JPS6143337A - Voice input device for japanese word - Google Patents

Abstract

PURPOSE:To improve the recognizing performance with no load of users by registering automatically a feature standard pattern obtained from input voices as a feature standard pattern of syllables corrected through collation with a dictionary, etc. CONSTITUTION:The input voices sent from an analog input part (a) undergo the spectrum analysis through a voice analysis part (b), and the syllables are separated through a syllable segmentation part (c). A feature pattern of the syllable separated part and the voice waveforms are stored in a feature pattern temporary memory (d). Thenthe distances are calculated at a single syllable recognizing part (f) among a feature pattern 1g in a pattern memory (g) and standard pattern memories 2g and 3g. The contents of the memory (d) are transferred to the memory 1g and collated with the contents of memories 2g and 3g respectively. The results of these collations are stored to a memory and also displayed to a display device (k). Then a mistake if detected is corrected through operations of a keyboard and collation with a dictionary. If no mistake is detected, a syllable feature pattern is registered automatically to the memory 3g.

Description

[Detailed description of the invention] Technical field> Honji F! A relates to the improvement of a Japanese speech input device that recognizes input speech in units of syllables, and in particular, to automatically register the characteristic standard pattern of the speech uttered at the time of input as the standard pattern of syllables obtained as a result of modification. This is what I did.

<Prior art> Conventional speech recognition devices are divided into a registration mode, in which standard nomiturns of speech characteristics are registered in advance, and a gR mode (input mode), in which input speech is recognized, and the input speech is analyzed in recognition mode. It was not possible to register the feature pattern obtained as a feature standard pattern. This is not much of a problem when using words as the unit of recognition, but when using syllables as the unit of recognition, each syllable is influenced by the syllables before and after it (articulatory combination), and when the unit of recognition is a word or sentence box, There was a problem in that the strength and pitch of the voice differed greatly depending on the syllable position, resulting in a decrease in recognition performance.

Conventionally, to solve these problems, one syllable was registered several times at the time of registration, and this was done by having many characteristic standard patterns. They vary, and it is rarely possible to register in advance a characteristic standard pattern for the syllable environment in all cases.

Therefore, the main purpose of the present invention is to automatically add or replace the characteristic standard pattern of the voice uttered at the time of input with the standard pattern registered in advance as the standard pattern of the syllable string corrected by dictionary checking etc. The objective is to improve recognition performance without burdening the user.

Another object of the present invention is to improve recognition performance by registering at the time of input only syllables with poor recognition rates determined by a counter that determines the recognition rate of each syllable standard pattern.

Another object of the present invention is to register standard patterns only in the main memory at the time of input, and further register only patterns that are effective by external instruction operations such as a keyboard in the auxiliary memory, thereby automatically registering patterns at the time of input. The objective is to minimize the negative effects caused by incorrect pattern registration.

<Example> Hereinafter, with reference to the drawings, the present invention will be described in which continuously uttered speech is recognized in units of syllables, and the recognition results are compared with a dictionary.
An example of a Japanese speech input device that has a function of transmitting the input data in units of words or the like to an external device after correction will be explained.

FIG. 1 is a block diagram showing the configuration of an apparatus according to an embodiment of the present invention.

In Figure 'I, the voice that is uttered and input is input to the analog input section a via a microphone, etc., and after being amplified by the amplifier in the analog input section a, it is converted into a digital signal by the analog/digital conversion section. The digital signal is input to the speech analysis section and the syllable segmentation section C.

The speech analysis section divides the input speech into frames of about 16m5 and performs spectrum analysis, and at intervals of about 8mS, the syllable segmentation section C analyzes characteristic patterns and information necessary for syllable segmentation (power, number of zero crossings, etc.).
Transfer.

The syllable segmentation section C cuts out syllables from the input speech using various information from the speech analysis section.

The feature pattern of the cut out portion and the audio waveform of that section are stored in the waveform/feature pattern temporary memo Ijd. Then, the fact that the syllable has been cut out is transmitted to the CPUe, and the address in the waveform/feature pattern temporary memory d is also transmitted at the same time.The waveform/feature pattern temporary memory d is configured to be able to store a plurality of syllables.

The processing of the syllable segmentation unit C is configured such that the start and stop of the processing is controlled by instructions from the CPUe.

f is a monosyllable recognition unit, and in the monosyllable recognition unit f, CP
According to the command from Ue, the characteristic pattern memo 171F and the standard pattern memory 2? and 31, and returns the results to the CPUe. Then, the CPU e stores the results in the recognition result storage memory l, and the recognition result storage memo IJ, which is displayed on the display device k, is configured to be able to store recognition results for a plurality of syllables. .

The pattern memory 2 is divided into three parts, and It is a feature pattern memory that can store only one feature pattern corresponding to the input syllable, the two after 0, 21 and 3? is the memory for the feature standard pattern.
), 5) The 2f area is the memory for the syllable feature pattern to be registered in the registration mode, and the ay area is the memory for the syllable feature pattern to be registered in the recognition (input) mode realized in this patent. Memory 0 The syllable characteristic pattern memories 22 and 31 are areas for storing each syllable base in the form of a code, respectively, as will be described later.
It is composed of a flag area that stores the presence or absence of registration, and a feature standard pattern area that stores feature standard pattern data. □i is an expansion memory, which is a word candidate created by expanding syllable candidates during dictionary matching. It is a memory that stores information. Further, j is an input unit composed of a keyboard, etc., and for example, as shown in FIG. 2, it is equipped with a registration mode key j1, a recognition mode key j2, a kana key j14, etc. ok is an instruction device such as a CRT, and t is an instruction device such as a CRT. I controls data transmission and reception when transferring recognition results to an external device
/F (interface) part, m is a syllable counter part consisting of a correct answer counter and a recognition counter as described later;
n is a dictionary memory that stores all the voices input for automatic correction, and O is an auxiliary storage section (for example, a floppy disk) for storing characteristic patterns.

Next, the operation of the apparatus configured as described above will be explained in the registration mode and the recognition mode. FIG. 3 shows the processing flow of (::PUe) in the registration mode.

In FIG. 3, when the main body of the apparatus is set to the registration mode by operating the registration mode key j+, the pattern memory 1 is first initialized in step nl, and all standard patterns are erased. Table 1 shows the configuration of the standard pattern memory 2t, and the standard pattern memory 8f is configured in the same way.The initialization process in step nl is the configuration example of the 0th ti standard pattern 27. This is achieved by putting "0" in the flag area indicating whether or not registration is available in No. 32. Next, the process moves to step n2, and the monosyllable to be uttered is displayed on the display device k as follows.

"A" Here, the subscript "1" indicates that it is the first in the "A" pattern.

The operator looks at the display on the display and inputs a predetermined monosyllabic sound by uttering it.

In response to this audio input, the process moves to step n3, where the syllable segmentation unit C is instructed to start audio segmentation, and the syllable segmentation unit C segments monosyllables.
The waveform of that section and the characteristic pattern obtained by the voice analysis section are stored in the waveform/characteristic pattern temporary memory d.

In step n4, the syllable segmentation unit C checks whether or not a single syllable has been segmented, and if it has been segmented, the process moves to the next step n5.In step n5, the syllable segmentation unit C is commanded to stop the segmentation process, and the registration is performed. Continue the process. At step n6, the audio portion corresponding to the syllable just output is read out from the waveform/characteristic pattern temporary memory d and reproduced and outputted via the audio output control section.

In step n7, it is determined whether the operator has been accurately cut out based on the reproduced audio, and in accordance with the resulting instructions from the keyboard j, it is determined whether to re-cut or register. If the operator listens to the playback output and determines that the clipping has been performed accurately, he or she will operate the execution key j7 and proceed to step n8, and if the operator instructs re-cutting, he or she will operate the release key j6. Accordingly, the process returns to step n3.

In step n8, the display device kK! The feature standard pattern corresponding to the indicated syllable is stored in the location of the turn memory 22, and a flag indicating whether or not the corresponding registration is present is set to "1".

In step n9, it is determined whether or not all standard patterns have been registered. If not, the process returns to step n2, and the next monosyllable is displayed, for example, "AzJt-".
FE is indicated and the same processing is performed.

In this way, when the registration is completed, several characteristic standard patterns of all monosyllabic characters are registered in the standard pattern memory 22.

Next, the operation in recognition mode will be explained.

(2) Explanation of the g-thread mode FIG. 4 shows the processing flow of the CPU e in the m-thread mode.

First, the device is set to the recognition mode by operating the recognition mode key j2, and when the operator utters the voice to be recognized, in step nil, the syllable is cut into the syllable segment M/part C according to the input voice. Give the command to start. Then, the syllable segmentation part C is a waveform
The feature pattern temporary memo IJdt is initialized, and thereafter feature patterns and waveforms corresponding to the cut out syllables are entered from the first address, and information on the waveform of each syllable and the start and end addresses of the feature pattern is given to the CPUe.

In step n12, a check is made to see if a syllable has been cut out, and if it has been cut out, the process moves to step nlB.

In step n13, the waveform/feature pattern temporary memory d
The characteristic pattern is transferred to the characteristic pattern memory 1f area of the pattern memory 1 for recognition. That is, by giving a command for monosyllable recognition unit fK recognition, recognition is performed by comparing the contents of the characteristic pattern memory lf with the contents of the standard pattern memos IJ2f and 3f, and the result is stored in the recognition result storage memo 'Jh. At the same time, it is displayed on the fitting device k (step n14).

For example, if the first recognition result when uttering "Kaimono" as an input voice is "Kagimomo", the display device kK
Hakagimomo- is displayed, and the recognition result storage memory hK stores a plurality of recognition result candidates for each syllable, for example, as shown in the second part.

Table 2 (The numbers in brackets indicate distances normalized by the value of the ``th'' candidate.) When the operator finishes inputting the above word such as ``kaimono,'' the operator presses the ``end'' key j5 on the keyboard j. Then, the syllable segmentation unit C is commanded to stop cutting (steps n15 and n16).

Then, if all the character strings are correct, by inputting the "transfer" key j8, the kana characters can be output to the external device via the I/F section t (step 118.n2
0).

Also, if you look at the display of the recognition results and find that most of the characters are incorrect or you have made a mistake, you can return to the initial state based on the judgment in step n17 by inputting the "cancel" key j4. In addition, if some of the recognition results are incorrect, correction is performed by dictionary comparison by pressing the word next candidate key j9 (step n19). When pressed for the first time, word candidates are expanded in order of decreasing total distance, as shown in Table 3.

3. When the table development is completed, the first candidate is searched for a match with the dictionary, and the matched one is displayed on the display section k (step n20)0
At this time, the candidates that were previously displayed are highlighted.

Furthermore, when the word next candidate key j9 is pressed for the second time or later, the above expansion is not performed and a match with the dictionary is determined from the next candidate of the previous matching candidate. When all the expansion candidates are matched in this way, the process returns to the beginning 9, displays the first candidate, and waits for the next key. In addition, the previous matching candidate can be displayed using the pre-word candidate key jlO.

Next, in step 21, the counter m corresponding to each syllable is operated.
It consists of a total of four types of counters: 1 correct answer counter, recognition counter, total correct answer counter, and total recognition counter.

Table 4: Structure of syllable counter Each of the above counters is operated according to the following items.

l] Add 1 to the recognition counter of the syllable of the output result 0 2) Add I to the correct answer counter for the syllable for which the gFJ1 candidate in the recognition result storage memory h and the output result are equal.

3) Add the number of recognized syllables to the total recognition counter.

4) Add the number of syllables for which the first candidate in the recognition result storage memory h and the output result are equal to the correct total force Kunter.

When the counter operation is performed based on each item described above, the process proceeds to step n22.

In step n22, the syllable feature pattern of the voice in the feature pattern temporary memory d defined by the following determination method is transferred to the standard pattern memory 3F of the pattern memo IJ f.
That is, it is registered in the main memory.

Judgment method> ■Register all feature patterns as standard patterns of syllables given in the output results, regardless of the counter.

■ Regardless of the counter, only syllables for which the output result and the first recognition result differ are registered as the syllable standard pattern specified by the output result0 ■ Only syllables whose syllable counter specified by the output result satisfies the following conditions register.

Mouth Correct answer counter/recognition counter < THI■ ■
Register only the syllables that meet the conditions of big ■ O ■ Correct answer: Overall power Kunta/! Whole knowledge counter > T
At H2, either method above ■■■■ or
9. Use the method of not registering. Otherwise, use one of the above methods (however, it is different from the previous one) t-.

The above method is an example of a device with a counter in Table 4, and a similar method is to use a ring counter for each syllable to determine the pronunciation syllable based on the recognition rate of the past few syllables. This can also be achieved by registering syllables that have been incorrectly recognized several times in a row in the past.

However, all of the above methods will be disabled by switching the mode from the keyboard, and you will be able to choose not to register.

The configuration of the standard pattern memo I73f is as shown in Table 5.

Table 5 Standard pattern 3? Configuration example In this example, since the long sword voice is registered as "i", the feature pattern is transferred to the area where there is no registration in the area of the syllable base "i", that is, "i", and the presence/absence of registration is determined.
Make it J. If the presence/absence of registration is "l", for example, in the case of "a" in Table 5, up to "a." are registered, so "a3" is the oldest in terms of time. Delete area data and then register in that area.

In this way, the recognition rate is calculated for each syllable using the correct answer counter and the recognition counter, and only the syllables with a poor recognition rate are automatically registered in the feature standard pattern memory 31, which is the main storage device. This prevents the standard pattern from being replaced. After that, when the pattern write key j13 is pressed, the contents of the feature standard pattern memory 3F are registered in an auxiliary storage device such as a floppy disk (steps n23 and n24). It is possible to achieve the following effects.

(6) Once the minimum amount of registration has been performed, additional registration is automatically performed upon input, so there is no need to register all standard patterns in advance.・・・・・・(The number of registered words can be reduced) (B) Even with the conventional method of registering all syllables, the pronunciation at the time of registration and the pronunciation at the time of input are often different, so updating registration etc. However, according to the present invention, the standard pattern is automatically updated at the time of input, so there is no need for update registration. The final recognition rate is reached quickly.・・・・・・(Recognition rate improves quickly) (c)
The utterances made in the registration mode may be quite different from the utterances made in the recognition mode, and automatic registration in the recognition mode eliminates this inconvenience and improves the final recognition.・
(Improve the final recognition rate)) Even if the time of registration and the time of input are different in time, or the voice changes due to a cold, etc., automatic registration at the time of input allows the voice to be used at that time. Since a standard pattern is created, it is possible to prevent the recognition rate from dropping drastically.・・・・・・(
(E) By having a recognition counter and a correct answer counter for each syllable, calculating the recognition rate for each syllable, and registering only those syllables with a poor recognition rate as feature standard patterns. It is possible to prevent the standard pattern of syllables with a good recognition rate from being replaced, and improve the recognition rate.

(f) If the pattern registered at the time of input is effective in actual input, it is written to the auxiliary storage device by external instruction operation, and the standard pattern is used for the next input, so automatic registration at the time of input Negative effects can be minimized.

In the above embodiment, the standard pattern memories 2F and 35
Although it was structured so that it was registered in two parts,
By using a counter that indicates the quality of standard patterns disclosed in Japanese Patent No. 7-217296, the worst feature standard pattern may be erased and the syllable at the time of input may be registered in that area.

As described above, according to the present invention, a plurality of types of syllable characteristic standards in which input speech is registered in advance.

A Japanese voice input device recognizes the syllable by syllable by calculating the similarity with a pattern, and then corrects the result by checking it with a dictionary or by operating an external command such as a keyboard to obtain the final input. Since the feature standard pattern obtained by analyzing the input voice is automatically registered as the syllable characteristic standard pattern corrected by dictionary comparison, etc., the input voice can be changed without any burden on the user. Features can be added to or replaced with standard patterns to improve recognition rates.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing the configuration of a device according to an embodiment of the present invention, FIG. 2 is a plan view showing an example of a keyboard, FIG. 3 is a processing flow diagram for explaining the operation in registration mode, and FIG. 4 FIG. 2 is a processing flow diagram for explaining operations in recognition mode. b is a speech analysis unit, C is a syllable segmentation unit, d is a feature pattern-order memory, and e is a CPU. 1 is a pattern memo'), il'lt expansion memory, m is a syllable counter section, and o is an auxiliary storage section. Agent Patent attorney Aihiko Fuku (and 2 others): $2 Figure

Claims (1)

[Claims] 1. Recognize the input voice syllable by syllable by calculating the similarity with pre-registered feature standard patterns of multiple types of syllables, and check the result with a dictionary or give external instructions such as a keyboard. In a Japanese speech input device that obtains the final input by modifying it through operations, the feature standard pattern obtained by analyzing the voice uttered at the time of input is used as the feature standard pattern of the syllable corrected by dictionary matching etc. A Japanese voice input device characterized by comprising a means for automatic registration. 2. The above-mentioned registration means includes means for registering only standard patterns with poor recognition rates obtained by a counter for obtaining recognition rates of syllable standard patterns. speech input device. 3. The Japanese language recited in claim 1, wherein the registration means includes means for storing the syllable standard pattern in the main memory in the auxiliary memory in response to an instruction from the outside. Voice input device.