JP3505364B2 - Method and apparatus for optimizing phoneme information in speech database - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention is applied to waveform synthesis in which optimum phoneme piece information is selected and connected from a voice database composed of phoneme piece information, which has been cut out from text utterances and accumulated in advance, to obtain a synthesized sound. The present invention relates to a phoneme piece information optimizing method in a speech database and a phoneme piece information optimizing device.

[0002]

2. Description of the Related Art Conventionally, a voice file structure for performing clustering on the phoneme piece information of a voice database applied to waveform synthesis for connecting voice waveforms to obtain a synthesized voice and registering optimum phoneme piece information in a voice file. The method is Japanese Patent Laid-Open No. 8
-263520.

FIG. 5 is a flowchart showing a conventional context clustering process. In the figure,
From the phoneme-labeled waveform data in the speech database 100, all the waveform data with the same phoneme label are taken out and set as the initial cluster 110 (step 201).

Next, the individual waveform data (elements) in this initial cluster 110 are subjected to feature analysis (step 20).
2). In this feature analysis, the order of feature parameters such as LPC (linear predictive coding) cepstrum is n, and the frame period of the analysis window function is variable, and the analysis is performed so that the number of frames is m frames. Thus, an n × m-dimensional feature parameter matrix is obtained for each element.

Next, using the result of this feature analysis, the cluster distortion of the initial cluster 110 is obtained (step 20).
3). Specifically, in the vector space of feature parameters, the sum of squares of the distances between all the elements of the initial cluster 110 and the previously obtained centroid is calculated, and this is defined as the cluster distortion of the initial cluster 110. To do.

In this way, the cluster distortion of the initial cluster 110 is obtained and registered in the context cluster table 208. In the context cluster table 208, as shown in the figure, for each cluster, the context belonging to it, its centroid, its cluster distortion, and
A set of element waveforms included in it is registered.

At the stage of obtaining the cluster distortion of the initial cluster 110, only the initial cluster 100 is registered in the context cluster table 208.

Next, the context cluster table 208
A cluster having the largest cluster distortion is obtained from the inside (step 204), the obtained cluster is taken out from the context cluster table 208, and further divided into two clusters according to the context (step 205).

In the initial stage, the initial cluster 110
Since only the initial cluster 110 is registered in the context cluster table 208, cluster division is performed on this initial cluster 110.

In this way, after the initial cluster 110 is divided, the context cluster table 208
In, the initial cluster 110 is deleted, and the two divided clusters are newly registered (step 20).
6).

By repeating the above processing (steps 203 to 206), the initial cluster 110 is gradually subdivided into smaller clusters. Then, the termination determination of the context clustering is performed for each of the repeated loops (step 207).

[0012]

However, even if the speech file (database) is created by reducing the phoneme piece information in the speech database 100, the sentences included in the speech database 100 and the sentences to be input to the speech synthesizer are However, since the appearance frequency of phoneme pieces is different, the problem that the context cluster table 208 still contains a lot of phoneme piece information that is not used at the time of speech synthesis remains.

Therefore, according to the present invention, a large amount of learning text information (sentences) is prepared in advance, which is preliminarily synthesized by a speech synthesizer using all the phoneme pieces, and from the result, each phoneme piece is used. The number of times (frequency information) is obtained, the sum of distances is calculated according to the distribution, and clustering is performed.

As a result, even if the speech database that is the target of the clustering process contains various pieces of phoneme information, a speech file (database that allocates many pieces of speech to frequently used speech). )
It is possible to build.

[0015]

A method for optimizing phoneme piece information in a speech database according to the present invention is applied to a phoneme piece in a speech database, which is applied to waveform synthesis to obtain a synthesized sound by connecting phoneme pieces cut out from a sentence utterance. In the information optimizing method, the text data for learning is synthesized in advance, the frequency of use of each phoneme piece information that constitutes the speech database is obtained according to the synthesis result, and the clustering process is performed based on the frequency of use. The feature is that the phoneme piece information of the audio file is optimized.

The phoneme piece information optimizing method in the speech database of the present invention is a phoneme piece information optimizing method in a speech database which is applied to waveform synthesis for obtaining a synthesized sound by connecting phoneme pieces cut out from a sentence utterance. In the first step, the phoneme selection unit receives the learning text information as an input, and selects the optimum phoneme unit from the speech database in which the phoneme units cut out from the sentence utterance are accumulated, and the selection probability calculation unit uses the phoneme selection unit. A second step of obtaining a selection probability of each phoneme unit selected by the unit, and a clustering processing unit performing a clustering process on the speech database in the predetermined parameter space using the selection probability as a distribution probability of the phoneme unit. The three steps, and the phoneme waveform segment registration unit are classified by the clustering processing unit. Taringu been characterized by a fourth step, selecting a representative phoneme from each cluster.

A phoneme piece information optimizing device in a voice database according to the present invention is a phoneme piece information optimizing device in a voice database applied to waveform synthesis for obtaining a synthesized sound by connecting phoneme pieces cut out from a sentence utterance, A voice database accumulating phoneme pieces cut out from the sentence utterance, a phoneme selecting section for selecting an optimum phoneme piece consisting of the voice database by inputting learning text information, and each selected by the phoneme selecting section, A selection probability calculation unit that obtains a selection probability of a phoneme unit, a clustering processing unit that performs clustering processing on the speech database in the predetermined parameter space using the selection probability as a distribution probability of the phoneme unit, and clustering by the clustering processing unit. Select a representative phoneme from each processed cluster. The phoneme waveform segments registration unit that, characterized in that it comprises a.

Further, the clustering processing unit is characterized by performing clustering processing so that a distance from a centroid in each cluster to all phoneme pieces included in the cluster is minimized.

The text information for learning is characterized by being composed of sentences.

Further, the phoneme piece is characterized by being composed of at least waveform information.

[0021]

BEST MODE FOR CARRYING OUT THE INVENTION An embodiment of the present invention is shown in FIGS.
Will be explained.

FIG. 1 is a schematic configuration diagram of an apparatus for realizing the present invention. FIG. 2 is a flow chart for realizing the phoneme unit information optimizing method in the voice database according to the present invention.

The process of FIG. 2 will be described below with reference to FIG.

In step S1, when the learning text information (sentence) is input to the language processing unit 1, the language processing unit 1
Performs morphological analysis and dependency analysis, and adds a phoneme symbol, a part of speech, and an accent symbol string to the analyzed phoneme.

In step S3, the prosody pattern generator 2
Estimates the fundamental frequency F ₀ , power, and phoneme duration in the vicinity of the phoneme center using the phoneme symbol, the accent symbol string, and the part-of-speech information of the input text obtained from the result of the morpheme analysis.

In step S5, the cost is obtained by using the formula shown in Formula 1 and the phoneme is selected. The specific selection of phonemes in this step is performed by selecting the fundamental frequency F ₀ estimated in step S3, the power, the cost representing the incompatibility with the estimated value of the phoneme duration, and the cost of connecting each phoneme. Is calculated, and a combination of phoneme pieces whose sum total is minimized is searched for by using Equation 1 according to the DP (dynamic programming) method. Here, the cost function D (F) indicating the cost is shown in Equation 1.

[0027]

[Equation 1]

Note that D _F0 , D _pow , and D _dur are the difference between the estimated value of the fundamental frequency, power, and phoneme duration in the vicinity of the phoneme center and the synthesis unit, and D _posi is the position in the sentence (start of word). ,
This is a numerical value of the difference between the word and the ending).
D ^c _F0 , D ^c _pow , and D ^c _cep are the difference in fundamental frequency, the difference in power, and the difference in cepstrum at the connection points of the two combined units to be connected. D ^c _ph is a distortion indicating the ease of connection (connection priority) determined in consideration of the utterance environment. Further, w _x and w ^c _x are weighting factors by which the respective parameters are multiplied.

Next, in step S7, the learning text information (sentence) is synthesized by using the speech database 7 including all the phonemes, and the number of times each phoneme is selected is calculated. Specifically, all phoneme pieces in the speech database 7 are divided into appropriate phoneme units. At this time, when an unvoiced consonant is included, it is divided into CV and VC, and when a voiced consonant is included, it is divided into VCV. In addition, "C" means
Represents a consonant, and "V" is a vowel (V
owel).

Next, the number of times phoneme pieces included in each phoneme unit are selected is averaged to obtain the number of phoneme unit selections, and the selection probability of each phoneme unit is calculated. In the present invention, a small probability is assigned to a phoneme unit that has never been selected. In step S9, the clustering processing unit 5 is executed based on the selection probability calculated in step S7.
Using the LBG algorithm when the distribution is known,
Clustering processing is performed on the specified number.

Here, step S9 will be described in detail with reference to FIG.

First, in step S21, a specified number of initial representative vectors A ₀ are arbitrarily determined. Further, it is assumed that the index m = 0 and the average distortion D- ₁ = -∞.

In step S23, all the phonemes in the speech database are divided into clusters P (A _m ) to which the nearest representative vector A _m belongs. Equation 2 is used to calculate the distance at this time.

[0034]

[Equation 2]

In step S25, the centroids (centroids) of all the phonemes belonging to each cluster are calculated,
Let m + 1 be the representative vector. The centroid is obtained by calculating an average for each element of the vector of acoustic parameters of the phoneme piece, and this average is calculated using the selection probability of each phoneme.

In step S27, the average distortion D _{m + 1} for the representative vector A _{m + 1} and the cluster P (A _{m + 1} ) is calculated. The distortion is a result obtained by averaging the equation 2 with a selection probability.

A step S29 increments the index by 1.

In step S30, the end judgment is made. The reduction rate of the distortion is calculated, and if it is less than a certain amount ε, the cluster P (A _m ) at that time is output and the process ends.

Now, returning to FIG. 2, the description will be continued.

Finally, in step S11, the centroid of each cluster obtained in step S9 is calculated,
By registering the phoneme closest to it as the selected phoneme by the audio file creating unit 6, a reduced (reduced) audio file (database) is newly created.

Next, an evaluation experiment was conducted in order to confirm the effectiveness of the present invention. In this experiment, we aimed to create a synthesizer that reads out place names. All the new postal code data except for the Kyushu region were used as the learning texts. From the new postal code data, the city / county name, ward / town / village name, and town area name were extracted and converted into the sentence "This is X city, Y ward, Z town."

From the result, "here is" and "is."
The selection probability of each phoneme was obtained from the rest of the results except for the part. Using the reduced files (database) clustered using the obtained selection probabilities and the reduced files (database) clustered with uniform selection probabilities, the place names used for learning, the place names not used for learning (Kyushu region), Three types of novel sentences were synthesized and evaluated.

The size of the audio file (database) used in the experiment and the experimental result are shown below.

[0044]

[Table 1]

In FIG. 4, the vertical axis represents 20 place names and novels.
It is the average distortion per phoneme obtained by dividing the total distortion cost of number 1 when 5 sentences are combined by the total number of phonemes in the sentence. Also, FIG.
The diagonal lines in the middle are the synthesis results (conventional) by the speech file (database) created with a uniform selection probability, the intersecting lines are the results by the proposed method (proposed), and the vertical lines are the speech files containing all the phonemes (database). ) Is the result of synthesis (all).

Place-name (closed) indicates a place name used for learning, place-name (open) indicates a place name not used for learning (Kyushu region), and novel indicates a novel result which is a sentence having a completely different environment.

From these results, it can be seen that in all cases, the proposed method has less distortion than the case where the selection probability is uniform, and the proposed method is effective. Comparing each sentence type, the distortion of proposed is similar to all in reading place names, but it is conventiona in reading novels.
close to l. This can be said regardless of open-closed, and it can be seen that optimization to the environment of reading the place name is being done.

[0048]

As is clear from the above description, according to the present invention, phoneme piece information optimization in a speech database applied to waveform synthesis for obtaining a synthesized sound by connecting phoneme pieces cut out from a sentence utterance. In the method, the text data for learning is synthesized in advance, the frequency of use of each piece of phoneme information that constitutes the speech database is obtained according to the synthesis result, and the clustering process is performed based on the frequency of use to determine the frequency of use. It is possible to create an audio file (database) including high phoneme piece information.

Furthermore, the present invention is a phoneme piece information optimizing device in a speech database applied to waveform synthesis for obtaining a synthesized sound by connecting phoneme pieces cut out from a text utterance, and the phoneme pieces cut out from the text utterance. A phoneme selection unit that selects an optimum phoneme piece composed of the phonetic database by inputting a speech database that has stored a learning text and a phoneme selection unit.
A selection probability calculation unit that obtains a selection probability of each phoneme piece, a clustering unit that performs a clustering process on the speech database in the predetermined parameter space using the selection probability as a distribution probability of the phoneme pieces, and the clustering unit. Further, by providing the phoneme waveform element registration unit that selects a representative phoneme element from each cluster, it is possible to allocate a large amount of phoneme element information to a frequently used speech.

[Brief description of drawings]

FIG. 1 is a schematic configuration diagram of an apparatus for realizing the present invention.

FIG. 2 is a flowchart for realizing a phoneme piece information optimizing method in a voice database according to the present invention.

FIG. 3 is a flowchart showing in detail the processing of step S9 shown in FIG.

FIG. 4 is a diagram showing a result of an evaluation experiment of the present invention.

FIG. 5 is a flowchart showing a conventional context clustering process.

[Explanation of symbols]

1-Language processing unit 2 ... Prosody pattern generator 3 ... Phoneme selection section 4 ... Choice probability calculator 5 ... Clustering processing unit 6-Sound file creation section 7: Voice database 8: Reduced audio file (database)

─────────────────────────────────────────────────── ─── Continuation of front page (72) Inventor Makoto Hashimoto 2-5-5 Keihan Hondori, Moriguchi City, Osaka Sanyo Electric Co., Ltd. (56) Reference JP-A-8-263520 (JP, A) JP Hei 7-13598 (JP, A) Yasuhide Hashimoto, Takashi Saito, Speech synthesis using syllables in consideration of environmental dependence, Proceedings of the Japan Acoustics Society Conference, Japan, ASJ, 1995 September, 1995 Autumn I, 245-246 Hiroyuki Hirai, Makoto Hashimoto, Shuji Nishida, Hiroki Onishi, Reduction of speech synthesis speech database based on frequency information of phoneme selection results, ASJ study Proceedings of the presentation, Japan, ASJ, September 1997, Autumn 1997, 259-260 (58) Fields investigated (Int.Cl. ⁷ , DB name) G10L 13/06

Claims (9)

(57) [Claims] 1. A phoneme piece information optimizing method in a speech database which is applied to a waveform synthesis for obtaining a synthesized sound by connecting phoneme pieces cut out from a sentence utterance, which are synthesized in advance using learning text information, A speech database characterized by optimizing the speech segment information of a speech file by obtaining the frequency of use of each piece of speech element information that composes the speech database according to the synthesis result and performing clustering processing based on the frequency of use. Phoneme piece information optimization method. 2. A phoneme piece information optimizing method in a speech database applied to waveform synthesis for obtaining a synthesized sound by connecting phoneme pieces cut out from a sentence utterance, wherein a phoneme selection unit receives learning text information as an input. A first step of selecting an optimum phoneme piece from a speech database in which phoneme pieces cut out from the sentence utterance are accumulated, and a selection probability calculation unit obtaining a selection probability of each phoneme piece selected by the phoneme selection unit And a third step in which the clustering processing unit performs clustering processing on the voice database in the predetermined parameter space using the selection probability as the distribution probability of the phoneme unit, and the phoneme waveform unit registration unit. Fourth selection of representative phoneme pieces from each cluster clustered by the clustering processing unit A method for optimizing phoneme piece information in a speech database, comprising: 3. A phoneme piece information optimizing device in a speech database applied to waveform synthesis for obtaining a synthesized sound by connecting phoneme pieces cut out from a text utterance, wherein a voice accumulating phoneme pieces cut out from the text utterance is accumulated. A database, a phoneme selection unit for inputting learning text information and selecting an optimum phoneme unit consisting of the speech database, and a selection probability calculation unit for obtaining a selection probability of each phoneme unit selected by the phoneme selection unit. A clustering processing unit that performs a clustering process on the speech database in a predetermined parameter space using the selection probability as a phoneme unit distribution probability, and a representative phoneme unit from each cluster that is clustered by the clustering processing unit. And a phoneme waveform segment registration unit for selecting Phoneme information optimizing apparatus in voice database. 4. The clustering processing unit performs clustering processing so that a distance from a centroid in each cluster to all phoneme pieces included in the cluster is minimized. A method for optimizing phoneme piece information in the speech database according to 2. 5. The phoneme piece information optimizing method in a speech database according to claim 1, wherein the learning text information is composed of a sentence. 6. The method for optimizing phoneme piece information in a speech database according to claim 1, wherein the phoneme piece is composed of at least waveform information. 7. The clustering processing unit performs clustering processing such that a distance from a centroid in each cluster to all phoneme pieces included in the cluster is minimized. A phoneme piece information optimization device in a speech database. 8. The phoneme unit information optimizing device in a speech database according to claim 3, wherein the learning text information is composed of a sentence. 9. The phoneme piece information optimizing device in a speech database according to claim 3, wherein the phoneme piece is composed of at least waveform information.