JPH0338699A - Speech recognition device - Google Patents

Abstract

PURPOSE:To recognize speech in an extremely large vocabulary with good accuracy by providing plural recognition parts which differ in recognition unit and a display part, identifying an input voice in phoneme units, and matching the voice with a dictionary by the recognition parts according to the identification results and recognizing the voice in the recognition units corresponding to the respective recognition parts. CONSTITUTION:An analog voice signal which is inputted continuously to an analog input part 1 is inputted to a voice analysis part 2 to generate voice feature parameter. The feature parameter is identified in phoneme units by pattern matching, etc. Then a paragraph recognition part 8 matches a combination of syllable candidates in successive syllable recognition results with the contents of a paragraph recognition dictionary 9 to recognize the input voice in paragraph units. A word recognition part 10 matches dictionary items of a dictionary 11 for word recognition with a syllable candidate group in the successive syllable recognition results to recognize the input voice in word units. Namely, the speech is recognized in units of an extremely large vocabulary such as paragraphs.

Description

DETAILED DESCRIPTION OF THE INVENTION <Industrial Application Field> The present invention relates to a speech recognition device that recognizes input speech in meaningful units such as words or phrases.

<Prior Art> Conventionally, as a speech recognition device, there is a speech recognition device that uses phonemes as identification units. This speech recognition device that uses phonemes as the identification unit performs speech recognition processing (bottom-up processing) by combining the phoneme identification results, so it is possible to recognize very large amounts of speech, including conjugations, such as phrases. , suitable for text input.

However, when recognition processing is performed in units of clauses using the speech recognition device that uses phonemes as identification units, it is difficult to increase the accuracy of clause recognition beyond a certain level. Therefore, attempts have been made to perform a recognition operation using words as a recognition unit in parallel with a recognition operation using a phoneme as a recognition unit, and to recognize phrases based on the results of both types of recognition. "Frequent Word Spotting" Proceedings of the 1986 Autumn Research Conference of the Acoustical Society of Japan 2-31O1 JP-A-153595, etc.).

<Problems to be Solved by the Invention> However, in a speech recognition device that performs the above-mentioned phoneme-based identification and word-based identification in parallel, it is necessary to perform an identification operation using words as identification units. However, even though it takes time to identify words when words are used as identification units, the words that can be identified are at most i, ooo
It is about a word.

Moreover, it is necessary to utter all the words to be identified once and then register a standard word pattern determined from the characteristic parameters of the word in the dictionary, which is inefficient.

SUMMARY OF THE INVENTION Accordingly, an object of the present invention is to provide a speech recognition device that can efficiently and accurately recognize super-power-supplied speech using phonemes as identification units.

<Means for Solving the Problems> In order to achieve the above object, the present invention acoustically separates a human-generated speech signal, converts it into feature parameters, and identifies the feature parameters in phoneme units using a predetermined method. In the speech recognition device that recognizes speech based on the identification result, a recognition unit recognizes the input speech by comparing it with a dictionary in meaningful recognition units such as words or phrases based on the identification result of the phoneme unit. The present invention is characterized in that it includes a plurality of recognition units each having a different recognition unit, and a display unit that displays the recognition results of each recognition unit, so that recognition processing of different recognition units can be used together.

Further, in the speech recognition device, it is preferable that the dictionaries used by each of the recognition units to recognize speech be configured to correspond to the recognition unit of each recognition unit, and provided for each recognition unit.

Furthermore, when performing recognition processing in the voice recognition device, it is desirable to provide a switching means for switching and selecting the recognition section to be operated.

Furthermore, the display means in the speech recognition device is divided into a plurality of areas, each area having one window for displaying recognition candidates by different recognition units, and selecting a desired recognition candidate from among the displayed recognition candidates. It is desirable that the recognition candidates can be selected by the recognition candidate selection means.

Furthermore, the display means in the speech recognition device has a plurality of windows each displaying recognition candidates by different recognition units, and a window selection means selects a desired window from among the plurality of windows. It is desirable that a desired recognition candidate be selected from among the recognition candidates displayed in the displayed window by the recognition candidate selection means.

Function> When an audio signal is input, the input audio signal is acoustically analyzed and converted into feature parameters, and these feature parameters are identified in units of phonemes by pattern matching or the like. Then, the identification results identified in phoneme units are input to a plurality of recognition units, and based on the above identification results, the recognition units that have meaning, such as words or phrases, according to each recognition unit are checked against a dictionary, etc. Human voice is recognized. Then, the recognition results of each recognition section are displayed on the display section.

Rather, it is possible to recognize speech by using recognition processing of different recognition units by a plurality of recognition units.

Further, in the above-mentioned speech recognition device, a dictionary used when recognizing speech in each recognition unit is configured to correspond to the recognition unit of each recognition unit and provided for each recognition KS, and the recognition unit to be operated is switched. If it is possible to switch between the recognition units manually, it is possible to recognize the voice more efficiently by combining the recognition units according to the content of the voice.

In addition, the above-mentioned speech recognition device divides a window opened into a plurality of areas into a plurality of areas, displays recognition candidates by different recognition units in each area, and selects a desired recognition from among the displayed recognition candidates. If the candidates can be selected by the recognition candidate selection means, the correct recognition candidate can be easily selected.

Further, the speech recognition device displays recognition candidates by different recognition units in each of the plurality of windows opened on the display means, and selects a desired window from among the plurality of opened windows by the window selection means. If the desired recognition candidate can be selected by the recognition candidate means from among the recognition candidates displayed in the selected window, the correct recognition candidate can be selected more easily and efficiently. .

<Example> Hereinafter, the present invention will be explained in detail with reference to illustrated embodiments.

This invention segments and identifies continuously uttered speech into syllable units, and performs recognition processing using two types of recognition units, word recognition processing and phrase recognition processing, based on the obtained syllable lattice. The results are used in conjunction with the phrase recognition results to recognize super-large dielectric speech.

FIG. 1 is a block diagram showing an embodiment of the speech recognition device of the present invention. An analog audio signal continuously input to the analog input section 1 is amplified by an amplifier (AMP)+1, converted into a digital audio signal by an A/D converter 12, and outputted to the audio analysis section 2. The digital audio signal inputted from the A/D converter 12 to the audio analysis unit 2 is subjected to spectrum analysis in units of 16 m5 frames at a cycle of 8 ms, and acoustic feature parameters are generated. The generated feature parameters are outputted to the syllable section extractor 3 together with information necessary for extracting the syllable section (power, number of zero crossings, etc.).

Then, the feature parameters manually entered from the speech analysis section 2 to the syllable section extraction section 3 are extracted in syllable units based on the information necessary for extracting syllable sections. Then, the syllable characteristic pattern cut out in syllable units is stored in the characteristic pattern memory 7I of the pattern memory 7 together with syllable interval information (syllable length, silent time length between the syllable immediately before/after the syllable, etc.). The pattern memory 7 is the feature pattern memory 71 that stores only one syllable feature pattern for l syllables extracted by the syllable section extraction unit 3 as described above.
and a standard pattern memory 72 that stores a plurality of standard patterns each consisting of acoustic feature parameters of standard syllables created in advance.

Stored in the feature pattern memory 7I! The similarity calculation between the syllable characteristic pattern of the syllable and the standard pattern stored in the standard pattern memory 72 is performed by the monosyllable recognition unit 6, and the monosyllable extracted by the syllable segment extraction unit 3 based on the calculation result is be identified. Then, the phrase recognition section 8 compares the combination of syllable candidates in the continuous syllable recognition results (syllable lattice) with the contents of the phrase recognition dictionary 9, and the input speech is recognized in units of phrases. On the other hand, the word recognition unit IO compares the dictionary items of the word recognition dictionary 11 with a group of syllable candidates in the consecutive syllable recognition results (syllable lattice).
The system recognizes human speech on a word-by-word basis. In other words, speech is recognized in ultra-large dielectric units such as phrases. Then, the recognition results (phrase candidates) by the phrase recognition unit 8 and the recognition results (word candidates) by the word recognition unit 10
is displayed on the display section 12. Then, the operator operates the keyboard 13 to select and output the correct recognition candidate from the phrase candidates and word candidates displayed on the display section 12.

The CPU (central processing unit) 5 includes the analog input section 1, the voice analysis section 2. Syllable interval extraction unit 3. Monosyllable recognition unit 6. Pattern memory 7, clause recognition unit 8. Word recognition unit IO1 controls display unit 12 and keyboard 13 to execute speech recognition processing operations. Further, the CPU memory 4 stores programs for the CPU 5 and information necessary for speech recognition processing.

FIG. 2 is a flowchart of the speech recognition processing operation executed under the control of the CPU 5. As shown in FIG.

Hereinafter, the speech recognition processing operation will be explained in detail with reference to FIG.

In step SL, the feature parameters generated by the speech analysis section 2 are extracted into syllable units by the syllable section extraction section 3. At that time, if a plurality of syllable section candidates exist, a plurality of syllable section candidates are extracted.

There are various methods for cutting out the syllable section, for example, as follows. That is, syllable boundaries are detected from the power dip and the change point of the spectrum accompanied by an increase or decrease in power.

If the number of syllables included in the section sandwiched by the detected syllable boundaries is more than twice the minimum value of the estimated average frame length, the above section is regarded as a plurality of syllable sections, and the spectral change point within the section is A plurality of syllable section candidates are cut out based on (FIG. 3).

At step S2, the syllable section candidates cut out at step SL are expanded to generate a plurality of syllable section strings (FIG. 4). At this time, each syllable interval sequence is given a reliability based on the certainty of each syllable boundary candidate.

In step S3, the characteristic pattern of one syllable interval of the plurality of syllable interval sequences generated in step S2 is stored in the characteristic pattern memory 71, and monosyllable recognition is performed by the monosyllable recognition unit 6 as described above. syllable candidates are output. In this way, the plurality of syllable interval sequences are recognized for each syllable interval.

In step S4, a syllable lattice is generated for each syllable interval string as a recognition result of the monosyllable recognition in step S3. At this time, each syllable candidate in the syllable lattice is given a degree of reliability based on its degree of similarity to the standard pattern, and the syllable candidates are arranged in descending order of degree of reliability (FIG. 5).

In step S5, based on the plurality of syllable lattices for each syllable interval sequence generated in step S4, the syllable lattice corresponding to the syllable interval sequence with the highest reliability (hereinafter referred to as the first syllable lattice) is determined. ) is sent to the phrase recognition unit 8. At the same time, execution of clause processing is instructed.

Then, in the phrase recognition unit 8, the first
A character string is created by combining the syllable candidates that make up the syllable lattice, and linguistic processing such as checking with the phrase recognition dictionary 9 is performed, and only grammatically valid character strings are recognized as (
The phrase candidates are output together with the cumulative reliability in the order of the cumulative reliability (Fig. 6). In other words, the phrase recognition process can be said to be a bottom-up process.

In step S6, all of the plurality of syllable lattices generated in step S4 are sent to the word recognition unit IO. The word recognition dictionary 11 is then compared and word candidates are output.

The above dictionary matching is performed as follows. In other words, it is determined for each dictionary item whether or not there is a syllable lattice in which all the syllables that make up the dictionary item (word) in the word recognition dictionary 2 are present in the group of syllable candidates that make up the syllable lattice. It is determined that As a result, if such a syllable lattice exists, its dictionary entry (word) is output as a recognition result (word candidate) along with its cumulative reliability in the order of the cumulative reliability (FIG. 7). In other words, s-word processing can be said to be top-down processing.

In step S7, the clause candidates output in step S5 and the word candidates output in step S6 are displayed on the display unit 12 (FIG. 8 or 9).

In step S8, with reference to the display contents in step S7, a correct recognition candidate corresponding to the human voice is selected from among the phrase candidates and word candidates using the keyboard 13, the recognition candidate is confirmed, and the speech recognition processing operation is performed. finish.

Next, the above-mentioned speech recognition processing operation will be explained in more detail with reference to FIGS. 1 to 9, taking as an example the case where the word "Haikei" is input to the analog input section 1.

FIG. 3 shows the syllable segment candidates extracted by the syllable segment extraction unit 3 together with the utterance content and phoneme symbols. The above-mentioned phoneme symbols are added to show the change in voice in an easy-to-understand manner, and are not directly related to this invention. In this case, in the human voice "yes", the syllable /ha/ is uniquely determined as the syllable interval candidate SO, and the syllable /i/ is uniquely determined as the syllable interval candidate Sl. However, in the input speech "kei", the syllable string /kei/ is determined as the syllable interval candidate S2, and the syllable /ke/ is determined as the syllable interval candidate S3, and the syllable /i/ is determined as the syllable interval candidate S2. 2 with the case decided in S4
Syllable section candidates for the street have been created.

...Step S1 in Fig. 2 The syllable interval candidates created as described above are expanded and the fourth
Generate a syllable interval sequence as shown in the figure.

In this case, two syllable interval sequences are generated: syllable interval sequence 0 consisting of syllable interval candidate sequence 5O-9L-92 and syllable interval sequence 1 consisting of syllable interval candidate sequence 5O-9l-83S4. In this case, as mentioned above, reliability is added to the syllable interval sequence, and the reliability added to syllable interval sequence 0 is higher than the reliability added to syllable interval sequence 1. .

...Step S2 in Figure 2 Each syllable interval candidate SO, St of the syllable interval sequence O shown in Figure 4
, S2 are monosyllable recognized by the monosyllable recognition unit 6, and a G syllable lattice is generated as shown in FIG. 5 based on the monosyllable recognition results. In this case, the syllable lattice based on the syllable interval sequence O is designated as syllable lattice 0, and the syllable lattice based on the syllable interval sequence 1 is designated as syllable lattice 1.

...Step S3 in FIG. Step S4 The syllable lattice 0, which is the first syllable lattice, is sent to the phrase recognition unit 8, and each syllable section candidate S O making up the syllable lattice 0 is sent.
, S 1 and S 2 are selected to create a character string, as will be detailed later. That is,
First, each syllable interval candidate SO, S1. The syllable candidate with the highest reliability in S2 is selected, and the character string "Haite" with the highest cumulative reliability is created. Thereafter, each character of the created character string is sequentially replaced with other syllable candidates to create a character string. Then, the phrase recognition dictionary 9 is checked to remove character strings that are not in the phrase recognition dictionary 9 and character strings that are not grammatically valid, and phrase candidates as shown in FIG. 6 are obtained. The candidate ranks in the figure are in descending order of the cumulative reliability of the syllable candidate strings that constitute the phrase candidate.
...Step S5 Furthermore, all the generated syllable lattices, that is, syllable lattice O and syllable lattice L
is sent to the word recognition unit 10, and the word recognition dictionary 11 is collated as described above to execute word recognition. In this case, the dictionary entry "Haikei" in the word recognition dictionary ll is used.
All of the syllables /ha/, /i/, /ke/, and /i/ that make up ``Haikei'' are in the syllable candidate group of syllable lattice 1, so ``Haikei'' is a word candidate in syllable lattice number 1. Similarly, all the syllables that make up the dictionary entry "Kite" are in the syllable candidate group of syllable lattice O, so the word candidate "Kite" is obtained with syllable lattice number "0".
”. In the same manner, word candidate character strings and syllable lattice numbers are obtained by associating them.
Word candidates as shown in the figure are obtained. The candidate ranks in the figure are in descending order of cumulative reliability of the syllable candidate strings that constitute the word candidate.

. . . The recognition candidates of phrase candidates and word candidates obtained in step S are displayed on the display unit 12. At that time, there are two display methods as shown below.

One method is to divide one window 21 opened on the display section I2 into two areas as shown in FIG.
By displaying word candidates in one area and phrase candidates in the other area, both types of recognition can be displayed simultaneously within one window 21. Another method is to open two windows 22 and 23 on the display unit 12, as shown in FIG. 9, and display word candidates in one window 22 and phrase candidates in the other window 23. be.

...Step S7 Then, in accordance with the display contents of FIG. 8 or FIG. 9, recognition candidates are selected in the following manner, for example.

That is, keys such as "↓" and "↑" on the keyboard 13 are used to select recognition candidates within one window, and keys such as "-" and "←" are used to select a window.

...Step S8 In this way, by using the phrase recognition process and the word recognition process together, it is possible to recognize a super large amount of phrases, etc., and at the same time, it is possible to recognize phrases that could not be recognized by the phrase recognition unit 8. Since the word "haikei" is recognized by the word recognition unit IO, a desired recognition result can be obtained by selection using the keyboard 13.

That is, in this embodiment, the syllable segment extraction unit 3 generates a syllable segment sequence based on the audio signal of the input voice. Then, for each generated syllable segment string, monosyllable recognition is performed by the monosyllable recognition unit 6 to generate a syllable lattice corresponding to each syllable segment string. Thereafter, the phrase recognition unit 8 executes phrase recognition processing for the syllable lattice with the highest reliability, and also executes word recognition processing for all syllable lattices. Then, the display section 12 displays the phrase candidates obtained by the phrase recognition process and the word candidates obtained by the word recognition process, and a desired recognition result is selected using the keyboard 13 with reference to the displayed results. There is.

In this way, by using the phrase recognition process by the phrase recognition unit 8 and the word recognition process by the word recognition unit 10 in combination, the following effects can be obtained. That is, by registering frequently occurring words in the word recognition dictionary 11, the frequently occurring words registered in the word recognizing dictionary 11 can be recognized with high accuracy through word recognition processing. On the other hand, words or phrases that are not registered in the word recognition dictionary 11 can be efficiently recognized by phrase recognition processing.

In other words, by using bottom-up phrase recognition processing and top-down word recognition processing together, it is possible to recognize a very large amount of speech with a balance between accuracy and efficiency. At this time, since the identification unit is a syllable as described above, the standard pattern can be registered more efficiently than when words are used as the identification unit.

That is, according to this embodiment, since syllables are used as identification units, speech can be efficiently identified. Furthermore, since recognition processing is performed on a word-by-word basis and on a phrase-by-phrase basis, it is possible to efficiently and accurately recognize speech using ultra-high power supply.

In the above embodiment, the syllable is used as the unit of speech identification. However, the present invention is not limited to this, and phonemes, semisyllables, etc. may also be used. That is, a so-called phoneme, which is a general term for phonemes, half-syllables, syllables, etc., may be used as the identification unit.

In the above embodiment, phrase recognition processing is performed on the syllable lattice with the highest reliability, and word recognition processing is performed on all syllable lattices. However, the invention is not limited thereto. The point is that it is sufficient if a plurality of recognition processes of different recognition units can be used together.

In the above embodiment, when displaying phrase candidates and word candidates on the display unit 12, they are displayed in kana characters, but they may be displayed in kanji by performing kana-kanji conversion.

In the above embodiment, the phrase recognition process and the word recognition process are executed simultaneously. However, the present invention is not limited to this, and a switching means may be used to switch between executing both recognition processes simultaneously or executing only one of the recognition processes. By doing this, for example, if only place names are registered in the word recognition dictionary 11, when manually writing general sentences that do not include place names, you can select only phrase recognition processing. By preventing unnecessary recognition candidates from being displayed on the display section (2), speech recognition can be performed more efficiently.

<Effects of the Invention> As is clear from the above, the speech recognition device of the present invention has the following effects:
Equipped with multiple recognition units and display units with different recognition units,
The input speech is identified in phoneme units, and based on the identification results, the plurality of recognition units mentioned above compare the dictionaries and recognize the speech in recognition units according to each recognition unit, and the recognition results are displayed on the display unit. Since it is displayed, recognition processing for different recognition units can be used OF. Therefore, in addition to identifying input speech in units of phonemes, by using multiple recognition processes in different recognition stages (stands) in combination, we can utilize the recognition characteristics of each recognition unit to efficiently and accurately generate a large amount of speech. I can recognize it.

Further, in the speech recognition device of the present invention, the dictionary used by each of the recognition units when recognizing speech is configured to correspond to the recognition unit of each recognition unit, and is provided for each recognition unit, and is operated. Since the recognition unit can be switched and selected by the switching means, speech recognition can be performed only by recognition processing according to the content of input speech, and speech can be recognized more efficiently.

Further, in the speech recognition device of the present invention, one window opened on the display section is divided into a plurality of regions, and recognition candidates by different recognition sections are displayed in each region, and among the displayed recognition candidates. Since the desired recognition candidate can be selected by the recognition candidate selection means, the correct recognition candidate can be selected from among the displayed recognition candidates, and the correct recognition/candidate can be easily selected.

Further, the speech recognition device of the present invention displays recognition candidates by different recognition units in a plurality of windows opened on the display unit, and selects a desired window from among the plurality of windows by a window selection means, Moreover, since the desired recognition candidate can be selected by the recognition candidate selection means from among the recognition candidates displayed in the selected window, the correct recognition candidate can be selected more easily.

[Brief explanation of drawings]

Fig. 1 is a block diagram of an embodiment of the speech recognition device of the present invention, Fig. 2 is a flowchart of speech recognition processing operation, Fig. 3 is a diagram showing an example of extracted syllable section candidates, and Fig. 4 is a generated 5 is a diagram showing an example of a generated syllable lattice, FIG. 6 is a diagram showing an example of the obtained phrase candidates, and FIG. 7 is a diagram showing the obtained word candidates. FIG. 8 is a diagram showing an example of recognition candidate display. FIG. 9 is a diagram showing an example of recognition candidate display.
The figure is a diagram showing another example of recognition candidate display. l...analog input section, 2...speech analysis section, 3...
-Syllable section extraction unit, 5...CPU, 6...Single syllable recognition unit, 7...Pattern memory, 8...
・Phrase recognition unit, 9...Phrase recognition dictionary, l
O...Word recognition unit, 11...Word recognition dictionary, 12...Display unit, 13...Keyboard.

Claims (5)

[Claims] (1) In a speech recognition device that acoustically analyzes the input speech signal and converts it into feature parameters, identifies the feature parameters in phoneme units using a predetermined method, and recognizes speech based on the identification results, the phoneme A recognition unit that recognizes input speech by checking a dictionary with meaningful recognition units such as words or phrases based on the unit identification results, and each recognition unit includes a plurality of recognition units each having a different recognition unit; What is claimed is: 1. A speech recognition device, comprising: a display section that displays recognition results obtained by the speech recognition system, and enables recognition processing of different recognition units to be used together. (2) The dictionary used by each recognition unit when recognizing speech is configured to correspond to the recognition unit of each recognition unit, and is provided for each recognition unit. Speech recognition device. (3) The speech recognition device according to claim 2, further comprising a switching means for switching and selecting the recognition unit to be operated when performing recognition processing. (4) The display means is divided into a plurality of areas, each area having one window for displaying recognition candidates by different recognition units, and recognizing a desired recognition candidate from among the displayed recognition candidates. 4. The speech recognition device according to claim 1, wherein the speech recognition device is made selectable by candidate selection means. (5) The display means has a plurality of windows each displaying recognition candidates by different recognition units, and the window selection means selects a desired window from among the plurality of windows, and displays the selected window. 4. The speech recognition device according to claim 1, wherein a desired recognition candidate can be selected from among the displayed recognition candidates by recognition candidate selection means.