JPH0364800A - Audio encoding and decoding method - Google Patents

Abstract

PURPOSE:To improve voice quality and shorten a processing time by encoding the value of an intersection level of each frame, the number of intersections, data showing which of positive and negative levels is crossed, and a time interval between an intersection and a last intersection. CONSTITUTION:This system has a means 10 which puts each specific number of samples in a frame and a means 20 which calculates a positive-negative intersection level (threshold value) from the values of the respective samples in the frame, and the value of the intersection level of each frame, the number of intersections, data indicating which of the negative and positive levels is crossed at each intersection, and the time interval (sample interval) between one intersection and its last intersection are encoded. Consequently, the voice quality is improved and the processing time is shortened.

Description

[Detailed description of the invention] branch tip summer The present invention relates to audio encoding and decoding systems.

Various voice transmission technologies, including long-distance telephones, have been deeply researched and have made progress. Furthermore, research into digital communication systems, which originated from the PCM system invented in 1938, is still ongoing in the direction of lowering the pit rate. Digital audio processing is indispensable when transmitting audio, etc. using this communication method. A typical audio processing method is ADPCM as a waveform encoding method.
(N, S, JayantP, Cumm1nskey
and J. L., Flanagan, Adaptiv.
e quan-tization in diffe
rental pen carding of 5p
e-ech, Be1l 5yst, Tech, J,, 5
2, 7.1105-1118.1973. ), and the auditory encoding method is ATC (R, Zelinskia
nd P, No11. Adaptive transf
orm coding of speech sign
als, IEEE Trans, ^5SP-25,2
99-309゜1977,), PARCOR (Hokuma, Itakura,
Saito.

Optimal code structure in Parcor type speech analysis and synthesis system.

IEICE Theory (A), J61-A, 2.119-126.1
978. ), L S P [Pipe Material, Itakura, Speech Information Compression Using Line Spectral Pair (LSP) Speech Analysis and Synthesis Method, IEICE Theory (A), J64-A, 8,599-606.198
1. ) have been proposed.

All of these encoding methods aim to reduce the amount of information transmitted while minimizing the deterioration of sound quality. There is a problem of increased processing time. but,
In recent years, advances in LSI technology have made it possible to perform high-speed arithmetic processing, and as a result, low bit rate high efficiency of 16 kbps or less is aimed at reducing transmission costs in digital communication networks and effectively utilizing frequency bands in mobile and satellite communications. Research on speech coding methods is actively underway.

And--"0 The present invention was made in view of the above-mentioned circumstances,
In particular, 16k is expected to be in demand in the future in mobile communications, etc.
This was developed with the aim of providing an encoding method that combines time domain and frequency domain and requires relatively little processing time as a method for audio encoding with a code amount of less than bps.

(2) In order to achieve the above object, the basic invention provides: (1) In a compression method for a PCM encoded audio signal, means for forming a frame every predetermined number of samples; It has a means for calculating positive and negative crossing levels (thresholds) based on the data, and a code indicating the crossing level value for each frame, the number of crossing points, and which level (positive or negative) is crossed for each crossing point. and encoding the time interval (sample interval) between the intersection and the previous intersection, or (2
) In the method for decoding the encoded signal in (1) above, when the intersection intersects with the previous intersection at a different code level and the interval is larger than a predetermined value, the distance between the two points is 3. Interpolate with a quadratic curve, and when the interval is smaller than a predetermined value, interpolate between the two points with a straight line, and when it intersects a level with the same sign as the intersection, interpolate between the two points with a quadratic curve. , or (3) In the compression method of a PCM encoded audio signal, the input PG
The M audio signal is framed in a predetermined section, the pitch is extracted in that section, and when pitch extraction is possible, the pitch section is used as a new frame and a positive/negative crossing level (threshold value) is set based on the value of each sample in the frame. ), and calculates the intersection level value for each frame, the number of intersections, a sign indicating which level (positive or negative) each intersection has crossed, and the relationship between the intersection and the previous intersection. Time interval(
When performing local decoding from these encoding results, if this intersection intersects at a different code level from the previous intersection and the interval is larger than a predetermined value, A cubic curve is used to interpolate between points, and when the interval is smaller than a predetermined value, a straight line is interpolated between the two points, and when the intersection intersects a level with the same sign as the intersection, a quadratic curve is used to interpolate between the two points. The difference between this local decoding result and the original signal is taken, and the signal is Fourier transformed, and the frequency and amplitude of the peak point of the spectrum of each harmonic, the value of the cross level for each frame, and the intersection point are calculated. , a code indicating which level (positive or negative) has been crossed for each intersection, and a time interval (sample interval) between the intersection and the previous intersection, or (4) the above ( In the method for decoding audio encoded data shown in 3), the value of the crossing level for each frame, the number of crossing points, a code indicating which level (positive or negative) is crossed for each crossing point, and the crossing point. Based on the time interval (sample interval) between the intersection and the previous intersection, if the intersection in the frame intersects at a different sign level from the previous intersection, and the interval is larger than a predetermined value, the distance between these two points is 3. Interpolate with the following curve, and if the interval is smaller than a predetermined value, the 2
A straight line is interpolated between the points, and when the intersection crosses a level with the same sign as the intersection point, decoding is performed by interpolating a quadratic curve between the two points, and the frequency and amplitude of the peak point of the spectrum of each harmonic are This method is characterized in that decoding is performed by adding a voice signal decoded by performing inverse Fourier transform from a code indicating . The present invention will be explained below based on examples.

The encoding method according to the present invention is based on the adaptive level crossing method, which is time-domain encoding [Koukashi, Akita, Mori, Coding of speech waveforms by two-level crossing method, IEICE National University, 1367.1986.
), (Kakashi, Mori, Speech Waveform Coding Using Adaptive Level Crossing Method, IEICE University, 1325, ↓987.) and Harmonic Coding, which is frequency domain coding (
R, J, McAulay and T, F, Qua-t.
iefi, 5peech analysis/5ynt
hesis based on 5i-nusoida
l representation, IEEE Tra.
ns, ASSP-34゜774-754.1986. )
, (L, B, A11nsida and J, M, T
ribolet, Non5tationary 5pe
ctral modeling of votes
peach, IEEE Trans, ASSP-31,
664-678, 1983).

[Inuyama, Two speech analysis and synthesis systems using the harmonic structure of speech and residual signals, Speech Research Fund, 584-50゜1984,
), (Matsuyuki, Tada, Study of medium band coding method using harmonic structure of voice, IEICE Technical Report, 5P88-57.1988
．． 3 is combined, pitch extraction is performed when using the adaptive level crossing method, and the pit rate is reduced by encoding one pitch period in one frame.

First, an overview of the adaptive level crossing method will be explained. It is known that when adjacent samples have values of different signs (positive or negative), they cross zero, and by using these zero crossings, it is possible to estimate the short-term average frequency. It is thought that it is difficult to reproduce good synthesized speech if encoding is performed using only these zero crossings. A pole point extraction method has been proposed as an encoding method that lowers the bit rate by encoding the pole points rather than the zero crossings in waveform encoding [Matsuzai.

Single motion, Yoshida, Suda, Speech coding using the polar point extraction method (PVP method), IEICE Technical Report, C385-32, 1985]. The encoding method of this pole point extraction method encodes the time interval between pole points and the difference in amplitude between pole points, and performs interpolation between pole points using these parameters.

Similar to the pole extraction method, an adaptive bell crossing method has been proposed in which coding is performed using crossings at positive and negative levels by applying zero crossings as a method of reducing the bit rate by waveform encoding [Takanashi, Aki UJ1. Mori, Coding of speech waveforms using two-level crossing method, IEICE, Engineering 367. 1986. ),
(Takanashi, Mori, Coding of speech waveforms by adaptive level crossing method, IEICE, 1325.1987.1゜In this coding method, the sign of the time interval at which the positive or negative set level is crossed is This encoding is a method of interpolating crossing points, so it only encodes the outline of the audio waveform.Therefore, it is used to encode the low-frequency components of the audio waveform.

Next, audio encoding will be explained.

FIG. 6 is a flowchart for explaining an example of audio encoding. First, from the audio waveform of 1 frame, the average amplitude a
mp is calculated using equation (1).

The positive and negative crossing levels are determined by multiplying this average amplitude by a coefficient. The intersection of the audio waveform that intersects the determined positive and negative levels is encoded. If a crossing level comes between two adjacent points, the one that is closer to the crossing level is determined as the crossing point.

The parameters to be encoded are (1) For each frame 1.Cross level 2. Number of intersections (II) At each intersection 1. Distinguishing whether the level is positive or negative 2. Time interval T between the intersection and the previous intersection.

Next, decoding will be explained.

Decoding is performed based on the parameters transmitted from the encoder, and FIG. 7 is a flowchart to explain the method. This section is interpolated with a cubic curve. 2. When the time interval is small, this section is interpolated with a straight line. (II) When the intersection has the same sign level as the previous intersection, this section is interpolated with a quadratic curve.

FIG. 8 is a diagram showing the basic configuration of adaptive level encoding. In the figure, 10 is an encoding section and 20 is a decoding section.

How to determine coefficients when interpolating between intersections will be explained below. This will be explained separately from the intersection.

First, we will explain when the intersection is at a different sign level from the previous intersection.The way to take the intersection at this time is when the time interval T is large (
There are two types: (Figure (a)) and when the time interval T is small (Figure (b)).

Now, assume that the current intersection is PL (T, -cd) and the previous intersection is PO (0, cd). When the time interval T is large, for example (T>10), interpolation is performed using a cubic curve.
Defined in (2). Coefficient at this time bt at
Let me explain how to determine d.

In order to make PO and PL equal, use the interpolation formula % formula %
(3) (4) Also, so that the decoded waveforms are smoothly connected, the slope of the previous intersection PO is made equal to the slope of the previous interpolated curve or straight line. Assuming that the slope of the intersection point PO on the previous interpolated curve or straight line is m, the interpolation formula can be differentiated to make it equal.

Then, pass through O at T/2 Then, the coefficients at b, c, and d are as follows.

C= m d=cd However, since the cubic curve to be interpolated should not have an intersection between PO and Pl, the interpolation formula is T/4 and d = c
d Also, when the time interval T is small, for example (T≦10)
When , interpolation is performed using a straight line, but if the interpolation formula using the straight line is defined as f(t)=at+b (12),
Since it passes through PO and PL, the coefficient a.

b becomes as follows.

should hold true. If the above equation does not hold, then set it to -c d / 2 at T/4, and remove from the condition that the slope is equal to the previous curve or straight line at the intersection PO, then the coefficient a.

b, c, and d are as follows.

b=cd Next, we will explain how to take an intersection when it has the same sign and level as the previous intersection.
As shown in FIG. 9(Q), one type of interpolation is performed using a quadratic curve. As before, let the current intersection be PL (T, cd)
, the previous intersection is PO(0, cd). When two intersections have the same sign, interpolation is performed using a quadratic curve, and the interpolation formula using the quadratic curve is defined as f(t)=a t2+b t+c (14). In order to make PO and PL equal, the interpolation formula is f(0)=cd=c (15) f(T)=cd=aT''+bT+c (16) Also, the decoded waveform is smooth. To connect, 1
In order to make the slope equal to the slope of the previous interpolated curve or straight line at the previous intersection PO, let m be the slope of the previous interpolated curve or straight line at the previous intersection PO. , we differentiated the interpolation formula, but we just need to make it equal.

Here, since the quadratic curve to be interpolated should not have an intersection between PO and Pl, the interpolation formula is T/2 and f (T/2
) should hold true when cd has a different sign. If the above equation (20) is expressed as %, and if we exclude from the condition that the slope of the previous curve or straight line at the intersection PO is equal, then the coefficient a.

b and c are as follows.

Therefore, the coefficients a, b, and c are as follows.

a = -1b=m (19) c=cd c=cd Also, if f(T/2) has the same sign at T/2 and becomes larger than 3cd, then f(T/2) is too large. In that case, T/2 is set to 2 cd, and if we remove from the condition that the slope is equal to the previous curve or straight line at the intersection PO, the coefficient a.

b and c are as follows.

c=cd Next, harmonic coding (HarmonicC
(ordering) will be explained.

In a voiced period of speech, the short-time Fourier spectrum generally has a harmonic structure (a periodic spectral structure consisting of a fundamental frequency and its harmonics) due to the periodicity of the signal. As an encoding method that focuses on this harmonic structure, Harm
The onic coding (hereinafter referred to as HC) method has been proposed (R, J, McAulay and T,
F, (luatiefi, 5peech analyzes
is/5 synthesis bas-ed on 5i
Nusoidal representation, I
EEE Trans.

ASSP-34, 774-754, 1986), (
L, B, A11neida and J.

M, Tribolet, Non5tationary
5 pectral modeling of voi
ced 5peech, IEEE Trans, ASS
P-31, 664-678゜1983,), (Inuyama, Two speech analysis and synthesis systems using the harmonic structure of speech and residual signals, Voice Research Fund, 584-50.1984.),
(Matsuyuki, Tada, Study of the middle '41F range code system using the harmonic structure of speech, IEICE Technical Report, 5P88-57, Eng. 98
8]. This expresses the harmonic structure in the frequency domain as a line spectrum and decodes the audio waveform using the weight of a sine wave, so the model is very simple and the number of parameter quantization bits can be significantly reduced.

FIG. 10 is a diagram showing a typical configuration of HC, and in the figure, 1
0 indicates an encoder, and 20 indicates a decoder. In the encoder 10, first, a Hamming window is applied to the input audio in frame units (approximately 20 to 32 msec), and then the frequency domain is processed by Fourier transform (DFT). Convert to next,
The peak point of each harmonic in the amplitude spectrum of each frequency is extracted (peakPicking), and the frequency, amplitude, and phase of the extracted peak point are transmitted as encoding parameters. The decoder 20 reproduces the audio signal based on the transmitted parameters. When the frequency of the decoded parameter is fl, the amplitude is ml, and the phase is φ, the encoded speech s (t) is expressed by the linear equation (25).

The harmonic components are extracted by extracting each peak of the harmonic structure from the amplitude component of the converted spectrum.

Here, the peak refers to the center point when the center of the amplitude spectrum values of five adjacent points among the FFT points is the maximum, and the points on both sides become smaller in order.

Next, encoding using the adaptive level crossing method that takes high frequency compensation into consideration will be explained, but first, the causes of audio deterioration caused by the adaptive level crossing method will be explained.

FIG. 11 is a diagram showing the amplitude spectrum of a voiced sound section of speech. Looking at the diagram, a clear peak of the spectrum called a formant can be seen.

Usually, voiced sounds, especially vowels, have about three characteristic formants, and even the same phoneme can vary considerably depending on the speaker. Therefore, it is necessary for the synthesized speech to have the same spectral envelope as the input speech.

Figure 12 shows the amplitude spectrum of the synthesized speech when encoding is performed using the adaptive level crossing method. Looking at the amplitude spectrum of the synthesized speech, we see that for low frequency components, the input speech There is almost no difference between
If you look at the high frequencies, they will deviate considerably from the input audio. Therefore, although the first and second formants are reproduced, the spectral envelope of the amplitude at higher frequencies cannot be reproduced, and the formants after the third formant cannot be reproduced. If encoding is performed using only the adaptive level crossing method, high frequency components cannot be reproduced, and this is considered to be a factor that leads to deterioration of synthesized speech.

FIG. 2(a) is a basic configuration diagram for explaining an embodiment of the encoding method according to the present invention.

(b) is a flowchart for explaining the operation of the encoder, and FIG. 3 is a flowchart for explaining the operation of the decoder. In FIG. 1, 10 is an encoder and 20 is a decoder.

As mentioned above, high frequency components cannot be reproduced simply by encoding using the adaptive level crossing method. Therefore, it is necessary to regenerate the high frequency acid. The aforementioned HC is used to reproduce high frequency components. Since the HC used here is intended to reproduce only high frequency components, the frequencies used when weighing a sine wave are only high frequencies. Normally, when encoding is performed using HC, frequency, amplitude, and phase are transmitted as parameters.

However, since low frequency components are considered to be reproduced by encoding them using the adaptive level crossing method, it is thought that with the HC used here, it is sufficient to encode only high frequencies. It will be done. Furthermore, human hearing is said to be insensitive to the phase of high frequencies, and there are encoding methods that focus on this property [Moriya, Honda, band encoding,
Theory of Faith (A), J 70. A, S, 445-
452°1987,). Here, we will take advantage of this property,
Transmits frequency and amplitude as parameters without transmitting phase.

After performing the adaptive level crossing method, encoding is performed using tIC, but this method uses the residual difference between the input speech and the partial decoding of the coded by the adaptive level crossing method. Encoding is performed by HC. At that time H.C.
The parameters transmitted when encoding are the frequency and amplitude of the high frequency.

FIG. 1 is a diagram showing the basic configuration of adaptive level encoding in which high frequencies are emphasized by this HC, and FIG. 4 shows the amplitude spectrum of synthesized speech in a voiced section using this encoding method. It can be seen that the amplitude spectrum closer to the original sound shown in FIG. 11 is reproduced compared to the one not using HC shown in FIG. 12. In addition, the original audio (
FIG. 3(a) shows the waveform of synthesized speech using the encoding method according to the present invention (see FIG. 3(b)).

Next, adaptive level encoding that takes pitch into consideration will be explained.

Since voiced sounds exhibit periodic waveforms, when performing adaptive level encoding, by performing adaptive level encoding only on the waveform for the period of time in one frame,
It is possible to lower the bit rate.

The method for extracting the period is to first perform adaptive level encoding at a high level and find the intersection point.

Next, adaptive level encoding is performed on one period of the waveform, and an intersection point is determined. Transmit these parameters to the decoder. The decoder first decodes one period of the waveform, and then decodes one frame so that the waveform matches the intersection point made at a high level.

Furthermore, for frames in which the pitch cannot be successfully extracted, such as in unvoiced sounds or transition parts, the frame is set at a specified time, and adaptive level encoding is performed on the entire frame.

Next, the procedures of the encoder and decoder will be specifically explained.

First, the encoder will be explained with reference to FIG. 2. The encoder's steps are as follows: 1. Frame extraction is performed using a frame length of 256 samples and a frame overlap of 15 samples.

2. Extract the pitch.

If the pitch cannot be extracted, adaptive level encoding is performed on the entire frame.

If the pitch can be extracted, adaptive level encoding is performed on one pitch in the frame.

The presence or absence of pitch extraction, the number of intersections for each, and the interval between intersections are transmitted. When a pitch is extracted, the length of the extracted pitch is also transmitted.

3. Perform partial decoding and take the residual difference from the input voice.

4. Perform FFT on the residual, find the frequency and amplitude of the FFT peak at the high frequency component, and transmit the results.

becomes.

Next, the decoder procedure will be explained with reference to FIG. 3. The decoder procedure is as follows: 1. Decoding is performed using the adaptive level crossing method from the average amplitude, number of intersections, and interval between intersections.

2. Perform HC decoding using frequency and amplitude.

3. The decoded data obtained by adaptive level coding and HC are added together to perform overall decoding.

becomes.

As is clear from the above explanation, according to the present invention, encoding is performed using HC after adaptive level crossing, so the original It is possible to reproduce audio amplitudes close to audio. Also, when performing adaptive level encoding, by performing adaptive level encoding only on the waveform for ■cycles in one frame,
Bitrate can be lowered.

[Brief explanation of drawings]

Figure 1 is a diagram showing the basic configuration of the encoding method according to the present invention, Figure 2 is a flowchart of speech encoding, Figure 3 is a flowchart of decoding, and Figure 4 is a diagram showing the encoding method according to the present invention. Figure 5 is a diagram showing the amplitude spectrum of the voiced sound obtained by the method. Figure 5 is a diagram showing the waveforms of the original voice (,) and voiced voice (b). Figure 6 is for explaining an example of voice encoding. FIG. 7 is a flowchart for explaining an example of decoding, FIG. 8 is a basic configuration diagram for explaining an example of a conventional technique of adaptive level encoding, and FIG. 9 is a flowchart for explaining an example of decoding. Figure 10 is a basic configuration diagram of HC, Figure 11 is a diagram showing the amplitude spectrum of the voiced original voice, and Figure 12 is a diagram for explaining the method when using the adaptive level crossing method. FIG. 3 is a diagram showing an amplitude spectrum of a voiced sound. 10... Encoder, 20... Decoder Figure 1 Figure 2 (a) ORDER DECORDER Figure 2 Figure 6 Figure 8 Figure 8 (a) (b) a P. (C) Figure 1 Frequency) Figure 2 Frequency

Claims (1)

[Claims] 1. In a compression method for a PCM encoded audio signal,
It has means for creating a frame for each predetermined number of samples, and means for calculating positive and negative crossing levels based on the values of each sample in the frame, and calculating the crossing level value for each frame, the number of crossing points, A voice encoding method characterized by encoding a code indicating whether a positive or negative level has been crossed for each intersection, and a time interval between the intersection and the previous intersection. 2. In the method for decoding encoded signals in claim 1, when the intersection intersects with the previous intersection at a different code level and the interval between the two points is larger than a predetermined value, interpolate with a cubic curve, and when the interval is smaller than a predetermined value, interpolate with a straight line between the two points,
When crossing the level with the same code as the intersection, the 2
An audio decoding method characterized by interpolating between points using a quadratic curve. 3. In the compression method for PCM encoded audio signals,
The input PCM audio signal is framed in a predetermined section, the pitch is extracted in that section, and when pitch extraction is possible, the pitch section is used as a new frame and the positive and negative crossing levels are calculated based on the values of each sample in the frame. It has means for calculating, and calculates the value of the intersection level for each frame, the number of intersections, a sign indicating whether the level is positive or negative for each intersection, and the time between the intersection and the previous intersection. When encoding the interval and performing local decoding from these encoding results, if the intersection intersects with the previous intersection at a different code level and the interval is larger than a predetermined value, the interpolate with a cubic curve, and when the interval is smaller than a predetermined value, interpolate with a straight line between the two points,
When crossing the level with the same code as the intersection, the 2
Interpolate between points using a quadratic curve, take the difference between this local decoding result and the original signal, perform Fourier transform on that signal, and calculate the frequency and amplitude of the peak point of the spectrum of each harmonic and the intersection for each frame. The value of the level, the number of intersections, and a sign indicating which level (positive or negative) was crossed for each intersection,
A voice encoding method characterized by encoding the time interval between the intersection and the intersection before it. 4. In the method for decoding encoded audio data set forth in claim 3, the value of the crossing level for each frame, the number of crossing points, and whether positive or negative levels are crossed for each crossing point are indicated. Based on the code and the time interval between the intersection and the previous intersection, if the intersection in the frame intersects at a different code level from the previous intersection, and the interval is larger than a predetermined value, the intersection between the two points is determined. Interpolates with a cubic curve, and when the interval is smaller than a predetermined value, interpolates with a straight line between the two points, and when it intersects a level with the same sign as the intersection, interpolates with a quadratic curve. An audio decoding method is characterized in that the audio signal is decoded by performing inverse Fourier transform from the code indicating the frequency and amplitude of the peak point of the spectrum of each harmonic, and the decoded audio signal is added to perform decoding. .