JP2910035B2 - Speech synthesizer - Google Patents

Abstract

PURPOSE:To synthesize the voices of various ways of talkings from phonetic symbols and accent type symbols by constituting a meter pattern estimating section of plural linear signal processing sections of multiple inputs-outputs and providing a weight coefft. memory, data input part and weight multiplying and data adding means thereto. CONSTITUTION:Data xi is inputted to an input section 1001 in the multiinputs- output signal processing section and the weight coefft. wi stored in a memory 1002 is multipled by a multiplier 1003 and is totaled by an adder 1004. Further, a threshold processing section 2000 is provided to limit the output of the linear signal processing section to a specified range. The phonetic symbols and accent type symbols of the inputs are converted to the meter pattern of the output by adequately determining the weight coefft. Many sets of the meter patterns corresponding to the phonetic symbols and accent type symbols are prepd. and the learning is repeated by gradually changing the weight coefft. until the meter patterns attains optimum values. The voices having dialectal characteristics and high power of personal expression are synthesized by this constitution.

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech synthesizer for converting an input character string into speech.

2. Description of the Related Art In recent years, the development of a speech synthesizer of the type of inputting a character string and converting it into speech has been active, and prototypes have been announced in academic conferences and newspaper reports. This type of speech synthesizer basically has a structure as shown in FIG.
A language processing unit 1 for converting a character string into a phonetic symbol and an accent type symbol, a prosody pattern estimating unit 2 for estimating a prosodic pattern of the voice from the phonetic symbol and the accent type symbol, and a voice synthesizing unit for synthesizing the voice from the information. It consists of three. In some simple devices, the language processing unit 1 may be omitted and phonetic symbols and accent symbols may be directly input.

A conventional method for realizing the speech synthesizer having such a configuration will be described below.

The method of realizing the language processing unit 1 is basically the same as the “Kana-Kanji conversion technology” of the Japanese word processor. That is,
The kana-kanji conversion technique is to perform a morphological analysis on an input kana character string and to search for a kanji code corresponding to a reading code from an independent word dictionary from an independent word dictionary to obtain a kanji-kana mixed sentence. On the other hand, in the processing in the language processing unit 1, the kana character string is morphologically analyzed and converted into phonetic symbols, and for an independent word, an accent type symbol corresponding to the reading is searched from the independent word dictionary instead of the kanji code, The pronunciation symbol + accent type symbol is used. If the input to the language processor 1 is a kanji-kana-mixed sentence, contrary to the method of using the independent word dictionary in the case of kana-kanji conversion, the independent word dictionary is searched using the kanji code and the reading code is extracted, and then the three words are read. Will be converted to phonetic symbols and accent symbols. For example, "I am" →
"I am" → "WATASIWA + flat type" is converted. The accent type symbol is, for example, NH published by the Japan Broadcasting Publishing Association.
For example, the accents of "chopsticks" and "bridges" are distinguished by "head-high type" and "Odaka type".

The prosody pattern estimation unit 2 outputs a prosody pattern representing the naturalness of speech, such as a pitch pattern, a formant pattern, a duration of a phoneme, and a dynamic pattern of a sound, from the phonetic symbols + accented symbols obtained by the language processing unit 1. . Conventional methods for estimating prosodic patterns are described in, for example, Fujisaki and Sudo, "Basic frequency patterns of Japanese word accents and their generation mechanism models."
Applied to the numerical model shown in Journal of the Acoustical Society of Japan, Vol. 27, No. 9, 1979, Higuchi and Yamamoto, "Control of prosodic features in a rule synthesis experiment system" Proceedings of the Spring Meeting of the Acoustical Society of Japan in 1986. It is obtained by inferring the empirically obtained rules shown in 2-2-14. The speech synthesizer 3 is described in, for example, Yamamoto et al., "Prototype Production of Speech Rule Synthesizer Using Phoneme as Synthesis Unit", Proceedings of the Acoustical Society of Japan, Spring Meeting, 1987, 3-6-2. That is, the acoustic parameters are determined based on the formant frequency prepared for each phonetic symbol from the phonetic symbol obtained by the language processing unit 1 and the pitch frequency, phoneme duration, and strength obtained from the prosodic pattern estimation unit 2. And drive the formant synthesizer.

However, in a configuration having a prosody pattern estimating unit based on the above method, it is very difficult to express dialects and individuality. Even if the pitch pattern is a numerical model, it can be obtained unless specialists take a very long time to analyze the actual voice data and analysis in order to recreate a model corresponding to the Osaka dialect or a model corresponding to the Hakata dialect. Absent.
Even more, if you want to get rules based on experience, it takes a lot of time to get enough rules. The same applies to estimating the prosody pattern of Mr. A's speech style and Mr. B's speech style. Therefore, in reality, there is a problem that only one type of speech can be synthesized according to one representative type of prosodic pattern.

In view of the above, an object of the present invention is to provide a speech synthesizer that synthesizes various dialects and personality voices by facilitating estimation of a prosody pattern.

Means for Solving the Problems The present invention is a speech synthesizing apparatus including a prosody pattern estimating unit in which a plurality of multi-input one-output signal processing units are configured in a circuit network.

Operation According to the present invention, the prosody pattern estimating unit outputs various prosody patterns from phonetic symbols and accent-type symbols, thereby synthesizing various spoken voices.

Embodiment FIG. 2 shows a general configuration diagram of a phoneme pattern estimating unit according to the present invention. FIG. 3 shows a specific configuration example. The operation of estimating a pitch pattern as a prosody pattern will be described below with reference to FIG.

FIG. 3 shows an example of a prosody pattern estimating unit 2 having a three-stage configuration for estimating a pitch pattern for 7 moras using a phonetic symbol and an accent-type symbol as inputs. The signal is propagated. The phonetic symbols and accent symbols used as input are divided into 13 inputs for each mora. 1
When a plurality of phonetic symbols are assigned to one input unit 101, they are distinguished by the magnitude of the numerical value. For example, 0 is assigned if not applicable, -1 is assigned other than the number of input mora, and (0-1) is assigned based on the number of phonetic symbols accepted by the unit. For example, if a mora is / nya /,
The second unit is 1/3 for repetitive sounds, the third unit is 6/8 for / n /, the eighth unit is 1 for vowels, and the remaining 10 units are 0 for 1 mora of input data. create. The upper layer unit 100 is a multi-input / one-output signal processing unit for obtaining inputs from all the lower layer units. As shown in FIG. 3, the input layer 10 is composed of 91 units of the input layer 101 and the intermediate layer 102. Is composed of 87 units, and the output layer 103 is composed of 84 units. Then, 91 units 101-u of the input layer 101 are combined with 87 units 102-u of the intermediate layer 102, and 87 units 102-u of the intermediate layer 102 are further combined with 84 units 103 of the output layer 103.
-U.

Each connection is multiplied by a weight obtained by learning in advance, and the output layer 103 obtains pitch values of 6 points for consonants and 6 points for vowels per 1 mora. Therefore, the output unit 103-u has 12 units per mora,
An output of 84 units is obtained per unit / mora × 7 moras.

FIG. 4 specifically shows the configuration of the linear signal processing unit based on only the linear operation in the multi-input one-output signal processing unit 100 constituting the prosody pattern estimating unit 2. In FIG. 4, reference numeral 101 denotes a multi-input one-output signal processing unit 100.
1002 is a memory for storing a weighting factor for weighting a plurality of inputs from the input unit 1001, 1003 is a multiplier for multiplying the weighting factor of the memory 1002 by the input from the input unit 1001, and 1004 is a multiplier 1003 This is an adder for adding the outputs of. That is, the multi-input one-output signal processing unit shown in FIG.
100 is calculated as y = xiwixi (1), where xi is the input value to the input unit 1001 and wi is the weight coefficient stored in the memory 1002. FIG. 5 specifically shows a configuration of a non-linear signal processing unit that also performs a non-linear operation in the multi-input / one-output signal processing unit 100 constituting the prosody pattern estimating unit 2. In FIG. 5, reference numeral 1000 denotes a linear signal processing unit described in FIG. 4, and reference numeral 2000 denotes a threshold processing unit for limiting the output of the linear signal processing unit to a value within a certain range. FIG. 6 shows an example of input / output characteristics of the threshold processing unit 2000. For example, the input / output characteristics of the threshold processing unit 2000 for limiting the output to the range of (0, 1) can be mathematically expressed as 0 = 1 / (1 + exp (-I)). Here, I and O are threshold processing units 20
00 input and output.

The prosody pattern estimation unit 2 having the above configuration outputs a pitch pattern to be finally obtained. Other prosodic patterns can be estimated in exactly the same way.

Next, a description will be given of why a prosody pattern can be estimated with the above configuration, and how a prosody pattern can be accurately estimated with the above configuration. The prosody pattern estimation unit 2 has a multistage circuit network configuration. The relationship between the input and output of the prosody pattern estimator 2 is stored in memory
It only depends on the weighting factor stored in 1002. Naturally, there is a strong correlation between the phonetic symbols / accented symbols of the input and the prosodic pattern of the output, so if this weighting factor can be determined appropriately, the phonetic symbols / accented symbols of the input will be converted to the prosody pattern of the output. It is possible to do. This is the reason that the prosody pattern can be estimated.

The second problem "how to accurately estimate the prosodic pattern" can be embodied into the problem "how to determine an appropriate weighting factor". This problem can be solved by gradually changing an arbitrary weighting coefficient and repeating learning from phonetic symbols / accented symbols until the rhythm pattern reaches an optimal value.
Such learning algorithms include, for example, backpropag
ation (DERumelhart, GEHint on and RJWilliams "
Learning Representations by Back-Propagating Erro
rs, "Nature, vol. 323, pp. 533-536, Oct. 9, 1986). Mathematical proofs are given in the references, but phonetic symbols / accented symbols are used as training data, and other means are used. A large number of prosodic patterns corresponding to the estimated phonetic symbols and accent type symbols are prepared as a set, and the input and output are used as input and output, and the input / output relationship is repeatedly learned by a backpropagation algorithm.

The results obtained experimentally are shown. In the experiment, the prosody pattern estimating unit 2 has three input layers, intermediate layers, and output layers as shown in FIG.
We learned to construct a circuit network consisting of layers and estimate the pitch pattern. However, the multi-input one-output signal processing units 100 in each layer were 84 and 87 non-linear signal processing units, respectively. Therefore, the weight coefficient stored in the memory 2 is (84
X 87) + (87 x 91). FIG. 7 shows the inputs and outputs of the prosodic pattern estimation unit 2. In FIG. 7, the dot points are the actually measured pitches, and the lines are the outputs of the prosodic pattern estimation unit 2. Just by inputting phonetic symbols and accent symbols, pitch patterns that are very close to actual measurements can be estimated. That is,
This shows that the speaker's dialect and personality can be expressed very well.

As described above, according to the present embodiment, by providing the prosodic pattern estimating unit including at least a plurality of multi-input / one-output signal processing units connected to the network, the phonetic symbols /
The corresponding prosodic pattern can be estimated from the accent marks.

In the prosody pattern estimating unit 2 in the embodiment, the multi-input / one-output signal processing unit in the upper layer is connected to all the units in the lower layer. However, it is not essential that all the units are connected. Any combination may be used.

Further, although the number of weighting factors stored in the memory 2 in the embodiment is the number of combinations of the number of units, 1 is always input to the multi-input / one-output signal processing unit 100 with a weight. May be. In this case, the number of weighting factors stored in the memory 2 increases by the unit. This input of 1 always transforms equation (1) into y = w0 + Σwixi (2). In other words, the expression ability is further increased by eliminating the restriction that the formula (1) always passes through the origin. That is, the estimation ability of the prosody pattern estimation unit 2 can be further improved.

Effect of the Invention As described above, according to the present invention, a prosody pattern can be easily estimated by configuring a prosody pattern estimation unit in a signal processing network including a multi-input one-output signal processing unit. For this reason, it is possible to synthesize speech that is rich in dialect and personality, so that its practical value is great.

[Brief description of the drawings]

FIG. 1 is a block diagram of a speech synthesizer according to one embodiment of the present invention, FIGS. 2 and 3 are block diagrams of a prosody pattern estimating unit in the device, and FIG. 4 is a configuration of a linear signal processing unit in the device. FIG. 5, FIG. 5 is a configuration diagram of a non-linear signal processing unit in the device, FIG. 6 is an input / output characteristic diagram of a threshold processing unit in the device, and FIG. 7 is a pitch pattern estimated by a prosody pattern estimating unit in the device. It is an example figure. DESCRIPTION OF SYMBOLS 1 ... Language processing part, 2 ... Prosody pattern estimation part, 3 ... Speech synthesis part, 100 ... Multi-input one-output signal processing part, 101 ... Input part of prosody pattern estimation part, 1000 ... Linear signal processing Part, 10
01: Input section of the multi-input / one-output signal processing section, 1002: Memory, 1003: Multiplier, 1004: Adder, 2000: Threshold processing section.

──────────────────────────────────────────────────続 き Continuation of the front page (56) References JP-A-63-46498 (JP, A) JP-A-61-259295 (JP, A) J. Sejnowski, C.R. Resenberg, "Parallel Networks That Learn to Prompt English Text", Comp lex, System, 1 (1987) p. 145-168 Hideyuki Takagi, Prosody Control Using Proto-Era Neural Networks (1988) 2008 Spring Meeting Presentation Papers-▲ I ▼-(58) Fields surveyed (Int.Cl. ⁶ , DB name) G10L 3/00 G10L 5/02 G10L 5/04 G06F 15/18 560

Claims (1)

(57) [Claims] 1. A language processing unit (1) for outputting phonetic symbols and accent types for each mora from an input character string, and an output from the language processing unit (1) as an input to estimate and output a pitch pattern. A pitch pattern estimating unit (2), and an output from the language processing unit (1) and an output from the pitch pattern estimating unit (2) as inputs, and a voice synthesizing unit (3) that synthesizes a voice signal.
And a pitch pattern estimating unit (2) comprising at least an input layer (10
1) a signal processing network including an intermediate layer (102) and an output layer (103). The input layer (101) includes a plurality of input layer units (101-u), and receives a signal from the language processing unit (1). Is input to the input layer unit (101-u), and the intermediate layer (102) includes a plurality of intermediate layer units (102-u). , Weighting processing, addition processing, and threshold processing for limiting the numerical value output from the input layer unit (101-u) to a certain range value are performed, and the output unit (103
-U), the output layer (103) includes a plurality of output layer units (103-u), and receives a value output from the intermediate layer unit (102-u) as an input and forms a pitch pattern forming a pitch pattern. A speech synthesizer for outputting a value.