JP4778402B2 - Pause time length calculation device, program thereof, and speech synthesizer - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a pause duration calculation device and its program, which calculates pause duration between voice component data by which natural listening feeling is obtained, when connecting the voice component data in voice synthesizing of a recording and editing system, and a voice synthesizer equipped with the pause duration calculation device. <P>SOLUTION: In voice synthesizing by connecting voice component data in which a voice waveform produced by uttering a predetermined unit of text is recorded, the pause duration calculation device 40 comprises: an acoustic feature amount detector 410 for detecting a predetermined acoustic feature amount in the voice waveform which is recorded in the voice component data; an acoustic distance calculation section 430 for calculating an acoustic distance between the acoustic feature amount of a preceding voice component data which precedes in voice component data connected together, and the acoustic feature amount of a following voice component data which follows it; and a pause duration calculation section 440 for calculating pause duration which is inserted between the voice component data by using a calculation formula which is set beforehand. <P>COPYRIGHT: (C)2008,JPO&amp;INPIT

Description

  The present invention relates to a speech synthesis technique based on a recording / editing method, in which speech synthesis is performed by connecting speech component data in which speech uttered by text is recorded in advance.

  In speech synthesis using the so-called recording and editing method, which connects speech component data in which words, phrases, fixed phrases, etc. are recorded in advance, the vocabulary and sentence types that can be synthesized are limited, but the unit of speech to be handled May be relatively long, resulting in higher quality audio, but the continuity of audio acoustic features (pitch frequency, speech speed, power, spectral envelope, etc.) at the connection between audio component data Had an effect.

  Therefore, for example, in the case where a word or phrase is sandwiched between fixed phrases or a sentence is composed by combining words or phrases, for example, the acoustic features before and after each audio component data at the time of connection are considered. In general, such as recording the voice that was uttered, when the audio component data is connected from the recording stage of the audio component data, it is devised so that the discontinuity of the acoustic features does not occur, or the discontinuity does not matter In this way, audio component data is connected with a long pause.

  On the other hand, in the conventional rule-based speech synthesis that can synthesize any text, on the other hand, in order to obtain a more natural synthesized speech, the pitch frequency, duration length, power, and pause between speech segments are obtained. It is necessary to finely control the time length (pause length), and various prosodic control methods have been proposed.

  Among them, a control rule based on analysis of actually spoken voice data has also been proposed for setting a pause length. For example, Patent Document 1 proposes a method for setting a pause length in accordance with input synthetic speech information. The speech rule synthesizer described in Patent Literature 1 is based on information such as the phoneme continuation length and the fundamental frequency of the input synthesized speech, and the dependency relationship between the preceding phrase and the succeeding phrase based on a preset pause setting rule. Depending on the presence or absence of punctuation marks, the presence or absence of pauses at the phrase boundary between the preceding phrase and the succeeding phrase, and if inserted, the reference value is set according to the type of pause and the pause length is affected. The pose length is corrected by grammatical factors and set to be an integral multiple of 1 mora length.

Further, in Patent Document 2, a phoneme string database is created in advance by extracting speech waveform data of a phoneme string from voice waveform data in which a person's utterance is recorded, and the phoneme string constituting the text information that is the target of speech synthesis. A recording and editing type speech synthesizer for connecting corresponding phoneme string data has been proposed. When connecting the phoneme string data, the speech synthesizer described in Patent Document 2 has a shorter silent part length at the rear end of the preceding phoneme string data and a shorter silent part length at the front end of the subsequent phoneme string data. By making the silent part length the silent part length of the phoneme string data connection part, natural speech synthesis can be performed even if any silent part length is extremely long.
Japanese Patent No. 3060422 (paragraphs 0019 to 0023, FIG. 3) JP 2002-41077 A (paragraph 0012)

  However, in the conventional speech editing method of speech synthesis, when recording is performed so as not to cause discontinuity in the acoustic characteristics of the speech at the connection part between the speech component data, the recorded speech is auditioned. In other words, the utterer or the recording worker was overburdened by referring to the stimulus that is the target of the utterance or by proceeding with recording while checking the degree of conformity with the recorded voice.

  In addition, there is a problem in that when the connection is made with a somewhat long pose in consideration of safety so that discontinuities are not anxious, the connection is extended and the naturalness is impaired. In order to avoid this, in order to set a pose having an appropriate length, it is necessary to repeat trials by changing the length and auditioning, which is inefficient. In general, since a fixed-length pose is used, the reading method is regular and tends to be a so-called mechanical impression.

  On the other hand, the speech rule synthesizer described in Patent Document 1 is used for rule synthesis. However, the dependency relationship between the input synthesized speech information and the preceding phrase and the succeeding phrase, the presence or absence of reading marks, and the pause Depending on the type, after setting a reference value, the pose length is corrected by a grammatical factor that affects the pose length, and application to a recording editing system is also conceivable.

However, when the device described in Patent Document 1 is applied to the recording and editing method, the same grammatical structure is used when a fixed sentence is composed by combining audio parts data or when a word or phrase is inserted between fixed phrases. It becomes a sentence, and the same pose length is given if it has the same grammatical structure.
In rule synthesis, the acoustic features in the environment before and after the pose can be controlled and aligned, so there is no problem even if the same pose length is given to sentences with the same grammatical structure. If an appropriate pause length is not set in consideration of the acoustic characteristics of the speech at the connection between the audio component data, discontinuity will occur at the connection of the audio component data, and the naturalness of the synthesized speech will be impaired. Become. In addition, by giving the same pose length to sentences having the same grammatical structure, there is a problem that the reading method is regular and a so-called mechanical impression is obtained.

  In addition, the speech synthesizer described in Patent Document 2 adopts the shorter silence part length between the phoneme string data to be connected, and the shorter silence part length at the end of the phoneme string data created by chance. The continuity of features was not fully considered.

  Therefore, the present invention relates to a pause time length calculation device for calculating a pause time length between voice component data that can obtain a natural audibility when voice component data is connected in voice synthesis of a recording and editing method, a program thereof, and An object of the present invention is to provide a speech synthesizer provided with this pause time length calculation device.

For this purpose, the pause time length calculating device according to claim 1 connects voice component data connected to each other when voice component data in which a voice waveform uttering a predetermined unit of text is recorded and voice synthesis is performed. a pause time length calculation device that calculates the rest time length to be inserted between, and the acoustic feature value acquiring means, acoustically distance calculating means, and the rest time length calculating unit, configured to include a.

According to such a configuration, the pause time length calculation device, as the respective acoustic feature amounts in the audio waveform recorded in the preceding audio component data preceding and the subsequent audio component data following in the audio component data connected to each other , At least one of the pitch frequency representing the pitch of the voice, the speech speed representing the speed of the speech, the power representing the volume of the speech, or the spectrum envelope representing the sound of the speech is obtained by the acoustic feature quantity acquisition means. Next, an acoustic distance representing the acoustic difference between the acoustic feature quantity of the preceding voice component data acquired by the acoustic feature quantity acquisition means and the acoustic feature quantity of the subsequent voice component data by the acoustic distance calculation means. Calculate the distance. Then, the pause time inserted between the preceding voice component data and the subsequent voice component data by using a preset calculation formula based on the acoustic distance calculated by the acoustic distance calculating unit by the pause time length calculating unit. Calculate the length.

The rest time length calculation apparatus according to claim 2 is a regression using the acoustic distance calculated by the acoustic distance calculation means as an explanatory variable as the calculation formula in the rest time length calculation apparatus according to claim 1. The formula was used.

  According to such a configuration, the pause time length calculation device can calculate the pause time length by performing a product-sum operation on the acoustic distance between the audio component data connected to each other and the regression coefficient.

According to a third aspect of the present invention, there is provided a program for calculating a pause time length between voice component data, wherein voice components connected to each other when voice component data recording a voice waveform obtained by uttering a predetermined unit of text is connected and synthesized. In order to calculate the pause time length to be inserted between the component data, the computer is configured to function as an acoustic feature quantity acquisition unit, an acoustic distance calculation unit, and a pause time length calculation unit.

According to such a configuration, the pause time length calculation program between the audio component data is recorded in the audio waveforms recorded in the preceding audio component data preceding and the subsequent audio component data following in the audio component data connected to each other. Acquire acoustic feature quantity as at least one of pitch frequency representing voice pitch, speech speed representing speech speed, power representing speech volume, or spectrum envelope representing sound of speech. Obtain by means. Next, an acoustic distance representing the acoustic difference between the acoustic feature quantity of the preceding voice component data acquired by the acoustic feature quantity acquisition means and the acoustic feature quantity of the subsequent voice component data by the acoustic distance calculation means. Calculate the distance. Then, the pause time inserted between the preceding voice component data and the subsequent voice component data by using a preset calculation formula based on the acoustic distance calculated by the acoustic distance calculating unit by the pause time length calculating unit. Calculate the length.
Thereby, the pause time length calculation program between audio component data can calculate an appropriate pause time length according to the acoustic distance of the audio component data connected to each other.

The speech synthesizer according to claim 4 is a speech synthesizer that synthesizes speech by connecting speech component data in which a speech waveform of a predetermined unit of text is uttered, comprising speech component data storage means, An information acquisition unit, a voice part data acquisition unit, and a pause time length calculation device are provided.

According to such a configuration, the speech synthesizer first includes text that is a target of speech synthesis and is continuously read out in a predetermined order by the reading information acquisition unit, or a predetermined unit that constitutes the text Read-out information consisting of information designating audio component data corresponding to the text of is acquired. Next, the desired audio component data is acquired by the audio component data acquisition unit from the audio component data storage unit that stores the audio component data in which the audio waveform has been recorded in advance based on the reading information acquired by the reading information acquisition unit. Then, the pause time length calculation device calculates the pause time length inserted between the voice component data constituting the text to be synthesized, and sets the pause time length as the pause time between the voice component data.
Thus, the speech synthesizer can create speech synthesis data in which a pause time corresponding to the acoustic distance between the speech component data connected to each other is inserted between the speech component data.

According to the first or third aspect of the present invention, the pause time to be inserted between the voice component data connected to each other is calculated according to the acoustic distance between the voice component data in the voice synthesis of the recording and editing method. Therefore, the sound reproduced by inserting the pause time can be a synthetic sound with a natural audibility that does not give a discontinuity or a mechanical impression.
In addition, the pause time to be inserted between the voice component data is the difference in pitch frequency representing the pitch of the voice, the difference in speech speed representing the speed of speech, the difference in power representing the size of the voice, or between the voice component data. In order to calculate the pause time length according to the difference in the spectral envelope, which is the sound of the voice, the voice played by inserting this pause time, the voice height represented by the acoustic feature, speech speed, voice Or a synthesized speech with a natural sensation that does not give a discontinuity between the size of the sound and the sound reverberation.
According to the second aspect of the present invention, since the pause time length is calculated by a regression equation having the acoustic distance as an explanatory variable, it is simply calculated by a product-sum operation of the acoustic distance and the coefficient of the regression equation. Can do.
According to the fourth aspect of the present invention, the pause time to be inserted between the voice component data connected to each other is calculated and set according to the acoustic distance between the voice component data in the voice editing of the recording and editing method. Therefore, by reproducing the voice synthesis data created by inserting this pause time, it is possible to obtain a natural audible voice that does not give a discontinuity or a mechanical impression.

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings as appropriate.
<Configuration of speech synthesizer>
First, with reference to FIG. 1, the structure of the speech synthesizer 100 provided with the pause time length calculation device 40 according to the present invention will be described. Here, FIG. 1 is a block diagram showing the configuration of the speech synthesizer of this embodiment.

  The speech synthesizer 100 according to the present embodiment shown in FIG. 1 includes a reading information input unit 10, a speech component data acquisition unit 20, a speech component data storage unit 30, a pause time length calculation device 40, a speech synthesis data storage unit 50, and a speech. A playback unit 60 is provided.

  Here, before explaining the details of each part, the principle of setting the configuration of the read-out information and the voice component data and the pause time length of the present embodiment will be described with reference to FIGS.

First, the configuration of the reading information will be described with reference to FIG. FIG. 2 is an explanatory diagram for explaining the configuration of the reading information.
In the example of the reading information shown in FIG. 2, “sentence” is a unit of “sentence”, and “sentence” is constituted by one or a plurality of phrases, and further, the reading information is constituted by a plurality of sentences. i, represents the number of sentences that constitute the reading information N, the phrase number in each sentence j, the number of clauses that constitute each sentence in M _i. Each sentence is continuously read out in the order of sentence number i, and the clauses of each sentence are continuously read out in the order of phrase number j.
Further, phr [i] [j] represents one piece of audio component data corresponding to the j-th clause in the i-th sentence.

  In the present embodiment, the audio component data is configured in units of phrases. However, the present invention is not limited to this, and may be in units of phonemes, words, morphemes, phrases, sentences, or the like. You may make it comprise audio | voice component data by mixing.

  Next, an example of the configuration of audio component data will be described with reference to FIGS. Here, FIG. 3 is a diagram illustrating the data structure of the audio component data, and FIG. 4 is a configuration diagram schematically illustrating the configuration of the audio waveform data included in the audio component data.

  The data structure of the audio component data shown in FIG. 3 includes, as basic data, an audio component number, reading (text) data, audio waveform data, number of beats (number of mora), and data length (total time length). The data related to the feature amount includes the leading end silent length, trailing end silent length, leading end unvoiced sound length, trailing end unvoiced sound length, leading end pitch frequency, trailing end pitch frequency, average speech speed, average power, leading end spectral envelope, trailing end. A spectrum envelope is included, and additional data (setting data) includes a pause time length.

  In the present embodiment, basic data is set in advance for voice component data used for voice synthesis in the recording and editing method, and is stored in the voice component data storage unit 30 (see FIG. 1).

  The basic data includes a voice part number for identifying a voice part, read-out data indicating the contents of the voice part, that is, text data, voice waveform data obtained by recording a voice uttered by the speaker, and text. The number of beats (number of mora) of data and the data length (total time length of speech waveform data) are set.

For example, in the example shown in FIG. 3, "123456" as the audio part number, as speech data "K Broadcast (alarm going)", a predetermined sampling frequency as the audio waveform data (see _{P A} in FIG. 4) (e.g., Digital data sampled at a frequency of about several kHz to several tens of kHz, “6” is set as the number of beats, and “1200 (ms)” is set as the data length.

  The data related to the acoustic feature amount is intermediate data for calculating the suspension time length in the suspension time length calculation device 40, and is calculated by the acoustic feature amount detection unit 410 in the process of calculating the suspension time length. Data to be set.

  The pause time length that is additional data is data that is calculated and set by the pause time length calculation device 40 (see FIG. 1). This data is data determined based on a difference (acoustic distance) in the acoustic feature amount from the audio component data connected behind in the audio synthesis, and is audio component data having the same audio component number. Also, the value differs depending on the position where the audio component data designated by the sentence number i and the phrase number j is used in the reading information.

Next, each data will be described with reference to FIG. 3 and FIG.
FIG. 4 shows the speech waveform data of the preceding speech component data P _A preceding in speech synthesis and the subsequent speech component data P _B connected after the preceding speech component data.

In the example shown in FIG. 4, the preceding audio component data P _A is the audio component data recorded shares of brand "K broadcast (Alarm likely)" succeeding audio component data P _B is stock "40 yen ( This is the audio component data that recorded "Yonjuen)". FIG. 4 shows the speech waveform.
In FIG. 4, the left-right direction is the time axis, and time passes from left to right.

As shown in FIG. 4, the data start position, the voice start position, the voiced sound start position, the voiced sound end position, the voice end position, and the data end position are respectively obtained for the voice waveforms of the voice component data P _A and P _B. Can be determined.

  The data length is the total data length from the data start position that is the start point of the voice waveform data to the data end position that is the end point of the voice waveform data.

  The leading end silent length and the trailing end silent length are the lengths of the silent sections “between the data start position and the voice start position” and “between the voice end position and the data end position” of the voice waveform data, respectively.

  The leading non-voiced sound length and the trailing non-voiced sound length are voiced sounds of “between the data start position and the voiced sound start position” and “between the voiced sound end position and the data end position” of the voice waveform data, respectively. It is the length of the non-voiced sound section that does not include. The unvoiced sound section includes a silent section and an unvoiced sound section.

  The voice section length can be calculated by subtracting the leading end silent length and the trailing end silent length from the data length.

  The voice start position, the voiced sound start position, the voiced sound end position, and the voice end position are obtained by acoustic analysis of the voice waveform data by the acoustic feature quantity detection unit 410 (see FIG. 1) of the pause time length calculation device 40. And the length of each section can be calculated based on the detected voice start position, voiced sound start position, voiced sound end position and voice end position, and data start position and data end position.

Quiescence period is the time length of the pause (silence) to be inserted into the connecting portion between the preceding sound component data P _A during voice reproduction and subsequent audio component data P _B, quiescence period of quiescence period calculating unit 40 calculated by the calculating unit 440 (see FIG. 1), it is set to the preceding audio component data P _a by rest time length setting unit 450 (see FIG. 1).

  Next, with reference to FIG. 5 (refer to FIG. 3 as appropriate), the principle of setting the pause time length according to the present invention will be described. Here, FIG. 5 is an explanatory diagram for explaining the principle of setting the pause time length according to the present invention.

The present invention, by quiescence period calculation unit 40 (see FIG. 1), the connecting portion between the preceding sound component data P _A and the subsequent audio component data P _B, and the preceding audio component data P _A and the subsequent audio component data P _B The pause time (silent state) is set based on the difference between the acoustic feature quantities. As a result, the audio waveforms included in the audio component data P _A and P _B can be connected and reproduced so as to have a natural audibility.

FIG. 5 shows a case where the pitch frequency F0 is used as an example of the acoustic feature quantity.
Using the pitch frequency F0, preceding audio component data P _A and the succeeding speech components when connecting the data P _B is preceding speech component data P _A rear pitch frequency ed at voiced end of. F0 _A and the tip pitch frequency st. At the voiced sound start position of the subsequent audio component data P _B. The difference between F0 _B, i.e. acoustic "distance" (hereinafter, appropriately referred to as "acoustic distance") based on, inserted into the connection portion between the preceding sound component data P _A and the subsequent audio component data P _B Set the pause time length to be used.

In the audio component data P _A and P _B connected to each other, when the acoustic distance is large, when the audio waveform is continuously reproduced as it is, the sound becomes discontinuous and unnatural. In addition, regardless of the acoustic distance, when a longer pause time is inserted, a feeling of being extended may occur and an impression of mechanical reading is given.

Therefore, in the present invention, prior audio component data P _A and the subsequent audio component data P _B and acoustic distance as set longer downtime to be inserted into the connecting portion larger, inserting the connecting portion as acoustic distance is smaller By setting the pause time short, it is possible to reproduce sound with a natural audibility.

  In addition, although the position which refers to the acoustic feature amount in the speech waveform varies depending on the acoustic feature amount to be used, details will be described later.

Returning to FIG. 1, the configuration of each unit of the speech synthesizer 100 will be described.
A read-out information input unit (read-out information acquisition means) 10 is an input unit for inputting read-out information to be subjected to speech synthesis. For example, a magnetic disk device, an optical disc device, or a flash memory in which read-out information is stored. Read-out information to be subjected to speech synthesis is read out and acquired from a storage device such as the above. Further, the reading information may be input via a communication line such as a network or a telephone line, or may be input via an input device such as a keyboard, and the input means is not particularly limited.
The reading information input unit 10 outputs the input reading information to the voice component data acquisition unit 20.

  In the present embodiment, the reading information input unit inputs voice component data corresponding to a phrase as information for specifying the text. However, normal text data is input as reading information, and the phrase is read using an appropriate method. Etc., and may be associated with voice component data prepared in advance.

  The speech component data acquisition unit (speech component data acquisition means) 20 receives the speech information output from the speech information input unit 10 and sequentially speeches the speech component data corresponding to “sentences” constituting the input speech information. The information is read out from the component data storage unit 30 and is obtained and output to the acoustic feature amount detection unit 410 of the pause time length calculation device 40.

The voice component data storage unit (voice component data storage means) 30 is a storage device such as a magnetic disk device, an optical disk device, or a semiconductor memory that stores voice component data. The set audio component data is stored in advance.
The audio component data stored in the audio component data storage unit 30 is appropriately read out by the audio component data acquisition unit 20.

The pause time length calculation device 40 includes an acoustic feature quantity detection unit 410, a preceding voice component data storage unit 420, an acoustic distance calculation unit 430, a pause time length calculation unit 440, and a pause time length setting unit 450. Yes.
The pause time length calculation device 40 receives the audio component data output from the audio component data acquisition unit 20, detects an acoustic feature amount from the input audio component data, and uses the detected acoustic feature amount to mutually Calculate the acoustic distance between the connected audio component data, calculate the pause time length between the connected audio component data based on the calculated acoustic distance, and use the calculated pause time length as the voice component data It is added (set) and stored in the speech synthesis data storage unit 50.
The detailed configuration of each part of the downtime calculation device 40 will be described later.

The voice synthesis data storage unit 50 stores the voice component data in which the pause time length is set by the pause time length setting unit 450 of the pause time length calculation device 40, for example, a storage such as a magnetic disk device, an optical disk device, and a semiconductor memory. Device.
The voice synthesis data storage unit 50 sequentially stores the voice component data in which the pause time length is set in association with the sentence number and the phrase number specified by the reading information input by the reading information input unit 10. Then, by storing the voice component data in which the pause time length is set corresponding to all the phrases included in the reading information, the voice synthesis data corresponding to the reading information is formed in the voice synthesis data storage unit 50. The
The voice synthesis data formed in the voice synthesis data storage unit 50 is read by the voice reproduction unit 60.

The voice reproduction unit 60 reads out the voice synthesis data formed in the voice synthesis data storage unit 50 and corresponding to the read-out information, and sequentially converts the voice waveform data included in the voice component data associated with the sentence number and the phrase number into analog analog data. The audio waveform signal is reproduced and the reproduced audio waveform signal is output to the speaker 70.
The audio reproduction unit 60 reproduces the audio waveform data included in the audio component data into the audio waveform signal, and inserts a silent state for the pause time length set in the audio component data, and then reproduces the next audio component data. I do.

  The speaker 70 receives the audio waveform signal output from the audio reproduction unit 60, converts the input audio waveform signal into a sound wave, and reproduces it in an audible manner.

  In the present embodiment, the speech synthesizer 100 reproduces the speech synthesis data formed by the speech component data in which the pause time length is set to the speech waveform signal by the speech reproduction unit, outputs it to the speaker, and is audibly reproduced. However, for example, the voice synthesis data may be transmitted via a communication line such as a network or a broadcast wave, and the voice synthesis data may be reproduced on the receiving device side.

<Configuration of pause time length calculation device>
Next, with reference to FIG. 6 (refer to FIG. 1 as appropriate), the configuration of the pause time length calculation apparatus 40 of the present embodiment will be described in detail. Here, FIG. 6 is a block diagram showing a configuration of the pause time length calculation apparatus of the present embodiment.

  The pause time length calculation device 40 shown in FIG. 6 includes an acoustic feature quantity detection unit 410, a preceding voice component data storage unit 420, an acoustic distance calculation unit 430, a pause time length calculation unit 440, and a pause time length setting unit 450. It is prepared for.

  The acoustic feature quantity detection unit (acoustic feature quantity acquisition unit) 410 includes a framing processing unit 411, a spectrum analysis unit 412, a pitch frequency detection unit 413, a speech speed detection unit 414, a power detection unit 415, and a spectrum envelope detection unit 416. The audio component data output from the audio component data acquisition unit 20 is input, the audio waveform data included in the input audio component data is analyzed to detect the acoustic feature amount, and Data relating to the acoustic feature quantity (see FIG. 3) is set in the voice part data, and the voice part data in which the data relating to the acoustic feature quantity is set is output to the acoustic distance calculation unit 430 and the preceding voice part data storage unit 420 To remember.

  The acoustic feature amount detection unit 410 according to the present embodiment includes a pitch frequency detection unit 413, a speech speed detection unit 414, a power detection unit 415, and a spectrum envelope detection unit 416, respectively. Four acoustic features are detected: “speech speed” representing the speaking speed, “power” representing the loudness of the speech, and “spectrum envelope” representing the sound of the speech.

  In the present embodiment, the above-described four acoustic feature quantities are detected. However, one or a plurality of acoustic feature quantities may be detected, for example, the duration of the phoneme at the end. You may make it detect other acoustic feature-values, such as length.

Next, details of each part of the acoustic feature quantity detection unit 410 will be described.
The framing processing unit 411 performs a framing process for cutting out the audio waveform data from the audio waveform data included in the input audio component data using a window function at a predetermined interval.
When performing the framing process, for example, the frame length is about 20 to 40 ms, the frame interval is about 5 to 20 ms, and a Hamming window, Hanning window, triangular window, or the like can be used as a window function.
The voice waveform data subjected to framing processing is output to the spectrum analysis unit 412.

The spectrum analysis unit 412 performs spectrum analysis on the speech waveform data subjected to framing processing output from the framing processing unit 411.
As a spectrum analysis method, for example, Fourier spectrum analysis, LPC analysis (linear prediction analysis), cepstrum analysis, or the like can be used, and a power spectrum, a prediction coefficient, a cepstrum, and the like are calculated as spectrum data.
The calculated spectrum data is output to the pitch frequency detection unit 413, the speech speed detection unit 414, the power detection unit 415, and the spectrum envelope detection unit 416 for detecting the acoustic feature amount.

Next, the configuration of the pitch frequency detector 413 will be described with reference to FIGS. Here, FIG. 7 is an explanatory diagram for explaining how the pause time length is set based on the pitch frequency.
As shown in FIG. 6, the pitch frequency detection unit 413 includes an end non-voiced sound length detection unit 413a and an end pitch frequency detection unit 413b.

In the present embodiment, as shown in FIG. 7, in the case of using the pitch frequency as the acoustic feature quantity, rear pitch frequency at the rear end of the speech waveform of the preceding audio component data _{P A (phr [i] [} j] and .Ed.F0), subsequent audio component data P tip pitch frequency at the tip of the voice waveform _B (phr [i] based on [j + 1] .st.F0) and, pause time length (phr [i] [j] .pau).
Since the pitch frequency cannot be extracted from unvoiced sound, the pitch frequency detected from the first frame containing voiced sound in the voice section is detected as the leading pitch frequency and detected from the last frame containing voiced sound in the voice section. Is detected as the rear end pitch frequency.

  The pitch frequency can be obtained, for example, by obtaining an autocorrelation function of the power spectrum, extracting the first peak of the autocorrelation function, and obtaining the frequency of the extracted first peak, performing cepstrum analysis, It can also be obtained by extracting the peak of the kerfrenchy part and calculating the reciprocal of the extracted kerfrenci. Further, the pitch frequency may be detected by other methods.

  The end non-voiced sound length detection unit 413a detects whether or not a voiced sound is included in each frame by analyzing the spectrum data of the framed speech waveform. Then, the position of the frame including the voiced sound that appears first is detected as the voiced sound start position. Further, the position of the last frame including the voiced sound is detected as the voiced sound end position.

  The tip non-voiced sound length can be calculated from the detected voiced sound start position and data start position. If the data start position is simply defined as “0 (ms)”, the voiced sound start position coincides with the tip non-voiced sound length. Further, the rear end non-voiced sound length can be calculated from the detected voiced sound end position and data end position. If the data start position is “0”, the data end position coincides with the data length. Therefore, the rear end non-voiced sound length can be calculated by subtracting the voiced sound end position from the data length.

  The end pitch frequency detection unit 413b detects the pitch frequency from the spectrum data corresponding to the frame of the voiced sound start position detected by the end non-voiced sound length detection unit 413a and sets it as the tip pitch frequency. The pitch frequency is detected from the spectrum data of the frame and set as the rear end pitch frequency.

  The pitch frequency detection unit 413 converts the leading non-voiced sound length and the trailing non-voiced sound length detected by the end non-voiced sound length detection unit 413a into phr [i] [j] .st.pos2 of the audio component data, respectively. And phr [i] [j] .ed.pos2, and the leading edge pitch frequency and the trailing edge pitch frequency detected by the edge pitch frequency detection unit 413b are set to phr [i] [j]. Set to st.F0 and phr [i] [j] .ed.F0.

  In the present embodiment, the pitch frequency at the end is used as the acoustic feature quantity. However, the average pitch of all frames (that is, frames including voiced sound) in which the pitch frequency can be detected in the speech waveform of each speech component. The frequency may be calculated and used as an acoustic feature amount. In particular, an average pitch frequency may be used in the case of an audio component with a short data length, and an end pitch frequency is preferably used when the data length is long. Thereby, it is possible to appropriately set the pause time in the connection part of the audio component data.

Next, the configuration of the speech speed detection unit 414 will be described with reference to FIGS. Here, FIG. 8 is an explanatory diagram for explaining how the pause time length is set based on the speech speed.
As shown in FIG. 6, the speech speed detection unit 414 includes an end silent length detection unit 414a and an average speech speed detection unit 414b.

In the present embodiment, the case of using the speech speed as the acoustic feature quantity, the average speech speed in the speech segment speech waveform of the preceding audio component data P _A, the average talk in the speech section of the subsequent audio component data P _B of the speech waveform Based on the speed, the pause time length (phr [i] [j] .pau) is calculated.
In the present embodiment, as shown in FIG. 8, the average speech speed calculated based on the number of beats (phr [i] [j] .mora) appearing in the speech segment and the speech segment length as the speech rate. However, the speech speed based on other definitions such as the number of phonemes per unit time may be used.

  The end silent length detection unit 414a detects whether or not the speech waveform signal has a power greater than or equal to a predetermined value for each frame by analyzing the spectrum data of the speech waveform that has been framed. First, the position of a frame having a power equal to or greater than a predetermined value is detected as the voice start position. Further, the position of the last frame having a predetermined power is detected as the voice end position.

Then, based on the detected voice start position and data start position, the leading end silent length can be calculated. Further, the trailing end silent length can be calculated based on the detected voice end position and data end position.
In addition, the voice interval length is determined from the data length (phr [i] [j] .time) to the leading silence length (phr [i] [j] .st.pos1) and trailing silence length (phr [i] [j] .ed, pos1) can be calculated by subtracting.

  Note that the voice start position and voice end position may be detected by detecting whether or not each frame has a phoneme instead of power, for example, or voice waveform data instead of spectrum data. Detection may be performed based on the signal level.

The average speech speed detection unit 414b calculates the number of beats (phr [i] [j] .mora) preset in the voice component data by the above procedure as shown in the equation (1). The average speech speed phr [i] [j] .SR is calculated by dividing by the section length.
phr [i] [j] .SR =
phr [i] [j] .mora / (phr [i] [j] .time -phr [i] [j] .st.pos1 -phr [i] [j] .ed.pos1)
... (1)

  The speech speed detection unit 414 converts the leading end silent length and trailing end silent length detected by the end silent length detection unit 414a into phr [i] [j] .st.pos1 and phr [i] [ j] .ed.pos1 and the average speech speed detected by the average speech speed detection unit 414b is set in phr [i] [j] .SR of the voice component data.

  In the present embodiment, the average speech speed of the voice section is used as the acoustic feature amount. However, when the data length is short, the silent length at the front end and the silent length at the rear end are ignored, and the number of beats is set as the data length. The average speech speed calculated by dividing may be used. In this case, it is not necessary to detect the silence length.

  In addition, when the data length is long, it may be detected and used instead of the average speech speed of the speech waveform. When using the speech rate at the end, for example, by analyzing the framed speech waveform, the duration of the first and last mora appearing is detected, and the reciprocal of the duration length is set to the tip speech rate, respectively. And it can be used as the rear-end speech speed. Or you may make it detect the appearance number of mora in the predetermined time from the front-end | tip and a rear end.

Next, the configuration of the power detection unit 415 will be described with reference to FIGS. 6 and 9. Here, FIG. 9 is an explanatory diagram for explaining the setting of the pause time length based on the power.
As shown in FIG. 6, the power detection unit 415 includes an end silent length detection unit 415a and an average power detection unit 415b.

In the present embodiment, as shown in FIG. 9, the case of using power as the acoustic feature quantity, and the average power in the speech section of the speech waveform of the preceding audio component data P _A, subsequent audio component data P _B of the speech waveform The pause time length (phr [i] [j] .pau) is calculated based on the average power in the voice section.

The end silent length detection unit 415a detects the voice start position and the voice end position and calculates the leading end silent length and the trailing end silent length in the same manner as the end silent length detection unit 414a of the speech speed detection unit 414. Therefore, detailed description is omitted. For example, the end silent length detection unit 414a may be shared by the power detection unit 415, the speech speed detection unit 414, and the spectrum envelope detection unit 416 described later.
Further, the voice section length can be calculated by subtracting the leading end silent length and the trailing end silent length from the data length.

The average power detection unit 415b detects the power (phr [i] [j] .pwr [k]: k indicates the frame number) for each frame using the spectrum data, as shown in the equation (2). The average power (phr [i] [j] .PW) is calculated by averaging the powers of all frames in the speech section.
phr [i] [j] .PW = sum (phr [i] [j] .pwk [k]) / number of frames in speech interval (2)
However, the sum () of the denominator on the right side indicates the total power of the frames in the speech section.

  The power detection unit 415 detects the leading end silent length and the trailing end silent length detected by the end silent length detection unit 415a, respectively, as phr [i] [j] .st.pos1 and phr [i] [j ] .ed.pos1 and the average power detected by the average power detector 415b is set in phr [i] [j] .PW of the audio component data.

  In the present embodiment, the average power of the voice section is used as the acoustic feature amount. However, instead of the average power of the voice section, the power of the front and rear end frames of the voice section may be detected and used. . Further, in this embodiment, power is used as an acoustic feature amount representing the loudness of the sound. However, instead of power, for example, a loudness level that is a sense amount representing an auditory volume may be used. .

Next, the configuration of the spectrum envelope detection unit 416 will be described with reference to FIGS. 6 and 10. Here, FIG. 10 is an explanatory diagram for explaining the setting of the pause time length based on the spectrum envelope.
As shown in FIG. 6, the spectrum envelope detection unit 416 includes an end silence length detection unit 416a and an end spectrum envelope detection unit 416b.

In the present embodiment, as shown in FIG. 10, in the case of using the spectral envelope as the acoustic feature quantity, rear spectral envelope at the rear end of the speech waveform of the preceding audio component data _{P A (phr [i] [} j] ed.SE) and the leading edge spectral envelope (phr [i] [j + 1] .st.SE) at the front end of the speech waveform of the subsequent speech component data P _B , the pause time length (phr [i] [j] .pau).

The spectrum envelope can be obtained based on the spectrum data calculated by the spectrum analysis unit 412. For example, when Fourier transform is used as a spectrum analysis method, a Fourier transform coefficient can be used. In addition, a band filter group, a correlation function, a coefficient of LPC analysis, a cepstrum, a mel cepstrum, and the like can be used. Furthermore, you may make it add dynamic feature-values, such as the primary differentiation of these coefficients, and a secondary differentiation.
The spectrum envelope is expressed as a vector quantity composed of a plurality of coefficients.

The end silence length detection unit 416a calculates the front end silence length and the rear end silence length by detecting the voice start position and the voice end position, similarly to the end silence length detection unit 414a of the speech speed detection unit 414. Therefore, detailed description is omitted. Note that, for example, the end soundless length detection unit 414a may be shared by the speech speed detection unit 414, the power detection unit 415, and the spectrum envelope detection unit 416.
Further, the voice section length can be calculated by subtracting the leading end silent length and the trailing end silent length from the data length.

  The end spectrum envelope detection unit 416b detects the spectrum envelope from the spectrum data corresponding to the frame at the voice start position detected by the end silence length detection unit 416a to obtain the tip spectrum envelope, and the spectrum data of the frame at the voice end position From this, a spectral envelope is detected and used as a rear end spectral envelope.

  The spectrum envelope detection unit 416 converts the leading end silent length and trailing end silent length detected by the end silent length detection unit 416a into phr [i] [j] .st.pos1 and phr [i] [ j] .ed.pos1 and the front-end spectrum envelope and the rear-end spectrum envelope detected by the edge spectrum envelope detection unit 416b are respectively set to phr [i] [j] .st.SE and phr [ Set to i] [j] .ed.SE.

  In the present embodiment, the spectral envelope at the end is used as the acoustic feature quantity. However, when the data length of the voice component is short, the average spectral envelope in the voice section may be used.

Returning to FIG. 6, the description of the configuration of the pause time length calculation device 40 will be continued.
The preceding audio component data storage unit 420 temporarily stores the audio component data in which the data related to the acoustic feature amount is set by the acoustic feature amount detection unit 410, and the audio component data calculates the next pause time length. In this case, the pause time length calculation unit 440 reads the previous voice component data. That is, the preceding audio component data storage unit 420 functions as a data delay unit.
As the preceding audio component data storage unit 420, for example, a semiconductor memory can be used, but a storage device such as a magnetic disk device or an optical disk device can also be used.

  The acoustic distance calculation unit (acoustic distance calculation unit) 430 inputs the audio component data in which the data related to the acoustic feature amount is set by the acoustic feature amount detection unit 410 as the subsequent audio component data, and the preceding audio component. The audio component data stored in the data storage unit 420 is read out and used as preceding audio component data. Then, the acoustic distance is calculated based on the data related to the acoustic feature amount set in the preceding speech component data and the data related to the acoustic feature amount set in the subsequent speech component data, and the pause time length calculation unit 440 Output.

The acoustic distance can be calculated by Expressions (3) to (8) according to the acoustic feature amount to be used.
First, when the pitch frequency is used as the acoustic feature amount, the rear end pitch frequency (phr [i] [j] .ed.F0) and the front end pitch of the subsequent audio component data are obtained by Expression (3). The acoustic distance (ΔF0 [i] [j]) is calculated based on the frequency (phr [i] [j + 1] .st.F0), and the rear end of the preceding audio component data is calculated by Equation (4). Based on the unvoiced sound length (phr [i] [j] .ed.pos2) and the leading unvoiced sound length (phr [i] [j + 1] .st.pos2) The temporal distance (ΔFp [i] [j]) between the ends where the pitch frequency used for the distance calculation is detected is calculated.

ΔF0 [i] [j] = | log (phr [i] [j] .ed.F0) -log (phr [i] [j + 1] .st.F0) | (3)
ΔFp [i] [j] = phr [i] [j] .ed.pos2 + phr [i] [j + 1] .st.pos2 (4)
Where log () represents a common logarithmic function.

  By detecting the pause time length in consideration of the temporal distance between the ends, the pause time length can be calculated more appropriately than when the pause time length is not taken into consideration.

  Next, when speech speed is used as the acoustic feature amount, the average speech speed (phr [i] [j] .SR) of the preceding speech component data and the average speech speed of the subsequent speech component data (equation (5)) phr [i] [j + 1] .SR) to calculate the acoustic distance (ΔR [i] [j]).

  ΔR [i] [j] = | phr [i] [j] .SR -phr [i] [j + 1] .SR | (5)

  Next, when power is used as the acoustic feature quantity, the average power (phr [i] [j] .PW) of the speech section of the preceding speech component data and the speech section of the succeeding speech component data are expressed by Equation (6). The acoustic distance (ΔP [i] [j]) is calculated based on the average power (phr [i] [j + 1] .PW).

  ΔP [i] [j] = | phr [i] [j] .PW -phr [i] [j + 1] .PW | (6)

  Next, when a spectral envelope is used as the acoustic feature quantity, the trailing edge spectral envelope (phr [i] [j] .ed.SE) and the leading edge of the following voice component data are obtained by Expression (7). Calculate the acoustic distance (ΔE [i] [j]) based on the spectral envelope (phr [i] [j + 1] .st.SE) and Calculate the acoustic distance based on the end silence length (phr [i] [j] .ed.pos1) and the end silence length of the following audio component data (phr [i] [j + 1] .st.pos1) The time distance (ΔEp [i] [j]) between the end portions where the spectrum envelope used in the above is detected is calculated.

ΔE [i] [j] = | phr [i] [j] .ed.SE -phr [i] [j + 1] .st.SE | (7)
ΔEp [i] [j] = phr [i] [j] .ed.pos1 + phr [i] [j + 1] .st.pos1 (8)
However, since the spectral envelope is a vector quantity, the Euclidean distance between the vector quantities is calculated as the spectral envelope distance in Equation (7).

  As in the case of using the pitch frequency, by detecting the pause time length in consideration of the temporal distance between the end portions, the pause time length can be calculated more appropriately than when the pitch frequency is not taken into consideration.

  The pause time length calculation unit (pause time length calculation means) 440 includes a multiple regression calculation unit 441 and a regression coefficient storage unit 442, and is based on the acoustic distance calculated by the acoustic distance calculation unit 430. The pause time length to be inserted into the connection between the preceding voice component data and the subsequent voice component data is calculated and output to the pause time length setting unit 450.

The multiple regression calculation unit 441 performs a resting time by performing a regression calculation based on the acoustic distance calculated by the acoustic distance calculation unit 430 and the regression equation coefficient stored in advance in the regression coefficient storage unit 442. The length is calculated and output to the pause time length setting unit 450.
In the embodiment shown in FIG. 6, the pause time length is calculated by a multiple regression equation using a plurality of acoustic feature quantities as explanatory variables. However, when one acoustic feature quantity is used as an explanatory variable, The pause time length is calculated by a single regression equation. The regression equation in the claims includes a multiple regression equation when there are a plurality of explanatory variables and a single regression equation when there is one explanatory variable.

  Here, the pause time length (phr [i] [j] .pau) is shown in Expressions (9) to (13) using regression coefficients a0 to a16 and the like according to the acoustic feature amount used. Calculated by multiple regression equation.

First, when only the pitch frequency is used as the acoustic feature amount, the multiple regression equation shown in Equation (9) is used.
phr [i] [j] .pau = a0 + a1 × ΔF0 [i] [j] + a2 × ΔFp [i] [j] (9)

Next, when only the speech speed is used as the acoustic feature amount, the single regression equation shown in Equation (10) is used.
phr [i] [j] .pau = a3 + a4 × ΔR [i] [j] (10)

Next, when only power is used as the acoustic feature amount, the single regression equation shown in Equation (11) is used.
phr [i] [j] .pau = a5 + a6 × ΔP [i] [j] (11)

Next, when only the spectral envelope is used as the acoustic feature amount, the multiple regression equation shown in Equation (12) is used.
phr [i] [j] .pau = a7 + a8 × ΔE [i] [j] + a9 × ΔEp [i] [j] (12)

In addition, when using four of the pitch frequency, speech speed, power, and spectrum envelope as acoustic feature quantities, the multiple regression equation shown in Equation (13) is used.
phr [i] [j] .pau = a10 + a11 × ΔF0 [i] [j] + a12 × ΔFp [i] [j] + a13 × ΔR [i] [j] + a14 × ΔP [i] [j ] + a15 × ΔE [i] [j] + a16 × ΔEp [i] [j] (13)

  Further, the present invention is not limited to the above regression formula, and the acoustic feature amount to be used may be appropriately combined to determine the multiple regression formula to calculate the pause time length.

  Here, with reference to FIG. 11, a method for obtaining the coefficients a0 to a16 of the regression equation will be described by taking a case where a pitch frequency is used as an acoustic feature amount as an example. Here, FIG. 11 is an explanatory diagram for explaining the relationship between the subjective evaluation experiment and the multiple regression analysis.

  In order to determine the coefficient of the multiple regression equation, first, an optimum pause time length for each combination of audio component data when the audio component data is combined in various ways is obtained by a subjective evaluation experiment. On the other hand, the acoustic distance and the temporal distance are calculated in the same procedure as the acoustic distance calculation unit 430 described above. The coefficient of the multiple regression equation can be determined by performing multiple regression analysis of the optimal pause time length obtained by the subjective evaluation experiment, the acoustic distance and the temporal distance obtained by calculation.

For example, as shown in FIG. 11, the optimal quiescence period Pause1 between subsequent audio component data P _B corresponding to the preceding audio component data P _A and clause 1b corresponding to clause 1a determined by subjective evaluation experiment. Similarly, the subsequent speech component data P _B optimal dwell time length between Pause2, preceding audio component data P _A and clauses which correspond to the clauses La corresponding to the preceding audio component data P _A and clause 2b corresponding to clause 2a The optimum pause time length PauseL and the like with the subsequent audio component data P _B corresponding to Lb is obtained by a subjective evaluation experiment.

  In the subjective evaluation experiment, the optimum pause time length can be quantified by, for example, the relative method or the limit method. The pause time length can also be obtained using a subjective evaluation experiment by another method.

The acoustic distance in combination with the respective preceding voice component data _{P A} and the subsequent audio component data _{P B (ΔF0 1, ΔF0 2} , ΔF0 L , etc.) and the temporal distance _{(ΔFp 1, ΔFp 2, ΔFp} L , etc.) Is calculated according to the procedure described above.

When these data are applied to equation (9), a relational expression such as equation (14) is obtained.
Pause1 = a0 + a1 × ΔF0 ₁ + a2 × ΔFp ₁
Pause1 = a0 + a1 × ΔF0 ₂ + a2 × ΔFp ₂
・
・
・
Pause1 = a0 + a1 × ΔF0 _L + a2 × ΔFp _L
(14)

By applying the least square method to the relational expression shown in Expression (14), the coefficients a0, a1, and a2 of the regression equation can be calculated and determined.
Even when other regression equations such as those shown in equations (10) to (13) are used, the coefficients of the regression equation can be determined by the same procedure.

  The regression equation coefficient determined in this way is stored in the regression coefficient storage unit 442 (see FIG. 6), and the regression equation coefficient is read from the regression coefficient storage unit 442 and used. Can be calculated.

The regression coefficient storage unit 442 stores a regression equation coefficient determined in advance based on the subjective evaluation experiment described above, and the stored coefficient is appropriately read by the multiple regression calculation unit 441.
As the regression coefficient storage unit 442, for example, a storage device such as a magnetic disk device, an optical disk device, or a semiconductor memory can be used.

  The pause time length setting unit 450 sets the pause time length calculated by the pause time length calculation unit 440 to the pause time length (phr [i] [j] .pau of the preceding voice component data when the pause time length is calculated. ) And the speech part data for which the pause time length is set are stored in the speech synthesis data storage unit 50 (see FIG. 1) in association with the sentence number and the phrase number.

  The speech synthesizer 100 described above can be implemented by creating a dedicated hardware part or all of the speech synthesizer 100. However, a general computer program is executed, and an arithmetic device, a storage device, an input device, It can also be realized by operating an image display device or the like. This program (pause time length calculation program) can be distributed via a communication line, or can be distributed on a recording medium such as a CD-ROM.

<Operation of speech synthesizer>
Next, the operation of the speech synthesizer 100 of this embodiment will be described with reference to FIG. 12 (refer to FIGS. 1 and 6 as appropriate). Here, FIG. 12 is a flowchart showing the flow of processing of the speech synthesizer of this embodiment.

  First, the speech synthesizer 100 inputs read-out information to be subjected to speech synthesis by the read-out information input unit 10 and outputs the input read-out information to the voice component data acquisition unit 20 (step S10).

  The speech synthesizer 100 uses the speech component data acquisition unit 20 to sequentially acquire speech component data corresponding to the phrase specified in the reading information input in step S10 from the speech component data storage unit 30, and calculate a pause time length. The result is output to the acoustic feature quantity detection unit 410 of the device 40 (step S11).

  The speech synthesizer 100 detects an acoustic feature amount by the acoustic feature amount detection unit 410 of the pause time length calculation device 40, sets data relating to the detected acoustic feature amount as speech component data, and this speech component data Is output to the acoustic distance calculation unit 430 as subsequent audio component data, and is stored in the preceding audio component data storage unit 420 as preceding audio component data in the connection of the next audio component data (step S12).

  Here, when the audio component data output by the acoustic feature quantity detection unit 410 is the audio component data for the first phrase constituting the reading information input in step S10 (Yes in step S13), this audio component Since there is no preceding audio component data corresponding to the data and there is no need to calculate the pause time length, the process returns to step S11 to acquire audio component data corresponding to the next phrase.

  On the other hand, when the audio component data output by the acoustic feature amount detection unit 410 is not the audio component data for the first phrase constituting the reading information (No in step S13), the speech synthesizer 100 determines the acoustic distance. The calculation unit 430 sets the acoustic feature amount set in the preceding audio component data stored in the preceding audio component data storage unit 420 and the subsequent audio component data output by the acoustic feature amount detection unit 410. The acoustic distance with the acoustic feature amount being calculated is calculated and output to the pause time length calculation unit 440 (step S14).

  Next, an acoustic distance between the multiple regression equation coefficient stored in advance in the regression coefficient storage unit 442 by the multiple regression calculation unit 441 of the pause time length calculation unit 440 and the audio component data calculated in step S14. Based on the above, the suspension time length is calculated and output to the suspension time length setting unit 450 (step S15).

  Then, the pause time length setting unit 450 sets the pause time length calculated in step S15 in the preceding voice part data (step S16), and associates it with the sentence number and phrase number of the reading information, and the voice synthesis data storage unit 50 (step S17).

  When the voice part data in which the pause time length is set in step S17 is stored, it is checked whether or not the next phrase remains in the reading information (step S18). If the phrase remains (Yes in step S18), step Returning to S11, the audio component data corresponding to the next phrase is acquired, and the processing up to step S17 is repeated.

On the other hand, if the next phrase does not remain (No in step S18), it is not necessary to set a pause time length for the subsequent audio part data corresponding to the last phrase, so this subsequent audio part data is read-out information. Is stored in the speech synthesis data storage unit 50 in association with the number of the last phrase in the last sentence (step S19).
Thus, the speech synthesis data for the read-out information is completed in the speech synthesis data storage unit 50.

  When the speech synthesis data is completed, the speech reproduction unit 60 sequentially reads out the corresponding speech component data from the speech synthesis data storage unit 50 according to the sentence number and the phrase number, and the speech waveform data included in the speech component data is obtained. It is converted into an analog voice waveform signal, outputted to the speaker 70, and reproduced so as to be audible. Then, after inserting a pause (silent state) of the pause time length set in the audio component data, the next audio component data is reproduced (step S20).

  By the procedure described above, the speech synthesizer 100 can reproduce the sound with a natural impression by inserting an appropriate pause time between the speech component data designated by the reading information.

  In the present embodiment, the acoustic analysis of the speech waveform data of the speech component data and the calculation of the pause time length are successively processed in the present embodiment, but the speech component data corresponding to all the phrases included in the reading-out information is processed. After the acoustic waveform data is acoustically analyzed, the pause time length may be calculated.

  In order to do this, for example, the acoustic feature quantity detection unit 410 performs acoustic analysis from the voice waveform data of the voice part data, sets data relating to the acoustic feature quantity in the voice part data, for example, voice synthesis data Store in the storage unit 50. When the acoustic analysis of the speech component data corresponding to all the clauses included in the reading information is completed, the acoustic distance calculation unit 430 sequentially reads the preceding speech component data and the subsequent speech component data pairs from the speech synthesis data storage unit 50. An acoustic distance is calculated, a pause time length is calculated by the pause time length calculation unit 440 based on the calculated acoustic distance, and a pause time length is set in the preceding voice component data by the pause time length setting unit 450 The data is stored in the voice synthesis data storage unit 50. When the setting of the pause time length between all the voice component data is completed, the voice synthesis data storage unit 50 completes the voice synthesis data constituted by the voice component data for which the pause time length is set.

In the present embodiment, the acoustic feature amount is acquired by performing acoustic analysis on the acoustic waveform data preset in the voice component data by the acoustic feature amount detection unit 410. The audio waveform data may be acoustically analyzed to detect data relating to the acoustic feature amount shown in FIG. 3, set as audio component data, stored in the audio component data storage unit 30, and used.
As a result, it is not necessary to detect the acoustic feature quantity of the voice component data selected by the voice synthesis every time the voice synthesis is performed, and the acoustic feature quantity can be obtained simply by referring to the data related to the acoustic feature quantity set in the voice component data. The feature amount can be acquired, and the processing time required for the speech synthesis process can be shortened.

It is a block diagram which shows the structure of the speech synthesizer of this embodiment. It is explanatory drawing for demonstrating the structure of reading-out information. It is a figure which shows the data structure of audio | voice component data. It is a block diagram which shows typically the structure of the audio | voice waveform data contained in audio | voice component data. It is explanatory drawing for demonstrating the principle which sets the idle time length by this invention. It is a block diagram which shows the structure of the idle time length calculation apparatus of this embodiment. It is explanatory drawing for demonstrating the mode of the setting of the pause time length based on a pitch frequency. It is explanatory drawing for demonstrating the mode of the setting of the idle time length based on speech speed. It is explanatory drawing for demonstrating the mode of the setting of the idle time length based on power. It is explanatory drawing for demonstrating the mode of the setting of the idle time length based on a spectrum envelope. It is explanatory drawing for demonstrating the relationship between a subjective evaluation experiment and multiple regression analysis. It is a flowchart which shows the flow of a process of the speech synthesizer of this embodiment.

Explanation of symbols

10 Reading information input part (Reading information acquisition means)
20 Audio component data acquisition unit (audio component data acquisition means)
30 Voice component data storage unit (voice component data storage means)
40 pause time length calculation device 100 speech synthesizer 410 acoustic feature quantity detector (acoustic feature quantity acquisition means)
413 Pitch frequency detection unit 414 Speech speed detection unit 415 Power detection unit 416 Spectrum envelope detection unit 430 Acoustic distance calculation unit (acoustic distance calculation means)
440 Pause time length calculation unit (pause time length calculation means)
441 450 rest time length setting unit Multiple Regression calculation unit P _A preceding speech component data P _B succeeding audio component data

Claims (4)

A pause time length calculation device that calculates a pause time length to be inserted between voice component data connected to each other when voice component data recording a voice waveform in which a predetermined unit of text is uttered is connected and voice synthesis is performed. There,
At least one acoustic feature amount of pitch frequency, speech speed, power, or spectrum envelope in the speech waveform recorded in the preceding preceding speech component data and the succeeding subsequent speech component data in the speech component data connected to each other. Acoustic feature acquisition means for acquiring each ;
The sound that is the difference between the acoustic feature amount of the preceding audio component data that precedes the audio component data connected to each other and the acoustic feature amount of the subsequent audio component data that is acquired by the acoustic feature amount acquisition unit. An acoustic distance calculating means for calculating a target distance;
Based on the acoustic distance calculated by the acoustic distance acquisition means, a pause time for calculating a pause time length to be inserted between the preceding voice component data and the subsequent voice component data using a preset calculation formula A length calculating means;
An idle time length calculation device comprising: 2. The pause time length calculation apparatus according to claim 1, wherein a regression equation using the acoustic distance calculated by the acoustic distance calculation means as an explanatory variable is used as the calculation formula. In order to calculate a pause time length to be inserted between speech component data connected to each other when speech synthesis is performed by connecting speech component data in which a speech waveform uttering a predetermined unit of text is connected,
At least one acoustic feature amount of pitch frequency, speech speed, power, or spectrum envelope in the speech waveform recorded in the preceding preceding speech component data and the succeeding subsequent speech component data in the speech component data connected to each other. Acoustic feature acquisition means for acquiring each ;
The sound that is the difference between the acoustic feature amount of the preceding audio component data that precedes the audio component data connected to each other and the acoustic feature amount of the subsequent audio component data that is acquired by the acoustic feature amount acquisition unit. Acoustic distance calculating means for calculating the target distance;
Based on the acoustic distance calculated by the acoustic distance acquisition means, a pause time for calculating a pause time length to be inserted between the preceding voice component data and the subsequent voice component data using a preset calculation formula Length calculation means,
An idle time length calculation program characterized in that it functions as: A speech synthesizer that synthesizes speech by connecting speech component data that records speech waveforms that utter a predetermined unit of text,
Voice component data storage means for storing voice component data in which a voice waveform is recorded in advance;
Read-out information acquisition means for acquiring read-out information consisting of text that is continuously read out in a determined order, or information that specifies the audio component data corresponding to the predetermined unit of text constituting the text;
Voice component data acquisition means for acquiring voice component data from the voice component data storage means based on the reading information acquired by the reading information acquisition means;
The pause time length calculation device according to claim 1 or 2 , which calculates a pause time length to be inserted between voice component data constituting the reading information acquired by the voice component data acquisition unit,
A speech synthesizer, wherein the pause time length calculated by the pause time length calculation device is set as a pause time length between the speech component data.

Images