JPS6022195A - Synthesization of voice - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

Detailed Description of the Invention (Technical Field) The present invention provides a method for preparing a plurality of re-key channels, sequentially and cyclically specifying one playback channel to play back speech segments one by one, and Regarding the method of synthesizing speech by superimposing the output of
In particular, the present invention relates to a speech synthesis method in which a syllable memory and a segment memory are prepared and a plurality of channels are set by controlling the addresses of these memories.

(Prior Art) Most of the methods of producing speech that are currently in practical use use natural waveforms or compressed waveforms (for example, ADPCM code,
PARCOR parameters), which are then edited and synthesized according to the output.Recently, a large number of words or unlimited arbitrary sounds such as people's names, company names, and place names can be output for general purposes. There is an increasing demand for equipment that requires In order to store such a large amount of linguistic information in a voice output device and output it as voice as necessary, it is unrealistic to rely on human vocalizations, and voice storage is also necessary. Files also require a huge amount of storage space. Therefore, research is being carried out on arbitrary word synthesis methods that synthesize speech waveforms only from character sequences.

Here, a waveform domain arbitrary word synthesis method with a simple synthesis circuit will be described.

By the way, when observing the waveform of speech, it can be seen that very similar waveforms are repeated in the sections of voiced sounds such as vowels.

This period is called a pitch period, and the waveform within the - period is called a speech unit in pitch units.

Changes in the content of this elemental piece express phonological properties, and the temporal changes in this cycle give accents.

As mentioned above, in voiced sound sections such as vowels, waveforms of almost the same shape are repeated, and waveforms of similar shapes appear in the same type of speech. Therefore, among the waveforms that appear in speech, characteristic speech waveforms necessary for producing speech are stored in a storage device, and by editing and joining these basic phoneme waveforms, arbitrary continuous speech can be created. It is thought that it is possible to synthesize

On the other hand, any sentence in Japanese can basically be expressed using over 100 different types of monosyllables. Therefore, by storing this many monosyllabic sounds in a storage device, it is possible to synthesize all of Japanese.

In this way, in order to output arbitrary words as speech, it is necessary to select something smaller than a word as the basic unit of speech, but in what units are words classified and memorized in this type of speech synthesizer? Whether or not to use the synthesizer is an important issue that determines the scale and sound quality of the synthesizer.

In other words, when it comes to phonology, smaller basic units like phonemes require fewer units to synthesize any linguistic message, but acoustics due to interactions with neighboring linguistic units (articulatory combination) are smaller. Changes in physical characteristics are rapid,
Simply connecting them together will not provide high-quality audio. If we use syllables as the basic unit, there are fewer operations to connect sounds.In addition, since sounds are memorized including the auditory parts, phonology is well preserved and rules for combining sounds are also preserved. Although it is expected that it will not be necessary to make it too complicated, there will be many duplications of phonemes in each single syllable, and the capacity of the storage device will increase.

On the other hand, phonological information as well as prosodic information such as accent, intonation, and duration are extremely important elements for obtaining natural synthesized sounds.

In a waveform domain synthesis device, in order to appropriately control the prosody using speech units stored in the storage device as raw materials, it is necessary to change the pitch, amplitude, and duration of the voice according to the instructions in the dictionary of control information. Create continuous audio.

Considering the controllability of prosody, it is more convenient to use the phoneme as the basic unit, and in order to perform prosody control using a monosyllable as the control unit, it is necessary to select an appropriate one from among multiple monosyllabic waveforms with different pitches and durations. There are no other useful options.

As described above, as the basic unit of speech in an arbitrary speech speech output device, it is better to use a speech segment in terms of storage capacity and prosody, and monosyllable level is better in terms of phonology.

For this reason, a method has been proposed in which the unit of control is a single syllable in terms of phonology, and the unit of control in terms of prosody is a speech segment. In other words, this is a method in which the speech segments recorded in the storage device are sequentially arranged within each monosyllable block. In this method, phonetic properties are maintained by sequentially extracting speech segments in chronological order for each single syllable, and speech segments can be used as a unit for prosody control.

However, in this method, the basic unit of memory in the storage device is the speech segment, but in terms of storage capacity, all monosyllabic waveforms are still stored in the storage device, and the storage capacity is limited. cause an increase. As an example, FIG. 1 shows a monosyllabic waveform.

Figure 1 a is a monosyllable /ma/, Figure 1 is a monosyllable /m 1
/, FIG. 1C is an example of the waveform of the monosyllable /ri/.

Each of these waveforms is recorded in the storage device as a monosyllabic waveform. Here, if we consider the phonetic segment as a unit, /ma/ and /
The m part of mi / and the i part of /mi/ and /r t/ can be considered to be almost the same elemental fragment, but each elemental fragment needs to be recorded sequentially in units of syllables. , even if the same elemental fragment has different syllables, each must be recorded separately.

(Objective of the Invention) The present invention stores speech segments in units of one pitch or speech segments representing a sample time series of attenuated waves obtained by converting the speech segments in a segment memory, and also stores all speech segments in a segment memory. It stores data representing a sample time series consisting of a fixed number, and further includes a syllable memory that stores a waveform corresponding to a single syllable unit as a starting address for each segment and the number of repetitions for each segment. A speech synthesis method that aims to reduce memory capacity by setting multiple channels and adding and superimposing all channels to reduce discontinuity at the joints of speech segments. This objective was achieved in the following.

(Summary of the Invention) First, waveform reproduction using a plurality of reproduction channels, which is a premise of the present invention, will be explained with reference to FIG.

In FIG. 2, E Ll, ~Es1s, p EL2
1 p EL22 and FT, ~PT5 indicate a sequence of speech segments and a series of pitch periods, respectively, created based on input information regarding speech to be synthesized. Sequence of speech segments ELII # EL12 r ELt3p E
L21 p EL22 and pitch period sequence P'r1t
P'r2 y' FT3 g FT4 t FT5 is shown as a pair. Four channels, 1st to 4th, are prepared as waveform playback channels, and the first speech segment EL1 is played back on the first playback channel, and the next voice segment EL12 is played back on the second playback channel.
play it on the channel, change the playback channel one after another, play it again, and then play it again on the first playback channel. In this way, by sequentially and cyclically selecting and reproducing one reproduction channel, for example, by reproducing a speech segment on the first reproduction channel, even if a short pitch period is set. Each speech segment is played back over its entire length, and a composite waveform can be obtained by adding all these playback channels. In addition, in FIG. 2, the same waveform is shown for simplification, but E L1
+ ~ EL13 and EL2t + , EL12 are different phonetic units '), EU, 11 to EL13 are the same phonetic unit EL1'), and EL21 and EL22 are the same phonetic unit EL2. Therefore, in FIG. 2, the phonetic segment E
L is repeated only three times, and speech segment EL2 is repeated only twice.

The present invention reduces the storage capacity in such waveform reproduction by providing a syllable memory and a segment memory for storing speech segments of a certain length. This can be done simply by setting the address variable of the segment memory and then using address control to read out one sample of data for each segment to each channel in synchronization with the sample clock based on the address variable. Examples will be described below.

(Example) Next, an example will be described using FIGS. 3 to 7.

FIG. 3 is a block diagram showing the speech synthesis device.
The figure is a flowchart showing the functions and control procedures executed by the microprocessor 1. Input information regarding the speech to be synthesized is input from the typewriter 2 to the microprocessor 1 in the form of a series of character codes indicating monosyllables for each word with appropriate pauses and the like. The prosody memory unit 3 stores prosodic control information such as accent type, intonation type, and duration regarding input words, as well as monosyllabic information, which are searched by the microprocessor 1 to form monosyllabic information according to the input information. An information sequence and a pitch period sequence PTj are created.

Further, the monosyllabic information stores information specifying the starting address of each monosyllable in the monosyllabic memory 4, a series of starting addresses for each unit such as the length of a single syllable, and the number of repetitions for each segment. The segment unit start address is information that specifies the start address of each segment unit in the segment memory 5. The segment memory 5 stores segment units necessary for reproducing arbitrary speech segments in all Japanese syllables, and the segment units are created by an appropriate compression technique such as DPCM, and All segment units are stored in an amount of data that can reproduce a sample time series of a predetermined length of 128 samples. The waveform regenerator 6 has a configuration compatible with the adopted compression technology, and reproduces the corresponding sample time series based on the data in units of fragments. There are 4 playback channels RG so that 128 samples can be played back.
1 to RG4 are provided. Waveform reproduction is performed by setting the pitch period sequence PTj and the sequence of segment unit information in correspondence on the real-time time axis set by interrupt activation of the sample clock, and reproducing the pitch period sequence PTj on the time axis every time the pinch period is updated. By selecting the corresponding element unit ELt of the element unit selection information and selecting one reproduction channel sequentially and cyclically every time the pitch period is updated, the element unit to be reproduced and the reproduction channel are made to correspond, and all 1 in the playback channel of
This is done by regenerating Sansol one by one and superimposing them.

5 to 7 show the waveform generation flow executed by the microprocessor l.

FIG. 5 shows the flow of waveform reproduction. Address variables ADI-%-AD4 for four channels, which are initialized and updated as described later, are prepared, and Stella f
In SP11, the address of the elemental piece memory 5 is specified with the first address variable AD1, one piece of data of one elemental piece is read out, and the data is sent to the first reproduction channel RG1 of the waveform regenerator 6. Play the sample.

In steps 5P12 to 5P14, the second to fourth
Addresses of the segment memory 5 are designated by address variables AD2 to AD4, and one sample is reproduced by the corresponding second to fourth reproduction channels RG2 to RG4, respectively. Step 5 In PI5, all playback channels RG,
By adding and superimposing the outputs of ~RG4, that is, 4 samples, one sample of the composite output is reproduced and output.

The synthesized output is reproduced by reproducing one sample each time the sample clock interrupt is activated.

FIG. 6 shows a flowchart for setting the start address of each elemental piece on the time axis of the sample clock.

In FIG. 6, it is started by an interruption of the sample clock, and in step 5T21 it is determined whether the single syllable currently being processed remains. Single note length variable PH1 = 0
Then, in step 5P22, the monosyllable number is incremented to extract the next monosyllable information from the previously created series of monosyllable information, and in step 5P23, the monosyllable length PHLa is loaded into the monosyllable length variable PH1. Then, the monosyllable start address is loaded into the address change i ADp of the monosyllable memory 4'. Next, step 5
At P24, the address of the monosyllable memory 4 is designated by the monosyllable start address, and the start address of the first segment unit ELj and the number of repetitions rJ of the segment unit are read out. In step 5P25, the segment unit start address is loaded into the address variable ADe of the segment memory 5.

In this step 5P25, the first segment unit head address in one monosyllable is set in correspondence with the sample clock. Further, in step 5P26, the remaining length of the segment unit ELj currently being processed is determined, and if the segment length variable ELt=0, the content of the address variable ADp is incremented in step 5T27 to specify the next address, and the Stellar Through fsP24.5P25, the start address of the next elemental unit ELj+1 in one single syllable is set on the time axis of the Sanzor clock. Steps 5P28 and 5P29 are performed in the process of counting the segment length on the time axis of the sample clock.The number of samples in all segment units is the same, 128, so in step 5P28, the segment length variable ELt is used to repeat the segment length. Number of times r
This is executed by loading the value multiplied by 128 e and decrementing it in synchronization with the sample clock in step 5P29. Also, in Stella 7° 5p30, the monotone cylinder length variable PH1 is decremented in synchronization with the sample clock. I'll keep it.

FIG. 7 shows a flow for setting the pitch period in real time and associating four segment units simultaneously read from the segment memory 5 with predetermined reproduction channels RGI-RG4. Referring to FIG. 7, each time an interrupt is activated by the sample clock, the remaining length of the pitch period PTj currently being set is determined in step 5P40, and if the pitch period variable 'pTt = 0, steps 5P41~ 5P
After the processing in step 43, the pitch number j is incremented in step 44 to extract the next pitch period PTj from the previously created pitch period series force≧, and step 5P4
In step 5, each pitch period PTj is set on the time axis of the Sanzor clock by converting it into a pitch period variable Pt, and by decrementing the pitch period variable pI(z. In steps 5P41 to 43, which are executed in synchronization with the update, one channel is sequentially and cyclically selected by incrementing the 4 channel specified value CH in step 5P41, and the 4 channels of the segment memory 5 are selected in the stellar fsP42.
The address variable ADch corresponding to the specified value CH is selected from among the address variables AD, ~AD4 for the channels, and in step 5P43, the address variable ADe indicating the start address of the currently set elemental unit ELi is selected for the selected address variable ADch. Each segment unit is assigned to each channel by loading it into the address variable ADch of the channel. In step fsP47, by incrementing address variables AD1%AD4 of each channel, each address variable AD1 to AD4 can sample data that reproduces samples in different unit of fragments and in different orders. It is possible to address sequentially in synchronization with the sol clock.

In the above embodiments, the processing of amplitude information that determines the strength of the voice is omitted, but it is monitored in the same way as monosyllabic information and is provided to each playback channel in synchronization with the update of the voice segment, thereby determining the strength and weakness of the voice. can be controlled.

As explained above, the speech synthesis device according to the present embodiment is equipped with an address memory in which addresses of speech unit data are stored in Japanese monosyllables as phoneme control units, that is, with each monosyllable as one block. Second, since the speech unit data is extracted continuously for each monosyllable, the phonology of the output monosyllabic speech is sufficiently preserved. However, since prosody control uses speech segments in the speech segment memory as units of control, it is possible to give appropriate prosody to synthesized speech.

Furthermore, since monosyllables are managed using addresses of speech segment data, there is no need to store speech segment data consecutively in monosyllable units in the speech segment memory, and therefore speech segment memories do not overlap phonologically. Since it can be configured with the minimum amount of speech segment data, the storage capacity of the storage device can be reduced. Also,
These advantages can be realized by a method of superimposing the outputs of multiple playback channels with simple address control.

(Effects of the Invention) As is clear from the above description, according to the present invention, since a plurality of playback channels are provided, speech segments are smoothly connected, and monosyllabic memory and segment memory are provided. Since address control corresponding to a plurality of channels is performed using the system, synthesized speech that is natural both phonologically and prosodically can be obtained with a small storage capacity and a simple procedure.

[Brief explanation of drawings]

Fig. 1 is a diagram showing an example of an audio waveform, Fig. 2 is an explanatory diagram showing the concept of waveform reproduction adopted in the present invention, Fig. 3 is a block diagram showing an embodiment of the present invention, and Fig. 4 is the overall flow. , and FIGS. 5 to 7 are diagrams showing detailed flows thereof, respectively. 1...Microprocessor, 2...Tiger writer,
3... Prosodic memory, 4... Syllable memory, 5... Piece memory, 6... Waveform regenerator. Patent Applicant: Oki Electric Industry Co., Ltd. Figure 3 Figure 4 Figure 5 Procedural Amendment (Self-restraint Showa! November 1, 2015 Director General of the Patent Office 1, Indication of Case 1988 Patent Application No. 129399 2, Name of the invention Speech synthesis method 3, relationship with the case of the person making the amendment Patent applicant's office (〒105) 1-7-12 Toranomon, Minato-ku, Tokyo
No. 4, Agent address (105) 1-7-12 Toranomon, Minato-ku, Tokyo
No. 5, Subject of amendment Column 6 of "Detailed Description of the Invention" in the specification, Contents of amendment Attached sheet, Contents of amendment (1) "Compressed waveform" and 4 on page 3, line 18 of the specification Correct the word "compression code" to "compression code". (2) From line 7 to line 8 of page 8 of the same book, the phrase ``even if they are the same elemental piece'' has been amended to ``even if they are similar elemental pieces.'' (3) In the 5th line of page 12 of the same book, it says ``any phonetic segment in Japanese syllables.'' I am corrected by saying, ``Japan Shigoon Jinbei''.

Claims (1)

[Scope of Claims] a) Contains a large number of segment units representing a sample time series associated with a speech segment of one pitch unit in natural voiced speech, and all segment units are divided into a predetermined number of segments. It is equipped with a segment memory that stores data representing a sample time series, and b) stores a large number of single syllables by the number of repetitions of each segment unit and the sequence of the start address where the segment unit is stored. c) having a plurality of predetermined playback channels for reproducing the sample time series based on the selected segment unit, and d) setting the address of the segment memory as a sample clock. Address variables to be updated in synchronization are prepared corresponding to each of the playback channels, and e) a series of monosyllabic information created based on input information regarding the speech to be synthesized is sequentially extracted once. specifies the start address of one single syllable, and reads the start address of one segment unit and the number of repetitions of that unit from the syllable memory, and also reads the number of repetitions or the number of repetitions of that unit, or the number of repetitions read immediately before. Every time a certain number of sample clocks are counted as the number of repetitions per unit of segment, the address of the syllable memory is advanced by one.
f) Selecting the start address of one segment unit read from the syllable memory in synchronization with the update of the bi-tsuchi cycle;
and sequentially and cyclically selecting one address variable from among the address variables in synchronization with the update of the pitch period;
Set the address of the selected segment unit in the selected address variable, g) update each address variable in synchronization with the sample clock, and update the segment memory with each address variable at each sample clock. Specify the address and give the data for each fragment unit to the corresponding playback channel, h) Synchronize with the sample clock and give the data for each playback channel
A voice synthesis method characterized by synthesizing voice by playing back samples one by one and then superimposing the output of each playback channel.