JPS593497A - Fundamental frequency control system for rule synthesization system - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION [Technical Field of the Invention] The present invention relates to a voice synthesis system +tic using a regular synthesis method, and is capable of obtaining sufficient performance through simple control.

[Prior Art of the Invention J There are two types of output methods that can be broadly classified into audio output devices. 1
One is to keep all the words or sentences to be output in a side window.
It outputs them sequentially in a desired combination,
When there are a large number of types of sentences to be output, there is a drawback that a huge amount of memory capacity is required.7 The output method that is the subject of this application for addressing sluggishness parameterizes the phonological units of natural speech by only the necessary types. This is a rule synthesis method that synthesizes arbitrary words or sentences from those phonological units.In this case, in addition to arranging each phonological unit, it also controls the pitch and amplitude of each phonological unit to create more natural speech. can do. 1 [Problems with the prior art] In order to obtain a smooth time-varying pattern of the fundamental frequency close to that of human speech by using the conventional method, it is necessary to It was necessary to set the fundamental frequency using suitable means.

Furthermore, the calculation process of these fundamental frequencies is performed using the logarithm value of the general frequency. This is because, in terms of human hearing, a better approximation can be obtained by operating according to the logarithm of the frequency. However, so-called speech synthesis r, 5xFi, which creates audio signals according to this fundamental frequency information, all require input of the fundamental period (reciprocal of the frequency), and therefore it is necessary to convert the logarithm of the frequency into a period. The sixth problem is that it requires a complicated circuit and conversion processing time.

Furthermore, when connecting two phonetic units, if the final amplitude of the first phonetic unit and the initial torsion of the second phonetic unit are different, it is necessary to interpolate between them to make it clear. This amplitude information company is generally given as a parameter in a format in which logarithmic representation is further compressed, but in order to interpolate this, it must be expressed as a single integer, then interpolated, and then returned to the original parameter format and input to the speech synthesis LSI. (Objective of the Invention) The present invention solves these problems, enables high-speed processing with a simple circuit, and has a sufficient processing time. The purpose of this invention is to provide a speech synthesis control method that can obtain excellent performance.

[Structure of the invention]

In order to solve the above drawbacks, the present invention first provides information for controlling the fundamental frequency as a rough approximation, and then smoothes that information (digital information) using a digital filter.
The attempt is made to obtain a smoothness close to that obtained by giving a fine approximation.

Furthermore, a tape lv index method is employed to convert this fundamental frequency information into parameters based on pitch periods, thereby achieving hardware simplification and ease of changing characteristics.

In addition, in the interpolation processing of amplitude information, the torsion width parameter is treated as a pure binary number and interpolated without converting it to an integer value, thereby achieving high-speed and high-performance interpolation processing. I have to.

[Embodiments of the invention]

Figure 1 is an explanatory diagram of the rule synthesis method.
This example uses the case of composing the words . There are several ways to form phonological units, but here we use a combination of so-called VCV (vowel, consonant, vowel).

[YAMAGATAJ is composed of five phonological units “YAJ.

1'-AMAJ, [AGAJ, [ATAJ,
[Obtained by connecting AJ. ((a) in the same figure) Since each unit can be connected using the same vowel, a natural connection can be obtained relatively easily.

FIG. 6B shows fundamental frequency information (logarithmic expression) and shows that there is an accent on "C" and "M".

FIG. 6(c) shows a pitch (periodic) representation of the fundamental frequency, where the logarithmic representation is a straight line, but the period is a curved line.

FIG. 2 is a block diagram of an embodiment of the present invention, in which 1 is a means for inputting words/sentences to be synthesized in character codes; 2 is a means for converting character strings into sequence information of phoneme units (VCV); 3 is a means for disassembling and arranging the sequence of phoneme units into V and C sequence information, 4 is a means for specifying accent and intonation patterns, 5 is a prosody table that provides parameters according to intonation, etc., and 6 is a fundamental frequency 7 is a digital filter, 8 is a conversion table, and 9 is a means for creating a time-series pattern of the connection time between phonemes; 9, each phoneme unit (VCV) is converted into a buff meter and stored at a sampling period of 10 m5ec, for example; file,
10 is a means for combining vC■ parameters, 11 is a means for synthesizing an audio signal according to the parameters, and 12 is a speaker.

When indexing the phonetic tape/L15 using intonation information etc. to obtain fundamental frequency information (log f),
Conventionally, in order to obtain a sufficiently smooth logf curve, the information in the phoneme table 5 must be prepared in detail at a period comparable to the above-mentioned sampling period, or the information on tape/%15 must be made coarse and replaced by a sample. A relatively complex predetermined function was used to interpolate between the two.

On the other hand, in the present invention, the information on the tape/L15 is coarsely divided, for example, every 100111161j, and the interpolation is performed using a simple function such as linear interpolation. Instead, the information is smoothed in digital fill 7.

FIG. 3 shows an embodiment of the digital filter 7, where the input X is the output Y delayed (Z"') by the subtracter 31.
The difference between X'□I is taken and multiplied by a (1
>a>O)L, and YX] are added by an adder 33 and output. The output Y is one sampling period (for example, 10m
mee) Delayed by delay element 34 and subtraction @? ;Feedback to 3X+adder 33. The transfer characteristic of this circuit is Y a , which is a pole when Z = 1-a, and when mapped to the S plane, it becomes Z = e”πfT, where f is the polar frequency and T is the sampling period. be.

Now, assuming a-1/4 and T == 0.01 seconds,
4.6 Hz in Ha f-J n (1a).

2πT i.e. cutoff frequency 4.6Hz, -6db10ct
low-pass filter tonal.

If such a filter is used, even a straight line approximation like the solid line in Figure 1(b) can be achieved with a smooth log f like the dotted line in the same figure.
It becomes information.

Furthermore, a conversion table 8 is used to convert the logarithmic expression information of the fundamental frequency obtained in this way into period information. FIG. 4 shows an example of a table in which a ROM (read only memory) is used, and the input is given, for example, an 8-bit binary number (0 to 255) + o, whereas the output is given as a 7-bit binary number. (111-21) 10 is output in binary.

According to such a table conversion method, it is possible to convert the device into a device with arbitrary characteristics simply by converting or switching the ROM. For example, changing from a male voice to a female voice can be easily performed. .

Further, amplitude interpolation between each phoneme unit is performed as follows.

In addition, each phoneme unit is 10 male in VCV file 9.
It is stored as a time-series set of various parameters for each time. Amplitude information is also one of the parameters. Figure 1 (a
) The amplitudes of the same vowel rAJ are generally different, such as the end of [YAJ and the beginning of [AMAJ], and when connecting them, it is necessary to perform interpolation to connect them smoothly. The amplitude W information among the parameters is generally expressed as shown in the following table.

That is, an IO integer is expressed as a floating point binary number, and since the first digit below the decimal point is always "1" (because it is normalized), it is omitted to compress the number of bits.

For example, the final amplitude puff meter of a certain phoneme is "10001".
"Then, the first amplitude parameter of the next phoneme is 10101"
In the case of
'', linear interpolation is performed between them, and the parameter format is returned again.

On the other hand, in the present invention, the parameter format is changed directly to pure 2
It is treated as a base number and linear interpolation is performed directly. In other words, in the example in item 1, the difference between ``10001'' and ``101 old'' is 0010.
0". If linear interpolation is performed between four points, it is sufficient to add a difference of "00001" to each point.

Therefore, as my ex-husband said, the parameter format is "10001"
From 10101", it is interpolated in steps of "00001". When converted to decimal, it becomes 610".
12', '14'', '16', '20'', which is an exponential interpolation.This method not only eliminates unnecessary conversion, but also has excellent auditory characteristics. It has the effect of being there. This is because human hearing has the characteristic that even if frequency, amplitude, etc. change logarithmically, it only feels like a linear change, and logarithmic interpolation is better when interpolating. -ing

An example of an interpolation circuit is shown in FIG. 5, and this circuit may be considered to be within the VCV coupling unit 10.

The final amplitude parameter of the previous phoneme unit is stored in the register 41, and the first amplitude puff meter of the next phoneme unit is stored in the V dimeter 42.
, and the subtracter 43 calculates the difference by treating each as a pure binary number.

On the other hand, the difference is divided by the connection time (number of samples) from the means 6 for determining the connection time between phonetic units by the divider 44,
The quotient may be added to the previous final amplitude value by the adder 45.

〔Effect of the invention〕

As described above, according to the present invention, by performing simple processing with a simple circuit, it is possible to obtain characteristics that are equivalent to or better than conventional ones, and are effective in reducing the cost of the audio output device.

[Brief explanation of the drawing]

Fig. 1 is an explanatory diagram of the rule synthesis method, Fig. 2 is a schematic block diagram of an embodiment of the present invention, Fig. 3 is a block diagram of an embodiment of the digital filter, and Fig. 4 is an illustration of the implementation of a conversion table. Example Glock diagram, Figure 5 is the 1st diagram 2 times

Claims (1)

[Claims] A plurality of phoneme units are sampled at a predetermined period, at least frequency information or amplitude information is converted into a normalized binary floating point representation, and stored in a format with the first digit below the decimal point deleted, to obtain the desired phoneme unit. In the rule synthesis method that connects and synthesizes speech, when interpolating information in the above format, linear interpolation is performed by regarding the exponent part and decimal point part of the above format with the first digit deleted as the upper digit of a pure binary number. Features an interpolation control method in the rule synthesis method.