JPH1097290A - Speech synthesizer - Google Patents

Abstract

PROBLEM TO BE SOLVED: To output a synthesized speech waveform of superior speech quality by reading an optimum unit speech waveform corresponding to a 1st vocal sound symbol part string divided in specific preferential order out of a waveform memory and connecting it. SOLUTION: A text speech synthesizer 10 includes a microcomputer 12. The microcomputer 12 receives an input character string consisting of a 1st vocal sound symbol string consisting of text document data, and uses a dictionary 14 for text analysis to convert it into a vocal sound symbol string consisting of the 1st vocal sound symbol part string and also generate the pitch pattern and power pattern of this input character string. Then the microcomputer 12 shapes, connects, and edits unit speech waveforms registered in a speech waveform data base 16 according to the pitch pattern and power pattern, and outputs the resulting synthesized speech. Language information corresponding to vocal sound symbols of a 2nd vocal sound symbol string which is divided in specific preferential order is added to the 2nd vocal sound symbol string.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech synthesizer, and more particularly to a speech synthesizer used for voice guidance, voice response, voice reading, etc., which synthesizes and outputs a voice waveform according to a phoneme symbol string corresponding to an input character string. Related to the device.

[0002]

2. Description of the Related Art A conventional general speech synthesizer is disclosed in Japanese Patent Application Laid-Open No. 7-92997. This corresponds to the phoneme sequence for speech synthesis, and from the speech units included in the database, based on the prosodic information added to the speech unit, a speech unit is selected based on acoustic features as a selection criterion, and Is to connect.

FIG. 6 shows a specific example of the configuration of the above-described speech synthesizer.

In FIG. 6, reference numeral 100 denotes an input terminal, 101
Is a preprocessing unit, 102 is a selection criterion parameter setting unit, 10
Reference numeral 3 denotes a unit selection unit, 104 denotes a condition setting unit, 105 denotes a unit parameter table, 106 denotes a unit file, 107 denotes a unit connection unit, and 108 denotes an output terminal.

The condition setting unit 104 sets various unit environment appropriate conditions to be used in the unit selection process in the unit selection unit 103, and can add, change, and delete these setting conditions.

Next, the processing of the conventional speech synthesizer will be described with reference to FIG.

The preprocessing unit 101 divides an input character string into phoneme units. The selection criterion parameter setting unit 102 sets a selection criterion parameter used for selecting a waveform segment that is a synthesis parameter based on the above-described phoneme unit. The unit selection unit 103 calculates the square error between the set selection criterion parameter and the unit parameter fetched from the unit parameter table 105, and selects the unit parameters in ascending order of the square error. Then, a primary candidate is generated, and a unit corresponding to a unit parameter that best meets the unit environment appropriate condition set in the condition setting unit 104 is determined as an optimum unit for the above-described phoneme unit.

[0008] At this time, the "element environment appropriate condition" is:
(1) The phoneme environment from which the unit was extracted matches the phoneme environment at the time of synthesis or the articulation method is similar, (2) the magnitude relationship of the average pitch matches the magnitude relationship of the selection reference parameter, (3) the pitch The sign of the slope (positive / negative / 0) matches the sign of the selection reference parameter, and the optimal segment is determined based on the prosody information.

Next, the unit connection unit 107 extracts the optimal unit for the determined phoneme unit from the unit file 106, connects the unit for each phoneme unit, and outputs a synthesized speech.

[0010]

However, the vowels at the end of Japanese tend to be unvoiced and have unique acoustic characteristics such as low power as a whole. At the time of unit selection, judgment is made only with prosodic information that can evaluate acoustic features such as the average pitch period and average power of the speech unit. There was a problem that the unit could not be selected.

Each speech unit is decomposed in syllable units, and speech units are selected in accordance with the selection criteria for each syllable unit. The sound quality was not available, which also hindered the improvement of the sound quality of the synthesized sound.

[0012] Therefore, a main object of the present invention is to provide a speech synthesizer capable of outputting a synthesized sound having excellent sound quality.

[0013]

According to the present invention, a synthesized speech waveform is synthesized by synthesizing unit speech waveforms corresponding to a plurality of first phoneme symbol substrings included in a first phoneme symbol string corresponding to an input character string. And a dividing unit that divides the first phoneme symbol string into a plurality of first phoneme symbol subsequences at a predetermined priority, and a second phoneme divided at the predetermined priority. A waveform memory storing a second phoneme symbol string including a symbol subsequence and a speech waveform including a unit speech waveform corresponding to the second phoneme symbol subsequence; and a unit corresponding to the first phoneme symbol subsequence. A waveform reading means for reading a voice waveform from the waveform memory; and a waveform connecting means for connecting a unit voice waveform read from the waveform memory to generate a synthesized voice waveform, wherein the second phoneme symbol string is provided. In
A feature is that linguistic information corresponding to each phoneme symbol is added to each phoneme symbol.

The second phoneme symbol subsequence is added with linguistic information indicating whether or not the phoneme symbol subsequence is the ending, and a second phoneme symbol subsequence that matches the first phoneme symbol subsequence is added. When the unit speech waveform corresponding to the phoneme symbol subsequence is read from the waveform memory, if the first phoneme symbol subsequence is the ending, the corresponding second phoneme symbol subsequence is suffixed based on the language information. Is selected.

The second phoneme symbol subsequence is added with linguistic information indicating whether or not the phoneme symbol subsequence is the ending, and the second phoneme symbol subsequence coincides with the first phoneme symbol subsequence. When reading a unit speech waveform corresponding to a phoneme symbol subsequence from the waveform memory, if the first phoneme symbol subsequence is not an ending, a second phoneme symbol subsequence corresponding to the first phoneme symbol subsequence is determined based on the language information. It is characterized in that a unit voice waveform that is not the ending is selected.

Further, the predetermined priority is a silent part,
It is characterized by an unvoiced part and a voiced part in this order.

[0017]

1 to 5 show an embodiment of the present invention.
This will be described with reference to FIG.

Referring to FIG. 1, text-to-speech synthesizer 1
0 includes the microcomputer 12. The microcomputer 12 receives an input character example composed of a first phoneme symbol string composed of text sentence data, and uses the text analysis dictionary 14 to convert the input character string into a first phoneme symbol substring in which a dividing point is set. And a pitch pattern and a power pattern of the input character string are generated.

At this time, in order to divide the first phoneme symbol sequence into the first phoneme symbol subsequences, it is preferable that a predetermined priority order is set, for example, a silent part, an unvoiced part, and a voiced part.

Next, the microcomputer 12 shapes and connects and edits the unit sound waveform registered in the sound waveform database 16 based on the pitch pattern and the power pattern, and outputs a synthesized sound generated thereby.

The speech waveform database 16 includes "speech waveforms", "phonological label information" for each speech waveform, "prosodic information" representing acoustic characteristics near the waveform connection point, and "suffix or not" indicating whether or not the ending is present. "Language information" is registered. The phoneme label information includes a phoneme symbol string (second phoneme symbol string) and a symbol string number. As a specific example, FIG. 2 lists information registered in the audio waveform database 16. Note that "-" included in the phoneme symbol string indicates a silent section of 5 msec or more.

The phoneme symbol sequence (second phoneme symbol sequence) registered in the speech waveform database 16 has a predetermined priority, for example, a silence portion, an unvoiced sound portion, and the like, like the first phoneme symbol subsequence. It is composed of a second phoneme symbol substring divided in the order of the voiced sound part.

FIG. 3 shows an algorithm for generating a phoneme character string, a power pattern, and a pitch pattern corresponding to an input character string.

First, in step S1, the microcomputer 12 writes an input character string into the memory 12a in units of one sentence. next,
In step S3, a morphological analysis of the character string is performed. That is,
The text analysis dictionary 14 stores information such as word notation and phonological symbol strings (reading), accents, parts of speech, and the like. The input character string is composed of what words using these information. Is analyzed.

Subsequently, in step S5, a phoneme symbol string of the input character string is generated based on the analysis result.

Then, in step S7, the pause (PAUSE) information of the input character string is analyzed using the text analysis dictionary 14, and a power pattern of the input character string is generated from the analysis result in step S9.

Further, in step S11, accent information of the input character string is analyzed using the text analysis dictionary 14, and a pitch pattern of the input character string is generated in step S13 from the analysis result.

Here, the power pattern is a well-known quantified type II model and the pitch pattern is also a well-known Fujisaki model (edited by Shizuo Hiki, "Speech Information Processing" University of Tokyo Press, 1973)
Is calculated by

Next, a phoneme symbol string corresponding to the input character string,
FIG. 4 shows an algorithm for generating an output voice based on the power pattern and the pitch pattern.

At step S15, the microcomputer 12 first determines a division point of the phoneme symbol string corresponding to the input character string, and divides the phoneme symbol string into a plurality of phoneme symbol subsequences.

Next, in step S17, the subsequence number n
Is set to “1”, and the unit speech waveform and the label information corresponding to the n-th phoneme symbol subsequence are extracted from the speech waveform database 16 in step S19.

Subsequently, in step S21, the phoneme duration and gain of the unit speech waveform are corrected by waveform shaping so as to match the power pattern corresponding to the input character string.

Then, in step S23, the pitch of the unit voice waveform is corrected by waveform shaping so as to correspond to the pitch pattern corresponding to the input character string.

Subsequently, the waveforms are connected in step S25,
The connected synthesized speech waveform is stored in the memory 12b in step S27.
To memorize.

Thereafter, in step S29, the sub-sequence number n is incremented, and in step S31, it is determined whether or not the n-th unit voice waveform exists. If "YES" here, the process returns to the step S19. However, if "NO", the synthesized voice waveform is converted into an analog voice waveform and output in a step S33. The data conversion in step S33 is performed by using a well-known P
SOLA method (F. Charpentier et al., “Pitch-Synchronous W
aveform Processing Techniques for Text-to-speech S
synthesis Using Diphones ", Proc. Eurospeech '89).

Here, step S1 which is a feature of the present invention is described.
5 will be specifically described below.

In the embodiment of the present invention, evaluation points determined by the following "evaluation function score" are calculated for all the phoneme symbol subsequences formed from the combination of the division points of the input phoneme symbol string, and each evaluation point is calculated. The division point is determined from the combination that minimizes the accumulation of the evaluation points corresponding to the phoneme symbol subsequence.

Here, the "evaluation function score" is a value type determined by the priority of the division point, a value link determined by the type of phoneme before and after the division point, and a value len determined by the length of the divided phoneme. , And a value f determined by the difference in pitch period from the theoretical value at the waveform connection point corresponding to the division point
0, and w1-w5 for each numerical value of the value term quantifying whether the selected phonological symbol subsequence is the ending or not.
Weighted and added together. The weights of w1 to w5 are real constants of 0 to 10, respectively.

Evaluation function: score = w1 * type + w2 * lin
k + w3 * len + w4 * f0 + w5 * term type = 0 (when the division point is the first priority) type = 1 (when the division point is the second priority) type = 3 (the division point is the priority) Type = 9 (otherwise) link = 0 (if the phonemes before and after the division point match) link = 9 (otherwise) len =-(separated by the division point) F0 = | log (pitch period of actual waveform)-log (theoretical pitch period) | / log (theoretical pitch period) term = 0 (input subsequence is not the ending, but is selected) Term = 1 (when the input subsequence is the ending and the selected subsequence is also the ending) term = 9 (other than the above) Hereinafter, the input character string / -ameno-tamedesu-
Regarding / (because of rain), the method of determining the division point is described.

In this embodiment, for simplicity of description, w1 = 1, w2 = 1, w3 = 1, w4 = 1, w5 = 1. The combination of phoneme symbol substrings is performed by a tree search shown in FIG.

In FIG. 5, the selected phoneme symbol subsequence (this phoneme symbol subsequence exists in the label information of the speech waveform database 16 and all phonemes before and after the phoneme division point coincide with each other) The score value is shown below. For the sake of explanation, the selected state of each phoneme symbol substring is referred to as "node 0" to "node 8" for convenience.

First, at node 0,
(Silence) and / -ameno. . . The phoneme symbol subsequence following "/" is retrieved from the label information of the speech waveform database 16, and a predetermined number m (two in the embodiment) of the phoneme symbol subsequence having the smallest score value is searched. ) Select and create m nodes below.

In FIG. 5, node 1 / -ameno- / and node 4 / -ameno-tam / were selected. /-
The score value of ameno- / ends at the division point of type = 9: out of priority.

Link = 0: Subsequent phonemic symbols match at t.

Len = −10 f0 = 1.2: pitch difference 1.2 times term = 0: the input subsequence is not the ending, and the selected subsequence is not the ending.

The score value of score = 9 + 0-10 + 1.2 + 0 = 0.2 / -ameno-tam / is such that type = 0: ends at the division point of the first priority.

Link = 0: Subsequent phonemic symbols match at m.

Len = -7f0 = 1.3: pitch difference 1.3 times term = 0: the input subsequence is not the ending, and the selected subsequence is not the ending.

Score = 0 + 0−7 + 1.3 + 0 = −5.7 Here, nodes 1 and 4 are set as phoneme division subsequence candidates.

Therefore, the total score at each node
The values are respectively: cumulative score at node 1 = 0.2 cumulative score at node 4 = -5.7. In this embodiment, the phoneme subsequences of node 1 and node 4 remain as candidates in order to leave search sequences of m phoneme subsequences starting from the one with the smallest total score for each division. Therefore, as the next search, nodes 2, 3, 5, and 6 are candidates, and the total score at node 2 = -1.6 The total score at node 3 = -5.2 The total at node 5 score = −4.5 Total score at node 6 is −6.6.

In this case, if there is a tie, priority is given to the smaller score value at that node.
Nodes 3 and 6 remain as candidates. Here, since the division of the node 3 has been completed, the total score at the node 3 always remains as a candidate. No further search from nodes 2 and 5 is performed. The division is repeated in the same manner, and the remaining
From FIG. 5, the nodes are Node 3, Node 7, and Node 8. The cumulative score of each node is: Total score at Node 3 = -5.2 Total score at Node 7 = -2.5 The total score at node 8 is -9.1.

Here, scor at node 7 and node 8
Comparing e, the calculation at node 7 is as follows.

Type = 0: ends at the division point of the first priority.

Link = 0: The end of the sentence is not connected.

Len = -6f0 = 1.1: pitch difference 1.1 times term = 9: the input subsequence is the ending, and the selected subsequence is not the ending.

Score = 0 + 0-6 + 1.1 + 9 = 4.1 At node 8, the calculation is as follows.

Type = 0: Ends at the division point of the first priority.

Link = 0: No connection is made after the end of the sentence.

Len = -6f0 = 1.5: pitch difference 1.5 times term = 1: the input subsequence is the ending, and the selected subsequence is the ending.

Score = 0 + 0-6 + 1.1 + 9 = -2.5, and if there is no term term, a node 7 whose pitch period before and after the connection is well approximated is finally selected, and the synthesized sound ends with / Desu- / is synthesized with the sound in the sentence, resulting in an unnatural sound.

Therefore, the phoneme division by searching up to the node 8 having the smallest score in consideration of the linguistic information as to whether or not it is the end is optimal, and the actual division is as follows: / −ameno − / − name / edsu- / is determined, and the voice waveform at the end of the database 16 is used as the voice waveform of / edesu- /.

[0062]

As is apparent from the above description, according to the present invention, the optimum unit speech waveform corresponding to the first phoneme symbol subsequence divided by the predetermined priority is read from the waveform memory by the reading means. Since it is read out and connected by the waveform connection means, it is possible to output a synthesized voice waveform having excellent sound quality.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a speech synthesizer of the present invention.

FIG. 2 is a diagram showing contents of a speech waveform database.

FIG. 3 is a flowchart showing a part of the operation of the embodiment.

FIG. 4 is a flowchart showing a part of the operation of the embodiment.

FIG. 5 is a flowchart showing a part of the operation of the embodiment.

FIG. 6 is a block diagram showing a conventional speech synthesizer.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 10 ... Text-speech synthesizer 12 ... Microcomputer 14 ... Dictionary for text analysis 16 ... Speech waveform database

 ──────────────────────────────────────────────────続 き Continuing from the front page (72) Inventor Hiroki Onishi 2-5-5 Keihanhondori, Moriguchi-shi, Osaka Sanyo Electric Co., Ltd.

Claims (4)

[Claims] 1. A speech synthesizer that synthesizes unit speech waveforms corresponding to a plurality of first phoneme symbol substrings included in a first phoneme symbol string corresponding to an input character string and outputs a synthesized speech waveform, Dividing means for dividing the first phoneme symbol sequence into a plurality of first phoneme symbol subsequences at a predetermined priority; and a second means including a second phoneme symbol subsequence divided at the predetermined priority. A waveform memory storing a speech waveform including a phoneme symbol sequence and a unit speech waveform corresponding to the second phoneme symbol subsequence; and reading a unit speech waveform corresponding to the first phoneme symbol subsequence from the waveform memory. Waveform reading means, and a waveform connection means for connecting the unit speech waveforms read from the waveform memory to generate a synthesized speech waveform, wherein the second phoneme symbol string includes, for each phoneme symbol, Languages corresponding to phonological symbols Speech synthesis apparatus characterized by multi-address is added. 2. The second phoneme symbol subsequence is added with linguistic information as to whether the phoneme symbol subsequence is an ending or not, and a second phoneme symbol subsequence that matches the first phoneme symbol subsequence is added. When the unit speech waveform corresponding to the phoneme symbol subsequence is read from the waveform memory, if the first phoneme symbol subsequence is the ending, the corresponding second phoneme symbol subsequence is suffixed based on the language information. 2. The speech synthesizer according to claim 1, wherein a unit speech waveform is selected. 3. The second phoneme symbol subsequence includes linguistic information indicating whether the phoneme symbol subsequence is an ending or not, and a second phoneme symbol subsequence that matches the first phoneme symbol subsequence. When reading a unit speech waveform corresponding to a phoneme symbol subsequence from the waveform memory, if the first phoneme symbol subsequence is not an ending, a second phoneme symbol subsequence corresponding to the first phoneme symbol subsequence is determined based on the language information. 2. The speech synthesizer according to claim 1, wherein a unit speech waveform other than the ending is selected. 4. The speech synthesizer according to claim 1, wherein the predetermined priority order is a silent part, an unvoiced part, and a voiced part.