JP2648138B2 - How to compress audio patterns - Google Patents

Description

Description: BACKGROUND OF THE INVENTION The present invention relates to audio processing, and in particular to the compression of audio patterns and the synthesis of audio patterns from such compressed patterns.

It is known that a bandwidth of at least 4 kHz is necessary so that an audio signal can be properly understood. In a digital speech processing system such as a speech synthesizer recognizer or encoder, the channel capacity required for digital storage of the entire waveform in the 4 kHz band is extremely large. Numerous approaches have been devised to reduce the number of digital codes required to represent a speech signal. Pulse code modulation (PCM), differential pulse code modulation (DPCM),
Waveform signals such as delta modulation or adaptive predictive coding have made it possible to obtain natural, high quality speech at bit frequencies of 16-64 kbps. However, the quality of the speech obtained by the waveform encoder is 16 kbps.
It will deteriorate as it goes down.

Another speech coding method disclosed in U.S. Pat. No. 3,642,302 uses a small number of slowly varying parameters, for example 12-16, which are processed to form a transcript of the speech pattern with less distortion. Such parameters, such as linear prediction coefficients (LPC) or log area parameters generated by linear prediction analysis, can be band-limited to 50 Hz without significant distortion due to band-limiting. Encoding LPC or log area parameters (log cross section parameters) generally requires a sampling frequency twice the bandwidth, and it is necessary to quantize each resulting frame of log area parameters. Each frame of the log area parameter can be quantized using 48 bits. Thus, 12 log area parameters, each with a 50 Hz bandwidth, would have a bit frequency of 4800 bits / sec.

If the band compression is further performed, the bit frequency decreases, but as a result, the distortion increases, and the intelligibility of the speech synthesized from the low band parameters is impaired. It is known that voices do not occur at a fixed rate in a voice pattern, and a method has been devised in consideration of such non-uniform occurrences. In U.S. Pat. No. 4,349,700, dynamic programming can be used to recognize speech with a wide range of speech patterns. U.S. Pat. No. 4,038,503 discloses a method of non-linearly changing the time width of an audio pattern in order to more uniformly express audio features. However, these configurations require that audio feature signals sampled at frequencies corresponding to the most rapidly changing features of the pattern be stored and processed. It is an object of the present invention to provide a speech expression and / or speech synthesis device with reduced digital processing and storage requirements.

SUMMARY OF THE INVENTION Sounds or events in the human voice occur on average at a rate that varies between 10 and 20 events per second. Such a speech event ("speech event" means one sound (for example, one vowel) in a sound sequence) occurs with a non-uniformly distributed time width, and various It has been observed that the movements of the vocal tract for sounds vary widely. Thus, significant compression can be achieved by converting speech feature parameters into units associated with short speech events located at non-uniform time intervals. By encoding such a voice event unit with high efficiency, high efficiency can be obtained without deteriorating the precision of pattern expression.

The present invention is directed to an audio compression device that analyzes an audio pattern and generates a set of signals at a first frequency that represents an acoustic feature of the audio pattern. In response to the speech feature signal, a sequence of signals representing the acoustic features of successive speech events in the speech pattern is generated, and digitally encoded digitally encoded corresponding to each speech event signal. The generated signal is generated at a lower frequency than the first frequency.

According to one aspect of the invention, the audio pattern stores a predetermined set of audio component signals and combines the audio component signals to form a signal representative of the acoustic features of the audio pattern, The voice pattern is synthesized by generating the voice pattern in response to the set of.
The predetermined audio component signal is formed by analyzing an audio pattern at a first frequency to generate a set of audio feature signals. A sequence of signals representing acoustic features of successive speech events in the speech pattern is generated in response to the sampled speech feature signal, and wherein a digital signal corresponding to a speech event representation signal at a frequency lower than the first frequency. A sequence of dynamically encoded signals is formed.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a flowchart illustrating the general method of the present invention; FIG. 2 is a block diagram of a speech pattern encoding circuit illustrating the present invention; FIGS. 3-8 are the circuits of FIG. FIG. 9 is a diagram of a voice synthesizer as an example of the present invention; FIG. 10 is a flowchart showing the operation of the circuit of FIG. 9; FIG. 11 is a circuit of FIG. FIG. 12 is a waveform illustrating a voice pattern and a voice event feature signal associated therewith.

General Description It is well known to those skilled in the art to represent speech patterns as a sequence of speech feature signals derived from linear prediction or other spectral analysis. Log area parameters sampled at close intervals to obtain efficient representations of speech patterns have been used in speech synthesis. According to the invention, the log area parameters are converted into a sequence of individual audio event feature signals φ _k (n) as follows.

The audio event feature signal φ _k (n) is sequential and occurs at an audio event frequency that is substantially lower than the frame frequency of the log area parameter. In equation (1), P is the total number of log area parameters y _i (n) determined by linear prediction analysis. m corresponds to the number of audio events in the pattern, n is the index of the sample in the audio pattern at the sampling frequency of the log area parameter, and φ _k (n) is k at the sample time n
A _ik is the combination factor corresponding to the contribution of the k th audio event function to the ith log area parameter. Equation (1) is expressed as Y = AΦ (2) in the form of a matrix. Where Y is (i, n) element y _i (n)
A is a P × N matrix, and A has its (i, k) element a _ik
Is a P × m matrix, and Φ is a matrix whose (k, n) element is φ
It is an m × N matrix that is _k (n). Since each speech event k represents only a small segment of the speech pattern, the signal φ _k (n) representing it is non-zero only for a small range of the sample interval of the entire pattern. Each log area parameter y _i (n) of equation (1) is a linear combination of sound events function phi _{k (n),} the bandwidth of each y _i (n) parameter sound event function phi _k of _(n) The largest band of any thing. Therefore, if y _i (n) is directly encoded, the encoding of φ _k (n) switch event signal and the equation (1)
It is easy to see that more bits are required than the combination coefficient signal.

FIG. 1 is a flowchart illustrating the general method of the present invention. According to the invention, the speech pattern is analyzed to form a sequence of signals representing the acoustic feature signal of the log area parameter. However, LPC, partial correlation (PARCO
R) or other audio features (eg, US Patent 3,624,3
02) can be used instead of the log area parameter. This feature signal is then converted to a set of speech events representing a signal encoded at a lower bit frequency for transmission or storage.

Referring to FIG. 1, at block 101, an electrical signal corresponding to a voice pattern is low-pass-filtered to remove unnecessary high-frequency voice and noise components, and the wave signal is filtered at a cutoff frequency of a filter. Sampled at twice the frequency. The samples of the audio pattern are then converted at block 110 to a sequence of digitally encoded signals corresponding to the pattern. Because the storage required for the sampled signal is too large for many practical applications, this is used in block 120 to generate the log area parameter signal, using linear prediction techniques well known to those skilled in the art. You. The log area parameter signal y _i (n) is generated at a constant sampling frequency sufficient to accurately represent the fastest expected event in a speech pattern. Typically, a sampling interval of 2 to 5 milliseconds is selected.

After the log area parameter signal is stored, the point in time of the occurrence of a continuous audio event in the pattern is detected, and block 130 generates a signal indicating the timing of the event. This is done by dividing the pattern into predetermined short segments, for example at intervals of 0.25 seconds. In each of the consecutive time with the start frame n _b and the end frame n _e, segment log area parameters
A matrix of log area parameter signals is formed corresponding to y _i (n). Matrix redundancy is reduced by factoring out the first four principal components as follows.

The main components of the first four terms are M.M. in Bell System Technical Journal, Vol. 54, No. 10, p. 169-1723 (December 1975). R. "An Efficie vocoder (" An Efficie ")
nt Linear Prediction Vocoder "). The resulting function of u _m (n) is such that φ _k (n) is best compressed in time. By selecting the coefficient b _km in a linear manner, the desired speech event signals are linearly combined so as to be defined.

In this way, the speech pattern is represented by a series of continuous component (small spread) speech event signals φ _k (n), each of which can be efficiently encoded.
Distance measures to get shape and location of audio event signals Is minimized to select the optimal φ (n), the position of which is Obtained from From Equations 5, 6, and 7, the speech event signal φ _k (n) having the smallest spread is centered at the negative zero crossing of ν (L).

After generating a [nu (L) signal at block 130, enters the block 140, [nu the audio event occurrence signal from the zero crossing in the negative (L), using the process of block 130 connexion sound event signal phi _{k (n} ) Is determined exactly. After generating the sequence of speech event representation signals, the combination coefficients _aik are generated in equations (1) and (2) by minimizing the mean square error.

Here, M is the total number of audio events within the range of the index n where the addition is performed. The partial derivative of the coefficient a _ik of E
In, the coefficients a _ik are obtained from the following linear simultaneous equations:

DETAILED DESCRIPTION FIG. 2 shows an electromagnetic transducer 201, a filter and a sampler 20 operable to convert a speech pattern into a stored sequence of digital codes representing the pattern.
3, the analog digital converter 205 and the audio sample store 210 are illustrated. The central processing unit 275 is a Motorola controlled by instructions permanently stored in read only memories (ROMs) 215, 220, 225, 230 and 235.
Includes microprocessors such as the MC68000.
Processor 275 is an arithmetic processor 280 and stores 210, 240,
It directs the operation of 245, 250, 255, and 260 to compress it into a compact set of speech event feature signals. Next, the voice event feature signal is input / output interface 26
It is supplied to the utilization device 285 through 5. The utilization device is a digital communication facility, or a storage device for delayed transmission or a storage device associated with a speech synthesizer. For Motorola MC68000 integrated circuits Motorola 1980
As described in the annual MC68000 16-bit Microprocessor Use Manual, the arithmetic processor is TRW's M
Consists of a PY-16HJ integrated circuit.

Referring to FIG. 2, the voice pattern is an electroacoustic transducer.
The resulting electrical signal is provided to 201 and supplied to a low-pass filter / sampler circuit 203, which operates to limit the upper limit of the signal band to 3.5kHz and sample the signal oscillated at a frequency of 8kHz. I do. The analog-to-digital converter 205 converts the sampled signal from the filter / sampler 203 into a sequence of digital codes, each of which represents the size of a signal sample. The resulting digital codes are sequentially stored in the audio sample store 210.

After storing the sampled voice pattern code in store 210, central processor 275 causes the instructions stored in log area parameter program store 215 to be transferred to a random access memory associated with the central processor. The flowchart of FIG. 3 illustrates the sequence of operations performed by the controller in response to instructions from store 215.

Referring to FIG. 3, when the block 305 is first entered, the frame count index n is reset to 1. Next the audio sample of the current frame is block 310
Is transferred from the store 210 to the arithmetic processor via the central processor 275. The occurrence of the audio sample signal end is checked in decision block 315. Control is passed to block 325 until the end of the voice pattern signal is detected, and the processors 275 and 280 perform LPC analysis on the frame. The LPC parameter signal of the current frame is then the log area parameter signal y at block 330.
_i (k) and the log area parameter signal is stored in the log area parameter store 240 (block 33).
Five). The frame count is incremented by one at block 345 and the audio sample of the next frame is read. (Block 310). When an audio pattern end signal occurs, control is provided to block 320, and a signal corresponding to the number of frames in the pattern is stored in processor 275.

After the log area parameter storage operation is completed,
The central processor 275 operates to transfer the instructions stored from the ROM 220 to the random access memory. The instruction code from store 220 is illustrated in the flowcharts of FIGS. These instruction codes are then used to generate a signal ν (L) from which the occurrence of a sound event of the sound pattern is detected and the position is detected.

Referring to FIG. 4, the frame count of the log area parameter is initially reset at block 403 by the processor 275, and the log area parameter y _i (n) of the initial time width n ₁ to n ₂ of the voice pattern is log. Transferred from area parameter store 240 to processor 275 (block 41)
0). After it is determined in decision block 415 whether the end of the voice pattern has been reached, block 420 is entered and factors of the first four main terms u _i (n), i = 1,. This eliminates the redundancy of the log area parameter signal.

The log area parameter at the current time width is And a set of signals from now on Is obtained. The u _i (n) signal over that time span is obtained by using the parameters b _i i = 1,. The u _i (n) signal over this time span is generated by making φ _{k the} most compact in the range of n ₁ to n ₂ . This is θ in equation (6).
This is realized by using the function of (L). A signal v (L) representing the voice event timing of the voice pattern is then formed at block 430 according to equation (7),
The (L) signal is stored in the timing parameter store 245. The frame counter n is incremented by a constant value, eg, 5 (block 435), depending on how close adjacent audio event signals φ _k (n) are expected to occur.
Blocks 410 are re-entered to generate φ _k (n) and ν (L) signals in the next time span of the voice pattern.

When the end of the voice pattern is detected at decision block 415, the frame count of the voice pattern is stored (block 440) and the generation of the voice event timing parameter signal for that voice pattern is completed. FIG. 11 illustrates a voice event timing parameter signal in the case of the occurrence of a message shown as an example. Each zero crossing in the negative direction of FIG. 11 corresponds to the center of the speech event feature signal φ _k (n).

Referring to FIG. 5, upon entering block 501, the voice event index I is reset to zero and the frame index n is reset to one. After the indexes I and n have been initialized, successive frames of the audio event timing parameter signal are read from the store 245 (block 5).
05), the zero crossing therein is detected by the processor 275 at block 510. Whenever a zero crossing is found, the voice event index I is incremented (block 51).
5) The voice event location frame is stored in the voice event location store 250 (block 520). Next, block 525 frame index n is incremented and the end of the voice pattern frame is checked at block 530. Block 530 returns to block 505 until the voice pattern frame end signal is detected, and each successive voice event location frame of the pattern is detected.

Upon detection of the audio pattern end signal at block 530, the central processor 235 operates to address the audio event feature signal generation program store 225 and transfer its contents to the processor. Central processor 275 and arithmetic processor 280 are thereby adapted to generate a sequence of voice event feature signals in response to the log area parameter signal in store 240 and the voice event position signal in store 250. The voice event feature signal generation command is illustrated in the flowchart of FIG.

Initially, the position index is set to 1 by block 601, and the location of the voice event in store 250 is transferred to central processor 275 (block 605). Block
Similar to 610, a limit frame of a predetermined number of audio event locations, eg, limited to five, is determined. The log area parameter of the duration of the audio pattern defined by the limit frame is entered in the memory section of the central processor 275 (block 615). Redundancy of the log area parameter is reduced by factoring out the number of key terms therein corresponding to a predetermined number of events (block 620). Immediately thereafter, a speech event feature signal φ _L (n) for the current position L is generated.

The minimization of equation (6) for determining φ _L (n) is performed by the following derivative.

here It is. m is a predetermined number of events, and r is one of 1, 2,... or m. Determining the minimum of the derivative of equation (13) at 0 gives Is obtained. From equation (14) It is. Therefore, equation (15) is Is converted to Φ (n) in equation (17) can be replaced by the right side of equation (14), Becomes here It is. By transforming equation (18) Get. u _i (n) is the main component of matrix Y Becomes Equation (20) is simplified as follows.

here It is. Equation (22) becomes R = ▲ ▼ (24) in a matrix expression. Here, λ = θ (L) (25). Equation (25) has exactly m solutions, and the solution that minimizes θ (L) is the solution that minimizes λ. λ = θ (L)
The coefficients b ₁ , b ₂ ,..., B _m that take the minimum value realize the optimal speech event feature signal φ _L (n).

In FIG. 6, a speech event feature signal φ _L (n) is generated at block 625 and stored in store 255. The loop including blocks 605, 610, 615, 620, 625 and 630 is repeated until decision block 635 detects the end of the audio pattern, forming a complete sequence of audio events for that audio pattern.

FIG. 12 illustrates waveforms illustrating audio patterns and audio event feature signals generated therefrom in accordance with the present invention. Waveform 1201 corresponds to the voice pattern part, and waveform 1205
-1 to 1205-n correspond to the sequence of the speech event feature signal φ _L (n) obtained from the waveform of the circuit of FIG. Each feature signal represents an acoustic feature of the audio event in the pattern of waveform 1201. The voice event feature signal is _{calculated by} the coefficient a _{ik of the} equation (1).
To reproduce a log area parameter signal representing the acoustic features of the audio pattern.

Upon completion of the operation shown in FIG. 6, the sequence of audio event feature signals for that audio pattern is stored in store 255. Each audio event feature signal φ _I (n) is encoded and used as shown in the flowchart of FIG.
Is forwarded to The central processor is adapted to receive a set of audio event signal encoding program instructions stored in ROM 235.

Referring to FIG. 7, the speech event index I is reset to 1 by block 701 and the speech event feature signal .phi.
_I (n) is read from store 255. Sampling frequency R _I of the current speech event feature signal Yotsute one of many ways known to those skilled in the art, it is selected in the block 710. For example, instruction code performs a Fourier analysis and then to determine the sampling frequency R _I, to generate a signal corresponding to the upper limit of the band of the characteristic signal. As is well known to those skilled in the art, the sampling frequency need only be such as to adequately represent the feature signal. Thus, a slowly changing feature signal may utilize a lower sampling frequency than a rapidly changing feature signal, and each feature signal may have a different sampling frequency.

For sound event feature signal φ _{I (n),} the sampling frequency signal is determined once, which is encoded in the frequency R _I in block 715. Any of the well-known coding schemes can be used. For example, each sample is
PCM, ADPCM or can modulate the any of the delta modulated signal, a signal representative of the characteristic signal position in the speech patterns, are joined together with the signal representative of the sampling frequency R _I. The encoded audio event feature signal is then transferred to the utilization device 285 through the input / output interface 265. Next, the speech event index I is incremented (block 720) and a decision block 725 is entered to determine if the last speech event signal has been encoded. Blocks 705-725 are repeated (I>) until the last audio event signal is encoded.
_IF ), when the last encoding is performed, the encoding of the speech event feature signal is completed.

In order to form a copy of the log area parameter signal from the audio event feature signal, the audio event feature signal must be combined according to equation (1). Accordingly, a combination coefficient of the voice pattern is generated and encoded as shown in the flowchart of FIG. After encoding the audio event feature signal, the central processor 275 reads the contents of the ROM 225. Instruction codes permanently stored in ROM control the formation and encoding of the combination coefficients.

This combination coefficient is generated for the entire speech pattern by the matrix processing of the central processor 275 and the arithmetic processor 280. Referring to FIG. 8, the log area parameter of the voice pattern is transferred to the processor 275 by the block 801. The speech event coefficient matrix G is generated by the following equation (block 805).

The correlation matrix c of Y-Φ is (Block 810). Next, a combination coefficient matrix is generated at block 815 according to the following equation.

A = G ⁻¹ C (28) The elements of the matrix A are the combination coefficients _aik of the equation (1). These combination coefficients are encoded in block 820 as is well known to those skilled in the art, and the encoded coefficients are transferred to utilization unit 285.

According to the present invention, the linear prediction parameters sampled at the frequency corresponding to its maximum rate of change are converted to a sequence of audio event feature signals encoded at a much lower audio event occurrence frequency, and the audio pattern is further Compressed
Reduce transmission and storage requirements without negatively affecting intelligibility. Utilization device 285 may be a communication facility connected to one of a number of speech synthesis circuits using LPC all-pole filters well known to those skilled in the art.

The circuit of FIG. 2 utilizes the spoken message using the device 285.
To a sequence of encoded speech event feature signals transmitted to the synthesizer. In the synthesizer, the combined coefficients of the speech event feature signal and the message are decoded and combined to form a message log area parameter signal. These log area parameter signals are then used to generate a copy of the original message.

FIG. 9 is a block diagram of a voice synthesizing circuit according to an embodiment of the present invention; FIG. 10 is a flowchart showing the operation thereof.
The store 915 of FIG. 9 is adapted to store the continuously encoded speech event feature signal and combination signal received from the utilization device 285 of FIG. 2 via the line 901 and the interface circuit 904. Store 920 receives the required sequence of excitation signals by combining via line 903. The excitation signal may consist of a voiced / unvoiced signal generated in response to a voice message in a manner well known to those skilled in the art and a series of pitch periods. The microprocessor 910 controls the operation of the synthesizer, and is the same as the Motorola type described above.
An MC68000 integrated circuit may be used. The LPC feature signal store 925 is used to store a continuous log area parameter signal of the spoken message, which comprises a combination of the speech event feature signal and the store 915. The formation of the uttered message transcript is the LPC feature signal from Store 925 and Store 92
In response to the excitation signal from zero, it is executed by the LPC synthesizer 930 under the control of the microprocessor 910.

The operation of the synthesizer is directed by the microprocessor 910 under the control of a permanently stored instruction code resident in its associated read-only memory. The operation of the synthesizer is described with reference to the flowchart of FIG. Referring to FIG. 10, the encoded speech event feature signal and the corresponding combination signal and the uttered message excitation signal are received by the interface 904, and the speech event feature signal and the combination coefficient signal store 915 are received. It is forwarded by block 1010 to the excitation signal store 920. The log area parameter index I is then reset to 1 in processor 910 (block 1020), which initiates the reproduction of the _first log area characteristic signal y ₁ (n).

In order to form a log area signal, it is necessary to combine the voice event feature signal with the combination coefficient of index I according to equation (1). The voice event feature signal position counter L is set to 1 in processor 910 by block 1025,
The current audio event feature signal sample is read from store 915 (block 1030). The sequence of signal samples is waved and the speech event feature signal is smoothed by block 1035, and block 1040 partially forms the current log area parameter signal. Voice event position counter L is incremented to address the next voice event feature signal in store 915 (block 1045), and the occurrence of the last feature signal is tested by decision block 1050. The loop including blocks 1030-1050 is repeated until the last voice event feature signal has been processed, thereby generating the current log area parameter signal and storing it in LPC feature signal store 925 under the control of processor 910. It is memorized.

Once the log area feature signal has been stored in the store 925, block 1050 enters block 1055, where the log area index signal is incremented to begin forming the next log area parameter signal (block 1055). In the loop from block 1030 to block 1050, the judgment block 10
You will enter through 60. After the last log area parameter signal is stored, the processor 910 causes a copy of the spoken message to be formed in the LPC synthesizer 930.

The synthesizer circuit of FIG. 9 stores a sequence of speech event feature signals corresponding to a plurality of spoken messages, and selectively generates copies of these messages in a manner well known to those skilled in the art. Can be easily changed. In such a device, the voice event feature signal generation circuit of FIG. 2 receives a predetermined vocal message sequence and uses
5 can be formed by a device for permanently storing the speech event feature signal and the corresponding combination coefficient of the message and a read-only memory containing the speech event and the combination signal of the said uttered message. The read-only memory containing the encoded speech event and the combination signal is incorporated as a store 915 in the synthesis circuit of FIG.

Continuation of the front page (56) References JP-A-51-93105 (JP, A) JP-B-53-26761 (JP, B2) Proceedings of the Acoustical Society of Japan Fall Meeting, 1981, I-1-1 -1 A trial of real-time data compression using the repetitiveness of voice format

Claims (6)

(57) [Claims] An audio pattern is analyzed (for example, 210, 2
15, 275, 280), a method for compressing a speech pattern comprising generating a set of acoustic feature signals representing said speech pattern at a first frequency, comprising: Generating an event time sequence of a signal representing the time of occurrence of the voice event by detecting the time of occurrence of a negative zero crossing in the signal representing the timing of the voice event in the voice pattern (eg, 220); Based on the event time sequence of the signal, the acoustic feature signals are linearly combined to generate a sequence of audio event signals, each representing a voice event of the audio pattern (eg, 225, 24).
0) sampling and encoding the audio event signal to generate a sequence of audio event signals encoded at a frequency lower than the first frequency and corresponding to an occurrence frequency of the audio event. (For example, 235, 250), a method of compressing a voice pattern, which includes each step. 2. The method of claim 1, wherein each audio event signal has a time spread such that its center is the time of occurrence of the corresponding audio event. A method of compressing an audio pattern, characterized in that the audio pattern is a linear combination controlled by the following. 3. A method for compressing a speech pattern according to claim 2, wherein said step of generating said encoded speech event signal comprises generating a signal representing a band of each speech representation signal.
5) A method of compressing an audio pattern, comprising: sampling the audio event signal at a frequency corresponding to the signal representing the band. 4. The method according to claim 1, 2 or 3,
A method for compressing a speech pattern according to any of the preceding claims, wherein the acoustic feature signal is a linear prediction parameter signal representing the speech pattern. 5. A method for compressing an audio pattern according to claim 4, wherein said linear prediction parameter signal is a log area parameter signal representing the audio pattern. . 6. A method for compressing a speech pattern according to claim 4, wherein said linear prediction parameter signal is a PA representing a speech pattern.
A method for compressing voice patterns characterized by being RCOR signals.