

In particular, two methods are further divided to encode the formant component and the excitation information. First, the formant component encoding method includes a linear prediction coding (LPC), a linear spectrum pair (LSP), and a PARCOR. This method has the advantage of efficient memory usage by transmitting feature parameters.

However, when the transition of the sound, the start / end of the sound, and the repetition of the voiced / unvoiced sound cannot be modeled using the characteristic parameters alone, deterioration of the sound quality occurs. From the pole model to the zero-pole model, there is a drawback that the sound is less natural and clear.

In addition, synthesis by analysis is mainly used to code excitation information, and a codec having a low bit rate using code excited linear prediction (CELP), vector sum linear prediction (VSELP), RELP, etc. codec) is implemented.

However, the encoding method also requires a repetitive calculation process and a comparison process in order to ensure a high sound quality at a low bit rate, resulting in a large amount of calculation and a complicated structure.

Recently, in order to solve the above disadvantages, the PSI method is used to reduce the amount of data of a speech signal having a quasi-periodical period. The PSI method pitches repeated pitch periods or excitation information. It is applied to synthesis or waveform editing techniques by rules by repeatedly synthesizing cycles repeated in synchronization with periods.

That is, (a) of the example waveforms according to the PSI method of Fig. 1 is a waveform diagram of the voiced sound of the voice signal, (b) shows a representative pitch interval of the repeated pitch set of the (a) voice signal, (C) is a waveform diagram obtained by repeatedly synthesizing the pitch period of (B) using the PSI method. When comparing the waveforms of (A) and (C), the synthesized audio signal of (C) is the original. It can be seen that it is not played back exactly like the voice signal as (A).

The conventional speech coding and synthesis method using the PSI method also uniformly synthesizes some waveforms, which causes energy compensation problems due to stress changes, and it is difficult to accurately detect pitches, and the sound changes and the pitches are severe. In the sound of changing intervals or strong stress and intonation, there is a problem that encoding of voice data is very difficult because not only serious degradation but also many calculation time are required.

Accordingly, the present invention has been devised to solve the above problems, and an object thereof is to provide an encoding process for transmitting only components within a representative pitch period and a decoding process by interpolation using a PSI method such as matrixing and decimation. The present invention provides a speech data encoding / decoding apparatus and a method for reducing the amount of data required through the real-time processing with a simple structure and a small amount of calculation, and improving sound naturalness and clarity.

The speech data encoding / decoding apparatus of the present invention for achieving the above object, when encoding the speech data, detects the pitch of the input speech data, and performs a matrixing process for constructing a pitch junction in units of the detected pitch; After performing the decimation of the sets between the first pitch set and the last pitch set, performing an encoding process for transmitting only the first pitch set and the last pitch set, and during decoding, the difference between the first pitch set and the last pitch set. The decimated pitch sets are restored by the signal to reproduce the original voice data.

Hereinafter, with reference to the accompanying drawings will be described in detail the speech data encoding / decoding apparatus and method thereof.

2 is a block diagram of a decoder for encoding voice data according to the present invention. As shown in FIG. 2, a digital voice data which is converted and output by the digital converter 5 is inputted and separated into voiced sound and unvoiced sound. On the basis of the unvoiced sound separation unit 10, the pitch detection unit 20 for detecting the pitch by inputting only the unvoiced sound separated by the voice / unvoiced sound separation unit 10, and the pitch detected by the pitch detection unit 20 as a unit The matrixing unit 30 constituting the pitch period set sequence and the first and the last pitch cycle set sequence among the pitch cycle set sequence configured in the matrixing section 30 are decimated and removed. When the unvoiced sound is separated and output from the decimation unit 40 and the voiced / unvoiced sound separating unit 10, the unvoiced sound is added to the decoder 60. The Gaussian noise generating unit 50 generates Gaussian noise for replacing with the Gaussian noise.

3 is a block diagram of a decoder for decoding voice data according to an embodiment of the present invention. As shown therein, a first pitch period set sequence and a last pitch period set sequence transmitted by the encoder 70 are respectively stored. A pitch period set sequence storage section 80 and a last pitch period set sequence storage section 90, an adder 100 for adding a difference between the first pitch period set sequence and the last pitch period set sequence, and the adder 100 A pitch set difference signal compensator 110 for compensating the decimated pitch sets by an amplitude and a number by a difference signal of the two pitch period set sequences added at, and an amplitude compensated by the pitch set difference signal compensator 110. The interpolation unit 120 is configured by restoring the decimated pitch period set due to the number of times.

Referring to the speech data encoding and decoding method of the present invention configured as described above are as follows.

First, before encoding the voice data, the voice data is separated into voiced sound and unvoiced sound, and then encoding is started.

If the input voice signal is voiced, the voice data is divided into one frame for 256 samples. The pitch is detected between the peaks and peaks or valleys and valleys in the frame. do.

That is, the correct pitch is detected by the pitch detection equation using the autocorrelation method as shown in Equation-1 below.

[Equation-1]

Subsequently, the detected pitches are reclassified in units of pitches to form a pitch period set sequence of the two-dimensional matrix.

That is, the index period p is assigned to each pitch period within one frame to form a pitch period set X _p .

And decimating the set sequence between the remaining two pitch set sequences except for the first pitch P _F first detected and the last pitch P _L last detected in the configured pitch period set sequence. Audio data is encoded.

At this time, the matrix is composed of the pitch period axis (X) and the pitch number axis (P) so that the voice data in the frame is expressed as in Equation-2 above.

[Equation-2]

s (n) = P _F (k) + P _D (k) + P _L (k), k = 1, 2,... , pitch period

That is, the process of the decimated components has a two-dimensional determinant after performing matrixing on a pitch period basis, as shown in Equation-3 and 4, respectively.

Equation-3

[Equation-4]

Here, S is a matrix divided into K divided by pitch period units (P), and the finally transmitted samples are the first and last pitch period sample sets.

Thus, the sample set transmitted to the decoder is only two sets P _F (X) and P _L (X).

In the encoding process using the PSI method by matrixing and decimation as described above, only components within a representative pitch period of the first pitch period set sequence and the last pitch period set sequence are transmitted to the decoder.

Meanwhile, in a decoder that receives two pitch period set sequences transmitted from the encoding, a difference between the first pitch period set sequence and the last pitch period set sequence is used to restore the pitch set P _D (n) decimated in the encoding. (Δd (k)) is obtained by the expression-5.

Equation-5

Δd (k) = P _F (k) = P _L (k), k = 1,2,... , pitch period

Here, P _F (k) is the sample string within the first detected one pitch period, and P _L (k) is the sample string within the last detected one pitch period.

This is to compensate for the amplitude and number of pitch sets removed by the decimation process in the encoder.

Subsequently, the number Occ of the decimated pitch period set sequence is calculated, and the interpolation samples are defined by dividing the value of the difference Δd (k) by the number Occ according to Equation-6.

[Equation-6]

P _D (k, t) = Δd (k) / Occ, t = 1,2,... , Occ

k = 1,2,... , pitch period

Where t is the number of decimation pitch sets and k is the sample in the pitch period.

The original speech data is reproduced by interpolation between the amount of change obtained in Equation 5 and the difference set between P _F (X) and P _L (X).

That is, as shown in the figure showing the interpolation process according to the present invention of FIG. 4, it is possible to accurately see the samples transmitted and the interpolated samples in the decoder of the present invention.

On the other hand, if the input voice is an unvoiced sound, the voiced / unvoiced detection stage replaces the unvoiced sound with Gaussian noise to process it.

It can be seen that the audio signal waveform diagram of FIG. 5 (b) has better clarity and naturalness than the waveform diagram of (a) through the encoder and the decoder according to the voice signal processing control method of the present invention.

As described above, the present invention can reduce the data required by interpolating the pitch period sample decimated from the sample in the first pitch period to the sample in the last pitch period after the matrixing is completed in the PSI method. Only a computational amount of detection and matrixing is required, so a simple structure can be achieved.

In addition, it can be applied to PCM communication as it encodes voice at low bit rate in digital communication. In particular, it can be applied to input / output methods such as mouse, monitor, and keyboard used for communication between human and machine. In the field of -to-speech, there is an effect that can be synthesized with a small amount of data.

Claims (2)

An apparatus for decoding and decoding voice data in a digital communication system, A voice / unvoiced sound separation unit 10 for inputting digital voice data to separate voiced and unvoiced sounds, a pitch detector 20 for detecting a pitch by inputting only voiced sounds separated by the voiced / unvoiced sound separation unit 10, and The matrixing unit 30 constituting the pitch period set sequence based on the pitch detected by the pitch detection unit 20, and the first pitch period set sequence and the last pitch period set of the pitch period set sequence configured in the matrix unit 30. When the decimation unit 40 transmits only the first and last pitch period sets after removing the set sequence between columns, and the unvoiced sound is separated and output from the voiced / unvoiced separator 10, the unvoiced sound is Gaussian noise. An encoder composed of a Gaussian noise generator (50) for generating Gaussian noise to replace with; A first pitch period set string storage unit 80 and a last pitch period set string storage unit 90 respectively storing a first pitch period set string and a last pitch period set string transmitted from the encoder; An adder 100 for adding a difference between the first pitch period set sequence and the last pitch period set sequence of the heat storage unit 80 and the last pitch period set sequence storage unit 90, and the two pitches added by the adder 100. Pitch set difference signal compensator 110 for compensating the decimated pitch set by amplitude and number by the difference signal of the period set sequence, and decimated by the amplitude and number compensated by the pitch set difference signal compensator 110. Encoding and decoding field of the voice data, characterized in that consisting of a decoder consisting of an interpolation unit 120 to restore the original pitch data by restoring the set of pitch periods . In the audio data encoding and decoding method in a digital communication system, A matrixing process of detecting a pitch of an input audio signal and constructing a pitch period set sequence based on the detected pitch as a unit, and a first pitch period set sequence and a last pitch period set among the pitch period set sequences configured in the matrixing process A decimation process of removing the set sequences between columns and transmitting only the first pitch period set sequence and the last pitch set, and a difference signal between the first pitch period set sequence and the last pitch period set sequence transmitted by the decimation process; And an interpolation process of reconstructing the sets between the two pitch period set strings by the number of decimated pitch period set strings and reproducing the original voice signal.