JP6840124B2 - Language processor, language processor and language processing method - Google Patents

Description

The present invention relates to a language processor, a language processor and a language processing method, and more particularly to, for example, a language processor, a language processing program and a language processing method that generate and output synthetic speech according to input text.

An example of the background technique is disclosed in Patent Document 1. According to Patent Document 1, a speech synthesizer that learns a deep neural network acoustic model from speech data and generates a synthetic speech using the learned deep neural network acoustic model is disclosed. In this speech synthesizer, a deep neural network acoustic model that inputs a language feature vector that expresses context data as a numerical vector and a speaker code, and outputs speech parameters corresponding to the speaker and context data. Is learned.

Further, another example of the background technique is disclosed in Patent Document 2. According to Patent Document 2, the time-length DNN (deep neural network) and the acoustic feature DNN are pre-learned from the speech corpus, and the speech waveform corresponding to the text is synthesized using the learned time-length DNN and the acoustic feature DNN. The voice synthesizer to be used is disclosed. In this speech synthesizer, the pre-learning unit generates phoneme language features, phoneme frame language features, phoneme time lengths, and phoneme frame acoustic features from the speech corpus, and assigns speaker labels and emotion labels. To do. Then, the pre-learning unit learns the time length DNN by giving the phoneme language features, speaker label, emotion label, and phoneme time length, and learns the phoneme frame language features, speaker label, emotion label, and phoneme frame. The acoustic feature amount DNN is learned by giving the acoustic feature amount of.

JP-A-2017-032839 JP-A-2018-146803

H.Zen et al, IEICE Trans.Inf. & Syst., Vol.E90-D, no.5, pp.825-834, May 2007 Zhizheng Wu et al, ISCA SSW9, vol PS2-13, pp.218-223, Sep 2016

In Patent Document 1 and Patent Document 2, nothing is disclosed in the normalization of language feature quantities, but in Non-Patent Document 1 and Non-Patent Document 2 referred to in these Patent Documents, from all the learning data. Normalization with calculated mean and variance or minimum and maximum values is used. However, in these normalization methods, in a speech synthesizer in which free text is input, outliers are generated because values that are not learned are included in the language features. Furthermore, since the extrapolation ability of the neural network is not sufficient, there is a problem that the prediction becomes unstable. As a countermeasure for this problem, generally, a countermeasure is taken to increase the learning data and widen the coverage. However, this measure cannot cover all input patterns. In addition, the cost required to collect a large amount of learning data becomes high.

Therefore, a main object of the present invention is to provide a novel language processing device, language processing program and language processing method.

Another object of the present invention is to provide a language processing apparatus, a language processing program, and a language processing method capable of preventing outliers from occurring.

The first invention is a language processing device that is input to a deep neural network of a speech synthesizer that generates synthetic speech and normalizes a language feature amount vector series composed of a plurality of different attributes for one utterance. It is a language processing device provided with a normalization means for normalizing the first attribute in the language feature quantity vector series with a second attribute different from the first attribute in the language feature quantity vector series for the one utterance. ..

The second invention is subordinate to the first invention, where the first and second attributes are linguistically relevant values.

The third invention is subordinate to the first or second invention, and the normalization means normalizes by dividing the first attribute by the second attribute.

The fourth invention is subordinate to any of the first to third inventions, and the absolute value of the first attribute is equal to or less than the absolute value of the second attribute.

A fifth invention is a language processing program that is input to the deep neural network of a speech synthesizer that generates synthetic speech and is executed by a language processor that normalizes a language feature vector series composed of a plurality of different attributes. Therefore, in the processor of the language processing device, the first attribute in the language feature quantity vector series for one utterance is set to a second attribute different from the first attribute in the language feature quantity vector series for the one utterance. A language processing program that executes a normalization step to normalize.

The sixth invention is a language processing method of a language processing device that is input to a deep neural network of a speech synthesizer that generates synthetic speech and normalizes a language feature quantity vector series composed of a plurality of different attributes. the processor of the language processing unit, the first attribute in the language feature vector sequence of 1 utterance, the process normalizes a different second attribute with the first attribute in the language feature vector sequence of the 1 utterances It is a language processing method to execute.

According to the present invention, the first attribute in the language feature vector series for one utterance is normalized by a second attribute different from the first attribute, so that outliers are prevented from occurring. Can be done.

The above-mentioned objectives, other objectives, features and advantages of the present invention will become more apparent from the detailed description of the following examples made with reference to the drawings.

FIG. 1 is a functional block diagram showing an example of the voice synthesizer of this embodiment. FIG. 2 is a diagram for explaining the pre-learning unit shown in FIG. FIG. 3 is a diagram for explaining the text analysis unit shown in FIG. FIG. 4 is a diagram for explaining the acoustic analysis unit shown in FIG. FIG. 5 is a diagram for explaining the synthesis processing unit shown in FIG. FIG. 6 is for explaining the relationship between the continuous length DNN and the acoustic features, the language features of the phonemes, which are the input data and the output data of the DNN, the language features of the phoneme frame, the continuation length of the phonemes, and the acoustic features of the phoneme frame. It is a figure. FIG. 7 is a diagram showing an example of language feature data. FIG. 8 is a flow chart showing an attribute value calculation process of a CPU built in the speech synthesizer. FIG. 9 is a flow chart showing a first attribute value calculation process of the CPU built in the speech synthesizer. FIG. 10 is a flow chart showing a second attribute value calculation process of the CPU built in the speech synthesizer.

FIG. 1 is a functional block diagram of the voice synthesizer 10 of this embodiment. The speech synthesizer 10 is a general-purpose personal computer or workstation, and also functions as a language processing device or a normalization processing device that normalizes language features, which will be described later. Although not shown, the speech synthesizer 10 includes components such as a CPU, a memory (HDD, ROM, RAM), and a communication device (network connection device).

Hereinafter, the speech synthesizer 10 will be described, but the CPU of the speech synthesizer 10 processes the processes related to learning and the generation of the synthesized speech according to various programs. Further, the storage units 12 and 16 are stored in the memory (HDD and / and RAM) of the speech synthesizer 10 or the memory built in the computer on the network accessible to the speech synthesizer 10 or on the accessible network. Means a database.

As shown in FIG. 1, the speech synthesizer 10 includes a storage unit 12, a pre-learning unit 14, a storage unit 16, and a synthesis processing unit 18. The storage unit 12 stores the voice corpus. A voice corpus is information about a voice in which a particular sentence is read aloud by one or more speakers. In this embodiment, the information about speech is text and speech waveforms. However, the voice waveform of the text and the voice read aloud thereof is stored in the storage unit 12 as a pair (associating with each other).

The pre-learning unit 14 learns the continuous length DNN by performing a predetermined text analysis on the text of the voice corpus read from the storage unit 12 and performing a predetermined acoustic analysis on the voice waveform for the text. Generates information such as language features and acoustic features for learning DNN. However, DNN means a deep neural network. The pre-learning unit 14 learns the continuous length DNN and the acoustic feature amount DNN stored in the storage unit 16 in advance by using information such as the language feature amount and the acoustic feature amount.

Since the text analysis method and the acoustic analysis method are known, detailed description thereof will be omitted in this embodiment. Further, in this embodiment, the continuous length DNN and the acoustic feature amount DNN are stored in the same storage unit 16, but they may be stored in different storage units.

In this embodiment, the continuous length DNN and the acoustic feature DNN are feedforward networks in which a plurality of nodes are composed of an input layer, a plurality of hidden layers (intermediate layers), and an output layer, respectively.

In the continuous length DNN, the linguistic features of phonemes are given to each unit of the input layer at the time of learning, and the continuous length of phonemes is given to the units of the output layer, so that each unit of the input layer, the hidden layer and the output layer Weights and the like are calculated, and phoneme unit learning is performed. In this embodiment, the phoneme linguistic features for learning include, for example, phoneme labels, mora information, accent phrase information, exhalation paragraph information, utterance information, and the like. However, the duration of a phoneme is represented by the number of phoneme frames that make up the phoneme. In this embodiment, the length of one frame of the phoneme frame is 5 msec.

When the speech synthesis process described later is executed, the language features of phonemes are given to each unit of the input layer of the continuous length DNN. Then, the continuous length of the phoneme corresponding to the language feature amount of the phoneme given to the input layer is output from the unit of the output layer of the continuous length DNN.

Further, in the acoustic feature DNN, the language feature of the phoneme frame is given to each unit of the input layer at the time of learning, and the acoustic feature of the phoneme frame is given to each unit of the output layer, so that the input layer and the hidden layer And the weight of each unit of the output layer is calculated, and the phoneme frame unit learning is performed. In this embodiment, the acoustic features of the phoneme frame include, for example, information such as a spectral coefficient, a noise coefficient, a pitch, and a voiced / unvoiced determination.

When the speech synthesis process described later is executed, the language feature amount of the phoneme frame is given to each unit of the input layer of the acoustic feature amount DNN. Then, each unit of the output layer of the acoustic feature DNN outputs the acoustic feature of the phoneme frame corresponding to the language feature of the phoneme frame given to the input device.

FIG. 2 is a diagram for explaining the pre-learning unit 14 shown in FIG. As shown in FIG. 2, the pre-learning unit 14 includes a text analysis unit 14a and an acoustic analysis unit 14b. As shown in FIG. 3, the text analysis unit 14a includes a text analysis means 140, a frame processing means 142, and a normalization processing means 144.

The text analysis means 140 performs text analysis such as morphological analysis on the text read from the speech corpus of the storage unit 12, generates a phoneme language feature for each phoneme, and frame-processes the phoneme language feature. In addition to outputting to the means 142 and the normalization processing means 144, the phoneme label included in the linguistic feature amount of the phoneme is output to the phoneme analysis unit 14b.

Here, the linguistic features of phonemes mean the information generated by the text analysis. For example, the linguistic feature quantity of a phoneme generated by text analysis includes various information such as a phoneme label, an accent position, part of speech information, an accent phrase information, an exhalation paragraph information, and a total number information for each phoneme. However, the phoneme label is information (phoneme information) for specifying the phonemes constituting the text, and in addition to the phonemes, the preceding and following phonemes are also included.

In the frame processing means 142, the language feature amount of the phoneme for pre-learning is input from the text analysis means 140, and the continuation length of the phoneme is input from the acoustic analysis unit 14b. The frame processing means 142 generates the language features of the phoneme frames for the number of phoneme frames indicated by the phoneme continuation length based on the phoneme language features and the phoneme continuation length for pre-learning. The linguistic features of the generated phoneme frame are output to the normalization processing means 144.

The normalization processing means 144 normalizes each of the language features of the phoneme and the language features of the phoneme frame, outputs the normalized phoneme language features to the continuous length DNN, and normalizes the phoneme frame. The language feature amount of is output to the acoustic feature amount DNN.

The normalization process in the normalization processing means 144 will be described in detail later.

As shown in FIG. 4, the acoustic analysis unit 14b includes a phoneme dividing processing means 150 and an acoustic analysis means 152. The phoneme delimiter processing means 150 inputs a phoneme label from the text analysis unit 14a, and performs acoustic analysis on the voice waveform read from the voice corpus of the storage unit 12 using predetermined learning data. The phoneme dividing processing means 150 specifies the position of the phoneme indicated by the phoneme label in the voice waveform, and obtains the phoneme dividing position. The obtained phoneme delimiter position is output to the acoustic analysis means 152.

Further, the phoneme dividing processing means 150 obtains the continuous length of the phoneme indicated by the phoneme label based on the phoneme dividing position. As described above, the duration of a phoneme is represented by the number of phoneme frames that make up the phoneme. The obtained phoneme continuation length is output to each unit of the output layer in the continuation length DNN of the storage unit 16 and is also output to the text analysis unit 14a (frame processing means 142).

The acoustic analysis means 152 receives the phoneme delimiter position from the phoneme delimiter processing means 150, performs acoustic analysis on the voice waveform read from the voice corpus of the storage unit 12, and performs acoustic analysis on the phoneme frame of a plurality of phoneme frames. For each, the acoustic features of the phoneme frame are generated. For example, the acoustic feature amount of the phoneme frame includes information such as a spectral coefficient, a noise coefficient, a pitch, and a voice / unvoiced determination. The generated acoustic feature amount of the phoneme frame is output to each unit of the output layer in the acoustic feature amount DNN of the storage unit 16.

Since the method of obtaining the phoneme dividing position and the phoneme continuation length by acoustic analysis and generating the acoustic feature amount of the phoneme frame is known, detailed description thereof will be omitted in this embodiment.

As described above, the text analysis unit 14a outputs the linguistic features of the phonemes for pre-learning to the input layer of the continuous length DNN, and the acoustic analysis unit 14b outputs the continuous length of the phonemes to the output layer of the continuous length DNN. Output to. As a result, the continuous length DNN is pre-learned. Further, the text analysis unit 14a outputs the language feature amount of the phoneme frame to the input layer of the acoustic feature amount DNN, and the acoustic analysis unit 14b outputs the acoustic feature amount of the phoneme frame to the output layer of the acoustic feature amount DNN. .. As a result, the acoustic feature DNN is pre-learned.

FIG. 5 is a diagram showing an example of a specific configuration of the synthesis processing unit 18 shown in FIG. As shown in FIG. 5, the synthesis processing unit 18 includes a text analysis unit 180, a continuous length generation unit 182, an acoustic feature amount generation unit 184, and a voice waveform synthesis unit 186.

The text analysis unit 180 performs the same processing as the text analysis unit 14a shown in FIG. Specifically, the text analysis unit 180 inputs a text in free sentences, performs text analysis on the text, generates a phoneme language feature for each phoneme, and normalizes the text. The text analysis unit 180 is similar to the phoneme language features for pre-learning generated by the text analysis unit 14a shown in FIG. 2 based on the phoneme language features generated and normalized by the text analysis. Generates phoneme linguistic features. Then, the text analysis unit 180 outputs the linguistic feature amount of the generated phoneme to the continuous length generation unit 182 and the acoustic feature amount generation unit 184.

Further, the text analysis unit 180 has a phoneme continuation length corresponding to the language feature amount of the phoneme output from the continuation length generation unit 182 and the acoustic feature amount generation unit 184 to the continuation length generation unit 182 and the acoustic feature amount generation unit 184. Is input, and based on the phoneme language feature amount and the phoneme continuation length, the language feature amount of the phoneme frame corresponding to the phoneme frame number indicated by the phoneme continuation length is generated. Then, the text analysis unit 180 outputs the language feature amount of the phoneme frame to the continuous length generation unit 182 and the acoustic feature amount generation unit 184.

The continuation length generation unit 182 receives the linguistic features of the phonemes from the text analysis unit 180, and uses the continuation length DNN of the storage unit 16 to generate the continuation length of the phonemes based on the linguistic features of the phonemes. Then, the continuous length generation unit 182 outputs the continuous length of the phoneme to the text analysis unit 180. Further, the acoustic feature amount generation unit 184 receives the language feature amount of the phoneme frame from the text analysis unit 180, and uses the acoustic feature amount DNN of the storage unit 16 based on the language feature amount of the phoneme frame. The acoustic feature amount is generated, and the acoustic feature amount of the phoneme frame is output to the voice waveform synthesis unit 186.

The voice waveform synthesis unit 186 receives the acoustic feature amount of the phoneme frame from the sound feature amount generation unit 184, synthesizes the voice waveform based on the phoneme feature amount of the phoneme frame, and outputs the synthesized voice waveform. Specifically, the voice waveform synthesis unit 186 generates a vocal cord sound source waveform based on information such as pitch and noise characteristics included in the acoustic features of the phoneme frame. Then, the voice waveform synthesis unit 186 performs vocal tract filter processing on the vocal cord sound source waveform based on information such as a spectral coefficient included in the acoustic feature amount of the phoneme frame, and synthesizes the voice waveform. That is, synthetic speech corresponding to the text is generated and output.

Since the method of synthesizing the voice waveform based on the acoustic features of the phoneme frame is well known, detailed description thereof will be omitted in this embodiment.

FIG. 6 is for explaining the relationship between the phoneme language features, the phoneme frame language features, the phoneme continuation length, and the phoneme frame acoustic features, which are the input data and output data of the continuation length DNN and the acoustic features DNN. It is a figure of.

As shown in FIG. 6, when the text for one utterance is "that is this and which is it.", The exhalation paragraph is "that is this" and "that is which". Also, in this case, the accent phrases of "that is this" are "that is" and "this is". Further, in this case, the mora of "that" is "a", "re", and "ga". In this case, the phoneme label of "a" is "a", the phoneme label of "re" is "r" and "e", and the phoneme label of "ga" is "g" and "a". .. In the example shown in FIG. 6, the phoneme continuation lengths of the phoneme labels "a", "r", "e", "g" and "a" are "6", "3", "5" and "a", respectively. Let it be "5", "5" and "3". In this way, utterances, exhalation paragraphs, accent phrases, mora, and phonemes have a hierarchical structure, and information related to these attributes is also a hierarchical structure. As described above, in this embodiment, the length of one frame of the phoneme frame is 5 msec.

As shown in FIG. 6, in the time interval of the phoneme label “a”, the language features of one set of phonemes (each of the above information) are generated corresponding to this one phoneme, and the languages of the six phoneme frames are generated. The feature amount (each information) is generated, and the acoustic feature amount (each information) of 6 sets of phoneme frames is generated. Further, in the time interval of the phoneme label "r", the linguistic features of one set of phonemes are generated corresponding to this one phoneme, the linguistic features of three sets of phoneme frames are generated, and three sets of phoneme frames are generated. Acoustic features are generated. In addition, in the time interval of the phoneme label "e", the linguistic features of one set of phonemes are generated corresponding to this one phoneme, the linguistic features of five sets of phoneme frames are generated, and five sets of phonemes are generated. The acoustic features of the frame are generated.

Thus, in the pre-learning, each unit of the input layer of the continuation length DNN is given the linguistic feature of the phoneme, and the unit of the output layer is given the continuation length of the phoneme, and this pre-learning gives the phoneme. It is done as a unit. That is, the continuation length DNN is given the language feature amount of the phoneme and the continuation length of the phoneme for each phoneme, and pre-learning is performed. Further, in speech synthesis, the continuation length of a phoneme is generated and output based on the linguistic features of the phoneme by using the continuation length DNN for each phoneme.

Further, as described above, in the pre-learning, each unit of the input layer of the acoustic feature DNN is given the language feature of the phoneme frame, and each unit of the output layer is given the acoustic feature of the phoneme frame. This pre-learning is performed in units of phoneme frames. That is, the acoustic feature DNN is given the language feature of the phoneme frame and the acoustic feature of the phoneme frame for each phoneme frame, and pre-learning is performed. In speech synthesis, the acoustic feature amount DNN of the phoneme frame is used for each phoneme frame, and the acoustic feature amount of the phoneme frame is generated and output based on the language feature amount of the phoneme frame.

FIG. 7 is a diagram showing an example of language feature data. As described above, the language feature quantity includes attributes related to phonemes, attributes related to mora, attributes related to accent phrases, attributes related to exhalation paragraphs, and attributes related to utterance, and is represented by a d-dimensional vector at time t. As the language feature quantity, the information of each attribute at time t is represented by a numerical value. Although the details of each attribute will be omitted, as an example, the attribute related to the accent phrase includes "ascending position of the mora in the accent phrase". However, "corresponding" means that it is the target of the processing when the normalization processing is performed.

Here, in the above-mentioned normalization processing means 144, the linguistic features of the phonemes text-analyzed by the text analysis means 140 and the linguistic features of the text-analyzed phonemes are processed by the frame processing means 142. The language features of the phoneme frame are input. The input language features are vector series and can be represented by Equation 1. As shown in Equation 2, the normalization processing means 144 has the first attribute (first attribute value) different from the first attribute and has a larger absolute value than the first attribute. Alternatively, the language features are normalized by dividing by the same second attribute (second attribute value), and the normalized language features are output. However, it is assumed that the first attribute and the second attribute are related. Further, as shown in FIG. 7, the language feature amount L for one utterance is a series of language feature amount vectors having a d-dimensional attribute at time t as an element. However, in Equation 2, | and | mean absolute values. In this embodiment, the absolute value of the first attribute value is equal to or less than the absolute value of the second attribute value.

Further, the above "related items" will be described with reference to FIGS. 6 and 7. When the text is analyzed by the text analysis means 140, as shown in FIG. 7, information having attributes related to utterances, exhalation paragraphs, accent phrases, mora, and phonemes can be obtained. These pieces of information have a hierarchical structure as shown in FIG. 6, and the information of each layer is mainly composed of the information of the lower layers. For example, the hierarchy of accent phrases has a hierarchy of mora and phonemes at a lower level, and the attributes of the accent phrase include the ascending position of the mora in the accent phrase and the total number of mora in the accent phrase. Therefore, in the normalization process in the normalization processing means 144 of this embodiment, basically, the attribute related to the position is divided by the attribute related to the total number of the same layer under the relation of the same layer, and the attribute related to the total number is called the total number. Under relevance, it will be divided by the attribute related to the total number of different hierarchies. The continuation length is the same as the total number. The normalization process is applied to all layers.

The normalization method shown in Equation 2 does not include conditions other than the utterance such as the mean and variance calculated from all the training data or the minimum and maximum values as in Non-Patent Document 1 and Non-Patent Document 2. 1. Since it is calculated under the limited conditions within one utterance, no outliers occur. Therefore, it is possible to avoid deterioration of the prediction performance due to outliers, and it is possible to predict the acoustic feature amount more stably than before.

FIG. 8 is a flow chart showing an example of the attribute normalization process of the CPU of the speech synthesizer 10 shown in FIG. The program for this attribute normalization process (corresponding to the "language processing program") is stored in the memory of the speech synthesizer 10 and executed by the CPU. The same applies to other programs and data required for learning processing and speech synthesis processing.

Further, in this embodiment, a process for normalizing the attribute of "ascending position of mora in the accent phrase" among a plurality of attributes included in the language feature will be described. Although detailed description is omitted, the same attribute normalization process is executed for other attributes.

As shown in FIG. 8, when the CPU starts the attribute normalization process, the CPU executes the first attribute value calculation process (see FIG. 9) described later in step S1 and calculates the second attribute value described later in step S3. The process (see FIG. 10) is executed.

In the next step S5, the variable t for time is initialized and the array array m [T] is set. That is, the CPU assigns 0 to the variable t and sets an array m [T] capable of storing the elements up to the maximum value T of the phoneme frame (here, the ascending position of the mora in the accent clause). However, the variable t is a variable for counting the number of frames. This also applies to the first attribute value calculation process (FIG. 9) and the second attribute value calculation process (FIG. 10), which will be described later. The array array m [T] is an array for storing normalized attribute values. The maximum value T is the total number of phoneme frames in one utterance.

In the next step S7, it is determined whether or not the variable t is smaller than the maximum value T of the phoneme frame. If the value is "NO" in step S7, that is, if the variable t is equal to or greater than the maximum value T of the phoneme frame, it is determined that the "ascending position of the mora in the accent phrase" is normalized for all mora, and the attribute is attributed. End the normalization process.

On the other hand, if "YES" in step S7, that is, if the variable t is less than the maximum value T of the phoneme frame, there remains a mora that does not normalize the "ascending position of the mora in the accent clause". Then, in step S9, the element m [t] whose ascending position is in the variable t is calculated (m [t] = x [t] / y [t]), and in step S11, the variable t is added by 1. (T = t + 1), the process returns to step S7.

FIG. 9 is a flow chart showing a first attribute value calculation process of the CPU executed in step S1 shown in FIG. 8 and step S51 of FIG. 10 described later. Hereinafter, the first attribute value calculation process will be described, but the same process as the process already described will be briefly described.

As shown in FIG. 9, when the CPU starts the first attribute value calculation process, the variable t and the variable i are initialized in step S21 (t = 0, i = 1), and the array array x [T. ] Is prepared (x [0], x [1],…, x [T-1]). However, the variable i is a variable for counting the ascending position of the mora in the accent phrase. The same is true in FIG. The array array x [T] is an array for storing the ascending position of the mora for each attribute value.

In the next step S23, it is determined whether or not the variable t is smaller than the maximum value T of the phoneme frame. If "NO" in step S23, the first attribute value calculation process is terminated, and the process returns to the attribute normalization process shown in FIG.

On the other hand, if "YES" in step S23, the numerical value of the variable i is assigned to the element x [t] in step S25. That is, the ascending number of the mora in the accent phrase is assigned. In the next step S27, it is determined whether the variable t is the end of the mora. Here, the CPU determines whether the variable t indicates the final frame in the mora. If "YES" in step S27, that is, if the variable t is not the end of the mora, the process proceeds to step S31. On the other hand, if it is "NO" in step S27, that is, if the mora ends in the variable t, the variable i is added by 1 in step S29 (i = i + 1), and the process proceeds to step S31.

In step S31, it is determined whether or not the variable t is the end of the accent phrase. Here, the CPU determines whether the variable t indicates the final frame in the accent clause.

If "NO" in step S31, the process proceeds to step S35 if it is not the end of the accent phrase in the variable t. On the other hand, if "YES" in step S31, if the end of the accent phrase in the variable t, the variable i is set to the initial value in step S33, and the variable t is added by 1 in step S35 (t =). t + 1), the process returns to step S23.

FIG. 10 is a flow chart of the second attribute value calculation process in step S3 shown in FIG. As shown in FIG. 10, when the second attribute value calculation process is started, the first attribute value calculation process shown in FIG. 9 is executed in step S51.

In the next step S53, the variable t and the variable i are initialized (t = T, i = x [T-1]), and the array array y [T] is prepared (y [0], y [1]]. ,…, y [T-1]). However, the array array y [T] is an array for storing each element of the second attribute value (here, the total number of mora in the accent clause) for normalizing each element of the first attribute value. Is. In this second attribute value calculation process, the array array y [T] is assigned a value from the last element y [T-1] to the first element y [0].

Subsequently, in step S55, it is determined whether or not the variable t is larger than 0. If "NO" in step S55, the second attribute value calculation process is terminated, and the process returns to the attribute normalization process shown in FIG.

On the other hand, if "YES" in step S55, the variable t is subtracted by 1 in step S57 (t = t-1), and in step S59, it is determined whether or not the end of the accent phrase in the variable t is reached. If "NO" in step S59, that is, if it is not the end of the accent phrase in the variable t, the process proceeds to step S63.

On the other hand, if "YES" in step S59, that is, if it is the end of the accent clause in the variable t, the element x [t] is assigned to the variable i in step S61 (i = x [t]), and further. , In step S63, the variable i is assigned to the element y [t] (y [t] = i), and the process returns to step S55.

According to this embodiment, the linguistic features are normalized by taking the ratio of linguistically related attributes using only the values in one utterance, so that outliers can be prevented from occurring. it can. Therefore, the prediction accuracy of the acoustic feature amount is good.

Further, according to this embodiment, since outliers do not occur, it is not necessary to increase the learning data in order to prevent outliers from occurring. That is, according to this embodiment, the prediction accuracy of the acoustic feature amount is good even with a small amount of learning data.

In this embodiment, the input value of the DNN is set on the condition that the absolute value of the first attribute value (L _td ) is equal to or less than the absolute value of the second attribute value (L _{t δ) (Equation 2).} The language features (each attribute) have been normalized so that they fall between 0 and 1 (or in the range), but are not limited to this. For example, the absolute value of the first attribute value with the proviso greater than the absolute value of the second attribute value, or by adding a predetermined constant to the first attribute value (L _td) and a second attribute value (L _T.DELTA.) The scale may be changed by multiplying the value so that the value after normalization exceeds the range of 0 to 1. In this case, since the scale of each attribute value is only changed, the same effect as before the scale is changed can be obtained. Further, depending on the range of values after normalization, the condition may be that the absolute value of the second attribute value is larger than the absolute value of the first attribute value (| L _td | <| L _{tδ |).}

It should be noted that the specific numerical values shown in the above-described examples are merely examples, and should not be limited, and can be appropriately changed depending on the product to be implemented and the like.

10 ... Speech synthesizer 12, 16 ... Storage unit

Claims (12)

A language processing device that is input to the deep neural network of a speech synthesizer that generates synthetic speech and normalizes a language feature vector series composed of multiple different attributes.
A normalization means for normalizing the first attribute in the language feature quantity vector series for one utterance with a second attribute different from the first attribute in the language feature quantity vector series for the one utterance is provided. , Language processor. The language processing device according to claim 1, wherein the first attribute and the second attribute are linguistically related values. The language processing apparatus according to claim 1 or 2, wherein the normalization means normalizes by dividing the first attribute by the second attribute. The language processing apparatus according to any one of claims 1 to 3, wherein the absolute value of the first attribute is equal to or less than the absolute value of the second attribute. A language processing program that is input to the deep neural network of a speech synthesizer that generates synthetic speech and is executed by a language processor that normalizes a language feature vector series composed of multiple different attributes.
To the processor of the language processing device, the first attribute in the language feature quantity vector series for one utterance is set to a second attribute different from the first attribute in the language feature quantity vector series for the one utterance. A language processing program that executes a normalization step to normalize. The language processing program according to claim 5 , wherein the first attribute and the second attribute are linguistically related values. The language processing program according to claim 5 or 6, wherein the normalization means normalizes by dividing the first attribute by the second attribute. The language processing program according to any one of claims 5 to 7, wherein the absolute value of the first attribute is equal to or less than the absolute value of the second attribute. It is a language processing method of a language processing device that is input to the deep neural network of a speech synthesizer that generates synthetic speech and normalizes a language feature vector series composed of multiple different attributes.
To the processor of the language processing device, the first attribute in the language feature quantity vector series for one utterance is set to a second attribute different from the first attribute in the language feature quantity vector series for the one utterance. A language processing method that executes normalization processing. The language processing method according to claim 9, wherein the first attribute and the second attribute are linguistically related values. The language processing method according to claim 9 or 10, wherein the first attribute is normalized by dividing by the second attribute. The language processing method according to any one of claims 9 to 11, wherein the absolute value of the first attribute is equal to or less than the absolute value of the second attribute.

Images