JP2009258498A - Speech synthesis device and speech synthesis method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To create a synthesized speech of high quality with a small storage capacity. <P>SOLUTION: The speech synthesis device comprises: a preceding section length determination part 2 which stores in advance each correspondence between phonological symbol and precedence section length, and determines a preceding section length corresponding to a phonological symbol obtained from an input text according to the correspondence; and a compressed speech waveform reading part 3 which reads, from compressed speech waveforms stored in a speech fragment dictionary 1, a compressed speech waveform including a speech waveform of a speech fragment corresponding to the phonological symbol obtained from the input text, with a speech waveform of a section length corresponding to the preceding section length determined by the determination part 2 added to the front of the compressed speech waveform. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a speech synthesizer and a speech synthesis method that are mounted on, for example, a car navigation system or a mobile phone and artificially generate speech signals from arbitrary sentences.

In text-to-speech synthesis technology that artificially generates speech signals from arbitrary sentences, phonetic symbols and prosodic information (for example, pitch, phoneme duration length) obtained by performing language analysis processing and prosody generation processing on input text Etc.) and a speech signal is generated from the phonetic symbols and prosodic information.
That is, in the text-to-speech synthesis technology, vowels are represented by “V”, consonants are represented by “C”, and for example, basic unit characteristic parameters such as “CV”, “CVC”, and “VCV” are expressed as speech units. When the phoneme symbol and the prosodic information are input, the phoneme unit corresponding to the phoneme symbol is selectively read out, the pitch and the phoneme duration are controlled according to the prosodic information, and the phoneme unit is read out. Are sequentially connected to synthesize voice.

In a conventional speech synthesizer, in order to hold a speech unit with a small storage capacity, compressed data of the speech unit is held.
However, if the speech segment is compressed by a compression method with a high compression rate, the recording capacity can be reduced, but the distortion at the beginning of the speech section increases, and the overall distortion tends to increase.
Such distortion leads to deterioration of the quality of the synthesized speech, and therefore there is a problem that the compression rate of the speech unit cannot be increased so much.

Therefore, when extracting and compressing the speech waveform of a desired speech segment from speech waveforms such as single sounds, words, and single sentences previously uttered by humans, the speech in the front section of the speech waveform of the speech segment is compressed. When the speech waveform of a speech unit is expanded by compressing the waveform including the waveform, the speech waveform in front of the speech unit of the speech unit is first expanded and read and discarded. A technique for reducing the distortion in the section has been developed (see, for example, Patent Document 1).
Note that in the speech synthesizer disclosed in Patent Document 1 below, the length of the speech segment to be expanded prior to the expansion of the speech waveform of the speech unit is dynamically determined by the compression distortion of the speech unit. In order to do this, information indicating the length of the voice section to be expanded in advance is stored separately as a pair with the voice segment.

JP 2002-287784 A (paragraph number [0058])

  Since the conventional speech synthesizer is configured as described above, the distortion in the speech segment section can be reduced, but it is necessary to store information indicating the length of the speech section to be expanded in advance. is there. For this reason, there is a problem that the compression effect of the storage capacity cannot be obtained sufficiently.

  The present invention has been made to solve the above-described problems, and an object of the present invention is to provide a speech synthesizer and a speech synthesis method that can generate high-quality synthesized speech with a small storage capacity.

  The speech synthesizer according to the present invention includes a speech unit dictionary storing a compressed speech waveform in which a speech waveform including a section preceding a speech unit section is compressed as a speech unit speech waveform, and an input A preceding section length determining means for determining a preceding section length corresponding to a phoneme symbol obtained from a text; and a phoneme corresponding to a phoneme symbol obtained from an input text out of compressed speech waveforms stored in a speech segment dictionary. Compressed voice waveform reading means for reading out a compressed voice waveform including a single voice waveform and having a voice waveform having a section length matching the preceding section length determined by the preceding section length determination means in front of the voice waveform; Provided, the voice waveform extracting means expands the compressed voice waveform read by the compressed voice waveform reading means, and extracts the voice waveform of the voice segment from the voice waveform after expansion. That.

  According to the present invention, a speech unit dictionary storing a compressed speech waveform in which a speech waveform including a section preceding a speech unit section is compressed as a speech unit speech waveform is obtained from an input text. Voice of speech segments corresponding to phoneme symbols obtained from input text out of compressed speech waveforms stored in the speech segment dictionary, and preceding segment length determination means for determining preceding segment lengths corresponding to the phoneme symbols A compressed speech waveform reading means for reading a compressed speech waveform including a waveform and having a speech waveform having a section length matching the preceding section length determined by the preceding section length determining means in front of the speech waveform; Since the waveform extraction unit is configured to expand the compressed speech waveform read by the compressed speech waveform reading unit and extract the speech waveform of the speech segment from the expanded speech waveform, By volume, the effect that can generate synthesized speech of high quality.

Embodiment 1 FIG.
FIG. 1 is a block diagram showing a speech synthesizer according to Embodiment 1 of the present invention. In FIG. 1, a speech segment dictionary 1 is a speech waveform of a speech segment. Stores a compressed speech waveform in which the speech waveform is compressed. Therefore, when a part of a word or simple sentence is used as a speech unit, the speech waveform is compressed and stored, including a section preceding the section of the speech unit.
For example, as shown in FIG. 3, when the CV segment “ta” is used from the compressed speech waveform uttered “Atama”, the speech segment dictionary 1 including the section preceding “ta” is used. Stored.

The preceding section length determination unit 2 stores a correspondence relationship between the phoneme symbol and the preceding section length in advance, and performs processing for determining the preceding section length corresponding to the phoneme symbol obtained from the input text with reference to the correspondence relationship. To do. The preceding section length determining unit 2 constitutes a preceding section length determining means.
The compressed speech waveform reading unit 3 includes a speech waveform of a speech unit corresponding to a phoneme symbol obtained from the input text from among the compressed speech waveforms stored in the speech unit dictionary 1, and precedes the speech waveform. A process of reading a compressed speech waveform to which a speech waveform having a section length that matches the preceding section length determined by the section length determination unit 2 is added. The compressed voice waveform reading unit 3 constitutes compressed voice waveform reading means.

The speech waveform extraction unit 4 expands the compressed speech waveform read by the compressed speech waveform reading unit 3 and performs a process of extracting the speech waveform of the speech unit from the expanded speech waveform. The speech waveform extraction unit 4 constitutes speech waveform extraction means.
The speech generation unit 5 controls the pitch and phoneme duration in accordance with the prosodic information obtained from the input text, while the speech waveform of the speech unit extracted by the speech waveform extraction unit 4 in accordance with the phoneme symbol obtained from the input text. Are sequentially connected to generate synthesized speech. The voice generation unit 5 constitutes a voice generation unit.
FIG. 2 is a flowchart showing the processing contents of the speech synthesizer according to Embodiment 1 of the present invention.

Next, the operation will be described.
The phonological symbols and prosodic information input to the speech synthesizer are information such as phonological symbols, pitches, phonological durations, and powers. For example, language analysis processing and prosodic generation processing are performed on the input text. Can be obtained.
A plurality of phonetic symbols and prosodic information obtained from the input text are sequentially input to the speech synthesizer.

The speech unit dictionary 1 stores the speech waveform of the speech unit, but the speech waveform of the speech unit is compressed and stored in order to reduce the storage capacity.
However, when a part of a word or simple sentence is used as a speech unit, the speech waveform is compressed and stored, including the section preceding the section of the speech unit.
For example, as shown in FIG. 3, when the CV segment “ta” is used from the compressed speech waveform uttered “Atama”, the speech segment dictionary 1 including the section preceding “ta” is used. Stored.

The preceding section length determination unit 2 stores a correspondence relationship between the phoneme symbol and the preceding section length in advance, and when a phoneme symbol obtained from the input text is input, the preceding section length corresponding to the phoneme symbol is referred to by referring to the correspondence relationship. The section length is determined (step ST1).
Specifically, for example, the preceding interval length for a phoneme symbol of a speech unit (eg, “CV”) starting with a consonant is “Nc”, and a phoneme symbol of a phoneme unit (eg, “VC”) starting with a vowel The preceding section length for is set as “Nv”, and such a correspondence is stored in the preceding section length determination unit 2.
If the phonological symbol obtained from the input text is a phonological symbol of a phoneme unit starting with a consonant, the preceding section length determining unit 2 refers to the above correspondence relationship, and the preceding interval length determining unit 2 corresponds to the phonological symbol obtained from the input text. The section length is determined as “Nc”.
On the other hand, if the phoneme symbol obtained from the input text is a phoneme symbol of a speech unit starting with a vowel, the preceding section length corresponding to the phoneme symbol obtained from the input text is determined as “Nv”.

Here, “Nc” and “Nv” are set, for example, such that the distortion of the segment of the speech segment is sufficiently reduced.
Alternatively, it is assumed that the deterioration feeling when the user listens is set to be sufficiently small.
The consonant (C) has a characteristic that the power of the voice signal is smaller than that of the vowel (V) on average, and even when distortion is large in the consonant section, the deterioration feeling when the user listens is smaller than that in the vowel section. It may be set so that “Nc” <“Nv”.
As a result, the preceding section length for the CV segment (the section length to be stored in the speech segment dictionary 1) can be shortened, which has the effect of reducing the storage capacity. In addition, since the length of the section in which the speech waveform of the CV segment is expanded is shortened, the processing amount required for expansion can be reduced.
Thus, since the preceding section length determination unit 2 determines the preceding section length that should be uniquely traced by the phoneme symbol, it is not necessary to store information on the preceding section length for each speech unit, and the storage capacity Can be suppressed.

When the preceding section length determination unit 2 determines the preceding section length, the compressed speech waveform reading unit 3 determines the speech corresponding to the phoneme symbol obtained from the input text from among the compressed speech waveforms stored in the speech unit dictionary 1. A process of reading a compressed speech waveform that includes a speech waveform of a segment and in which a speech waveform having a section length that matches the preceding section length determined by the preceding section length determination unit 2 is added in front of the speech waveform. (Step ST2).
For example, if the speech unit corresponding to the phoneme symbol obtained from the input text is “am” and the preceding segment length determined by the preceding segment length determining unit 2 is “Nv”, the speech unit dictionary 1 stores it. Among the compressed speech waveforms that have been recorded, a compressed speech waveform that includes the speech waveform of the speech unit “am” and that has a speech waveform with a section length that matches “Nv” in front of the speech waveform is read out. Like that.

When the compressed speech waveform reading unit 3 reads the compressed speech waveform, the speech waveform extracting unit 4 expands the compressed speech waveform (step ST3). That is, the compressed speech waveform of the preceding section + speech segment section is expanded.
Next, the speech waveform extraction unit 4 extracts the speech waveform of the speech unit (for example, the speech waveform of “am”) from the decompressed speech waveform (step ST4).

When the speech waveform extraction unit 4 extracts the speech waveform of the speech unit, the speech generation unit 5 controls the pitch and phoneme duration in accordance with the prosodic information obtained from the input text, and the phoneme symbols obtained from the input text. Accordingly, the speech waveforms of the speech units are sequentially connected to generate synthesized speech (step ST5).
The processes of steps ST1 to ST5 are repeated until the input of phoneme symbols / prosodic information is completed (step ST6).

  As apparent from the above, according to the first embodiment, a compressed speech waveform in which a speech waveform including a section preceding the speech unit section is compressed is stored as the speech waveform of the speech unit. The speech segment dictionary 1, the preceding segment length determination unit 2 that determines the preceding segment length corresponding to the phoneme symbol obtained from the input text, and the input text from the compressed speech waveform stored in the speech segment dictionary 1 The speech waveform of the speech unit corresponding to the phoneme symbol obtained from the above is included, and the speech waveform having the section length matching the preceding section length determined by the preceding section length determining unit 2 is added in front of the speech waveform. A compressed voice waveform reading unit 3 for reading the voice waveform, and the voice waveform extracting unit 4 expands the compressed voice waveform read by the compressed voice waveform reading unit 3, and the voice waveform of the speech unit from the expanded voice waveform Extract And then, it is, with a small storage capacity, it offers an advantage of being able to generate a synthesized speech of high quality.

That is, since the preceding section length determination unit 2 uniquely determines the preceding section length based on the phoneme symbol, it is not necessary to store information on the preceding section length for each speech unit, and the storage capacity increases. The effect is that it is possible to generate high-quality synthesized speech with less distortion.
Also, the phoneme symbols are classified according to the characteristics, and the length of the preceding section of the phoneme symbol that is less deteriorated when the user listens is shortened, so that the quality of the synthesized speech is avoided and the compressed speech waveform is avoided. It is possible to reduce the storage capacity and the processing amount required for decompression.

Embodiment 2. FIG.
In the first embodiment, the preceding section length determination unit 2 has set the preceding section lengths “Nc” and “Nv” so that the distortion of the speech segment section is sufficiently reduced, for example. However, the preceding section length may be set as follows.
For example, in the CV segment, consonants are classified into unvoiced consonants and voiced consonants, the preceding section length for unvoiced consonants is set to “Ncu”, and the preceding section length for voiced consonants is set to “Ncv”. Then, it is determined as “Ncu” <“Ncv”.
This is because the unvoiced consonant is a noisy signal and uses the characteristic that even when distortion is large in the unvoiced section, the deterioration feeling when the user listens is smaller than in the voiced section.
As a result, the length of the preceding section can be shortened for CV segments starting with unvoiced consonants, so that the storage capacity of the compressed speech waveform and the processing amount required for decompression can be reduced.

Also, for example, in a CV segment starting with an unvoiced consonant, the consonant is classified into a bursting unvoiced consonant and a rubbing unvoiced consonant. Set to “Nf”. Then, “Ns” <“Nf” is determined.
This is because the rupturable unvoiced consonant has a closed section in which the speech signal is close to silence, and the distortion of the rupturable unvoiced consonant does not increase so much even if the length of the silent section is reduced.
Accordingly, the length of the preceding section can be shortened for a CV segment starting with a bursting unvoiced consonant, so that the storage capacity of the compressed speech waveform and the processing amount required for decompression can be reduced.

For example, there are cases where “C” is silent as a special CV segment.
Thus, the preceding section length when “C” is silent is set to “Nsi”, and the preceding section length when “C” is sound is set to “Nvo”. Then, “Nsi” <“Nvo” is determined.
This uses the characteristic that even if the preceding silent section length is reduced, the distortion of the speech section does not increase so much.
As a result, the length of the preceding section can be shortened for a CV segment in which “C” is silent, so that the storage capacity of the compressed speech waveform and the processing amount required for decompression can be reduced. .

  It should be noted that the factors that determine the preceding section length are not limited to those described above, and it is needless to say that arbitrary information held in correspondence with the speech segment such as prosodic information can be used in addition to the phonetic symbol. .

Embodiment 3 FIG.
4 is a block diagram showing a speech synthesizer according to Embodiment 3 of the present invention. In the figure, the same reference numerals as those in FIG.
The speech unit dictionary 11 stores a compressed speech waveform in which a speech waveform including a section preceding the speech unit section is compressed as a speech waveform of the speech unit.
However, the speech unit dictionary 11 is different from the speech unit dictionary 1 of FIG. 1, when the speech segment speech waveform is time-reversed and the preceding section length is shorter than the time-reversed speech waveform, Instead of the speech waveform that is not time-reversed, a compressed speech waveform that is compressed by adding a speech waveform to a preceding section ahead of the speech waveform that is time-reversed is held.
For example, for the speech waveform of the VC segment, the speech waveform of the VC segment is time-inverted and regarded as the speech waveform of the CV segment, and the compressed data of the speech waveform is held.

  Similar to the preceding section length determining unit 2 in FIG. 1, the preceding section length determining unit 12 performs a process of determining the preceding section length corresponding to the phoneme symbol obtained from the input text. In addition, the preceding section length determination unit 12 performs a process of determining whether or not the time reversal of the speech waveform of the speech unit corresponding to the phoneme symbol is present. The preceding section length determining unit 12 constitutes a preceding section length determining means.

The compressed speech waveform reading unit 13 includes the speech waveform of the speech unit corresponding to the phoneme symbol obtained from the input text from the compressed speech waveforms stored in the speech unit dictionary 1, and precedes the speech waveform. A process of reading a compressed speech waveform to which a speech waveform having a section length that matches the preceding section length determined by the section length determination unit 12 is added.
However, unlike the compressed speech waveform reading unit 3 in FIG. 1, the compressed speech waveform reading unit 13 determines the speech waveform of the speech unit that is time-reversed when the preceding section length determination unit 12 determines that time reversal is present. In addition, a process of reading a compressed speech waveform to which a speech waveform having a section length that matches the preceding section length determined by the preceding section length determination unit 12 is added in front of the speech waveform is performed. The compressed voice waveform reading unit 13 constitutes a compressed voice waveform reading unit.

The time reversing unit 14 performs processing for reversing the speech waveform of the speech segment extracted by the speech waveform extracting unit 4 when the preceding section length determining unit 12 determines that time reversal is present. The time reversing unit 14 constitutes a time reversing unit.
FIG. 5 is a flowchart showing the processing contents of the speech synthesizer according to Embodiment 3 of the present invention.

Next, the operation will be described.
For example, when comparing the preceding interval length corresponding to the phoneme symbol of the CV segment and the preceding interval length corresponding to the phoneme symbol of the VC segment, the preceding interval length corresponding to the phoneme symbol of the VC segment is longer, The storage capacity of the compressed speech waveform increases.
In a general voice compression technique, even if a voice waveform of a VC segment is time-reversed and compressed, there is no significant difference in compression efficiency and quality compared to the case where the voice waveform of a CV segment is compressed.

Therefore, the speech unit dictionary 11 compresses the compressed data of the speech waveform that is not time-reversed when the time length of the speech waveform of the speech unit is shorter than that of the speech waveform that is not time-reversed. The compressed speech waveform that is compressed by adding the speech waveform to the preceding section in front of the speech waveform that is time-reversed is retained.
For example, the speech waveform of the VC segment is not held, the speech waveform of the VC segment is time-reversed and regarded as the speech waveform of the CV segment, and the compressed data of the speech waveform is retained.

Similar to the preceding section length determining unit 2 in FIG. 1, the preceding section length determination unit 12 stores a correspondence relationship between phonological symbols and preceding section lengths in advance, and when a phonological symbol obtained from the input text is input, the correspondence With reference to the relationship, the preceding section length corresponding to the phoneme symbol is determined (step ST11).
Since the method for determining the preceding section length corresponding to the phoneme symbol is the same as in the first embodiment, description thereof is omitted.

Further, the preceding section length determination unit 12 determines whether or not the time reversal of the speech waveform of the speech unit corresponding to the phoneme symbol is present (step ST12).
For example, comparing the preceding interval length corresponding to the phoneme symbol with the preceding interval length corresponding to the inverted phoneme symbol, if the preceding interval length corresponding to the inverted phoneme symbol is shorter, the time inversion is “present”. If the preceding section length corresponding to the inverted phoneme symbol is longer, it is determined that the time reversal is “none”.

The compressed speech waveform reading unit 13 is stored in the speech unit dictionary 11 as in the case of the compressed speech waveform reading unit 3 of FIG. A compressed speech waveform includes a speech waveform of a speech unit corresponding to a phoneme symbol obtained from input text, and a section that matches the preceding section length determined by the preceding section length determination unit 2 in front of the speech waveform A process of reading a compressed speech waveform to which a long speech waveform is added is performed (step ST13).
On the other hand, if the determination result of the preceding section length determination unit 12 is time reversal “Yes”, the speech waveform of the speech unit that is time-reversed from the compressed speech waveforms stored in the speech unit dictionary 11 is selected. In addition, a process of reading a compressed speech waveform to which a speech waveform having a section length that matches the preceding section length determined by the preceding section length determination unit 12 is added in front of the speech waveform is performed (step ST13).

When the compressed speech waveform reading unit 13 reads the compressed speech waveform, the speech waveform extracting unit 4 expands the compressed speech waveform as in the first embodiment (step ST14). That is, the compressed speech waveform of the preceding section + speech segment section is expanded.
Next, the speech waveform extraction unit 4 extracts the speech waveform of the speech segment from the decompressed speech waveform as in the first embodiment (step ST15).

When the speech waveform extraction unit 4 extracts the speech waveform of the speech segment, the time reversing unit 14 is extracted by the speech waveform extraction unit 4 if the determination result of the preceding section length determination unit 12 is “time reversal”. The speech waveform of the speech unit is time-reversed, and the speech waveform after the time reversal is output to the speech generation unit 5 (step ST16).
On the other hand, if the determination result of the preceding section length determination unit 12 is time reversal “No”, the speech waveform of the speech segment extracted by the speech waveform extraction unit 4 is output to the speech generation unit 5 as it is.

When the speech generation unit 5 receives the speech waveform of the speech unit from the time reversal unit 14, it controls the pitch and phoneme duration in accordance with the prosodic information obtained from the input text, and converts the phoneme symbol obtained from the input text. Therefore, the synthesized speech is generated by sequentially connecting the speech waveforms of the speech units (step ST17).
The processes of steps ST11 to ST17 are repeated until the input of phoneme symbols / prosodic information is completed (step ST18).

  As is apparent from the above, according to the third embodiment, when the compressed speech waveform that is time-reversed is included as the compressed speech waveform stored in the speech segment dictionary 11, the preceding section length is determined. When the unit 12 determines the preceding section length corresponding to the phoneme symbol, it determines whether or not the speech waveform of the speech segment corresponding to the phoneme symbol is time-reversed. Since the speech waveform of the speech unit extracted by the extraction unit 4 is configured to be time-reversed, the storage capacity of the compressed speech waveform and the processing amount required for decompression can be reduced without deteriorating the quality of the synthesized speech. There is an effect that can.

Embodiment 4 FIG.
In the first to third embodiments, the preceding section length determination units 2 and 12 have determined the preceding section length based on the phoneme symbol of the speech unit. However, the preceding section length is determined based on the phoneme symbol. Instead, the preceding section length may be determined based on the starting edge power of the speech unit.

As a method of determining the preceding section length corresponding to the starting power of the speech segment, for example, a method is considered in which the preceding section length is shortened if the starting power is small and the preceding section length is increased if the starting power is large.
This utilizes the characteristic that when the power of the audio signal is low, even when the distortion is large, the user is less likely to deteriorate when listening.
As a result, the length of the preceding section can be shortened for a speech unit having a low starting end power, and the storage capacity of the compressed speech waveform and the processing amount required for decompression can be reduced.

  In addition, when the power at the start and end of the speech segment is compared, and the power at the end is smaller than the power at the start, the compressed speech waveform that is compressed by time-reversing the speech waveform is held, and the compressed speech waveform is When synthesizing speech, if the speech waveform obtained by decompression is time-reversed and used for synthesis, the storage capacity of the compressed speech waveform and the processing amount required for decompression can be further reduced. There is an effect that can be done.

BRIEF DESCRIPTION OF THE DRAWINGS It is a block diagram which shows the speech synthesizer by Embodiment 1 of this invention. It is a flowchart which shows the processing content of the speech synthesizer by Embodiment 1 of this invention. It is explanatory drawing which shows the compression speech waveform stored in a speech unit dictionary. It is a block diagram which shows the speech synthesizer by Embodiment 3 of this invention. It is a flowchart which shows the processing content of the speech synthesizer by Embodiment 3 of this invention.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 1 Speech segment dictionary, 2 Leading section length determination part (preceding section length determination means), 3 Compressed speech waveform reading part (compressed speech waveform reading means), 4 Speech waveform extraction part (speech waveform extraction means), 5 Speech generation part (Speech generating means), 11 speech segment dictionary, 12 preceding section length determining section (preceding section length determining means), 13 compressed speech waveform reading section (compressed speech waveform reading means), 14 time reversing section (time reversing means)

Claims (4)

  Corresponds to the phoneme symbol that is obtained from the input text and the phoneme unit dictionary storing the compressed phonetic waveform in which the phonetic waveform including the segment preceding the phoneme segment is compressed. A preceding section length determining means for determining a preceding section length; and a speech waveform of a speech unit corresponding to a phoneme symbol obtained from the input text from among the compressed speech waveforms stored in the speech unit dictionary, Compressed speech waveform reading means for reading a compressed speech waveform in which a speech waveform having a section length matching the preceding section length determined by the preceding section length determining means is added in front of the speech waveform; and the compressed speech waveform reading means A speech waveform extraction means for expanding the compressed speech waveform read out by the step, and extracting the speech waveform of the speech unit from the expanded speech waveform; and a sound obtained from the input text Speech synthesis apparatus and a speech generation means for generating sequentially connected to synthesized speech audio waveform of speech units extracted by the speech waveform extracting means in accordance with the symbol.   When the compressed speech waveform stored in the phoneme unit dictionary includes a compressed speech waveform that is time-reversed, the preceding section length determination means determines the preceding section length corresponding to the phoneme symbol. If the speech waveform of the speech unit corresponding to the phoneme symbol is determined to be time-reversed and the time-reversal is determined by the preceding section length determining means, the speech of the speech unit extracted by the speech waveform extracting means 2. The speech synthesizer according to claim 1, further comprising time reversing means for reversing the waveform with time.   The speech synthesizer according to claim 1 or 2, wherein the preceding section length determining means determines the preceding section length corresponding to the power obtained from the input text instead of the phoneme symbol.   The preceding section length determining step for determining the preceding section length corresponding to the phoneme symbol obtained from the input text by the preceding section length determining means, and the speech waveform including the section preceding the speech unit section as the speech waveform of the speech unit From the speech unit dictionary storing the compressed speech waveform in which the compressed speech waveform is compressed, the compressed speech waveform reading means includes the speech waveform of the speech unit corresponding to the phoneme symbol obtained from the input text, and the front of the speech waveform A compressed speech waveform reading step for reading a compressed speech waveform to which a speech waveform having a section length matching the preceding section length determined by the preceding section length determining means is added; and a speech waveform extracting means comprising the compressed speech waveform reading means A speech waveform extraction step of decompressing the compressed speech waveform read out by step (a) and extracting the speech waveform of the speech unit from the decompressed speech waveform; Speech synthesis method and a speech generation step means generates a sequentially connected to synthesized speech audio waveform of speech units extracted by the speech waveform extracting means in accordance with the phoneme symbols obtained from the input text.

Images