JPS63316900A - Voice recognition system - 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 (a) Field of Industrial Application The present invention relates to a speech recognition system provided with automatic gain control of input speech sound pressure for speech recognition.

(B) Conventional technology In conventional speech recognition systems, in order to avoid the inconvenience of not being able to accurately extract the characteristics of the voice due to excessive sound pressure of the input voice, AGC (automatic gain control) is installed after the input means such as a microphone. ) circuit was installed.

(c) Problems to be solved by the invention Conventional AGC circuits feed back the output from the amplifier, and perform AGC based on the sound pressure at the moment of input, so high sound pressure audio is input. In this case, the AGC adjusts the gain to set the sound pressure level to the optimal level for voice recognition, but since this adjustment is only performed after the voice is input, there may be a time lag between the input and the adjustment. This causes problems such as not being able to perform optimal level adjustment, hindering voice feature extraction, and potentially lowering the recognition rate.

(d) Means for solving the problems The speech recognition system of the present invention combines a speech recognition device and a recording/playback device, and adds an automatic gain control device equipped with a memory unit as shown in FIG. This allows the same sentence to be input twice from the recording/playback device, recording the fluctuations in sound pressure during the first input, and adjusting the sound pressure of the human voice during the second playback.

(e) Effect: According to the system of the present invention, by combining a speech recognition device, a recording/playback device, a memory ms, and an automatic gain control means, it is possible to manually produce the same sentence twice without any fluctuation in utterance. Therefore, the sound pressure fluctuations are recorded in memory during the first input, and the gain is adjusted based on the sound pressure data in the memory so that the sound pressure of the input sound becomes suitable for speech recognition processing during the second playback. can be adjusted automatically.

That is, as shown in FIG. 18, the first input voice (S)
The fluctuation state of the sound pressure level V is detected from the sound pressure level V and stored in the memory.Then, for the second input sound (S), the sound pressure level in the memory is detected, for example, a low level (L) or a high level (
H), the control voltage Va of the automatic gain control means is set to a high level (Vh) in inverse proportion to this level value.
, set to be a low level (V!), and set the gain G-AX
VG, (A is a constant) is multiplied by a gain. Therefore, as shown in the output audio (S) in the figure, gain control can be performed substantially in real time without excessive levels.

(f) Embodiment FIG. 1 shows an external view of a dictionary-cheating machine which employs the present invention and creates sentences by voice input, and FIG. 2 shows a functional block diagram of the machine.

In Figure 2, (1) is the 1st t! This is a voice recognition section equipped with a circuit in the main body (100) of the I, the details of which are shown in the block diagram of FIG. ], a feature extraction unit (12) [Fig. 5] that extracts parameters representing the acoustic characteristics from the sound pressure-attenuated audio signal from the processing unit (11), and feature parameters obtained from the extraction unit (12). A word recognition unit (13) [Figure 6] and a phrase recognition unit (14) [Figure 7] that perform word recognition of input speech based on the following, and recognition from any of these recognition units (13) and (14) It consists of a candidate creation section (15) that creates candidates for recognized word character strings or recognized syllable characters based on the results.

Furthermore, in Figure 2, (2) is the main body (1) as shown in Figure 1.
00) is a mechanically and electrically detachable recording and reproducing device such as a tape recorder, (3) is a headphone type microphone as shown in FIG. 1, and (4) is a recording and reproducing device (2) and a microphone (3). ) and the voice recognition unit (4) [Figure 8]
It is. (6) is a display device for displaying character strings etc. generated based on the recognition results; (7) is a keyboard for inputting various control signals of the dictionary cheating machine;
(8) is a storage device such as a magnetic disk device that stores the character strings generated by the dictionary cheating machine, and (9) is a speaker (10) that uses the character strings in the storage device by regular synthesis.
This is a speech synthesis unit that reads out the text.

Note that (5) is a control section consisting of a microprocessor, which controls the operations of the above-mentioned sections.

There are two ways to create sentences using the dictating machine configured as described above, and each will be explained in detail below.

The first method is to collect all voices from the microphone (3) into the voice recognition unit (
1), performs voice recognition, converts the input voice into a character string, displays it on the display device (6), and at the same time inputs it into the storage device (8).
).

The second method is to record the text you want to input in advance using a recording and playback device (2
＞, connect this recording and playback device (2) to this device, input the recorded text to the voice recognition unit (1), perform voice recognition, convert the input voice into a character string,
The results are displayed on the display device (6) and simultaneously stored in the storage device (8).

As mentioned above, there are two ways to input audio.
The input switching section (4) switches the input. In addition to switching the input, the input switching section (4) also switches whether the recording signal (a) is to be recorded on the recording/playback device (2) or the microphone 3. > You can also switch whether to record the input audio.

The operations from voice recording to text creation will be explained in detail below.

(i) f-Voice Registration Process Before performing voice recognition, voice registration is performed in order to create a standard pattern of voices necessary for voice recognition.

First, the syllable registration mode will be described.

The standard pattern described here is the speech recognition unit (1)
This is the reference pattern for pattern matching in the phrase recognition unit (14) of
).

The method for registering voice in this dictionary-cheating machine is to first operate the switch (14sl) shown in Fig. 7 to connect the parameter buffer (14a) and the syllable recognition unit (14b).
There are three methods described below.

The first method is to attach a microphone (3) to the main body (100) of the machine.
A method of directly inputting registered speech, analyzing this registered speech with the speech recognition unit (1), creating a standard pattern, and storing the created standard pattern in the syllable standard pattern memory (14d) and storage device (8). It is.

The second method is to connect a recording/playback device (2) on which registered voices have been recorded in advance to the main body (100), and to listen to this recording 'f.
lInput the registered voice by playing the recorded voice,
This input registered speech is analyzed by the speech recognition unit (1), a standard pattern is created, and the created standard pattern is stored in the syllable standard pattern memory (14d) and the storage device (8).

The third method is to connect the microphone (3) to the main body (100) of this machine.
At this time, the recording/playback device (2) is connected to the main unit (100), and while recording the input audio to the recording/playback device (2), the registered audio is input directly from the main unit (100) side. Now, analyze the input gs voice from the microphone (3), create a standard pattern, store the created standard pattern in the storage device (8), and then input the voice to this microphone (3). Once finished, the audio recorded on the recording/playback device (2) will be played back,
This recorded registered voice is analyzed by the voice recognition unit (1),
At the same time as creating a standard pattern and storing the created standard pattern in the syllable standard pattern memory (14d),
This is a method in which the syllable standard pattern of the directly registered voice from the above-mentioned microphone (3) is also stored in the storage device (8).

In this third method, the sound recorded on the recording/playback device (2) is subject to the frequency characteristics of the recording/playback device (2), so the standard pattern created from the recorded sound and the microphone (3) When comparing the internal standard pattern with the standard pattern created from the voice input directly from the internal standard pattern, differences appear between the internal standard patterns. Therefore, when recognizing a recorded voice, it is necessary to use a standard pattern created from the recorded voice, and the microphone (3)
When you want to recognize the voice input directly from the microphone (3)
), it is necessary to use the standard pattern created from the voice input directly from the microphone (3), so by using the method described above, both the standard pattern registered directly from the microphone (3) and the standard pattern created from the recorded voice can be used. can be created and stored with a single voice registration operation. In addition, once the registered voice is recorded on the recording/playback device (2), the recorded voice can be played back and input on the dictionary cheating machine for which no standard pattern has been created, without requiring the registrant's voice input. You can create a standard pattern. Also, if you record the registered voice on the recording/playback device (2) and then record the text after this registered voice, you can connect this recording/playback device (2) to the main unit (100) later and listen to the recorded voice. Just by playing the audio, everything from registering the audio to creating sentences can be done automatically.

Note that the registrant's voice input for creating the standard pattern of a-voice is performed by the registrant reading out the characters displayed by the present device on the display device 6) in a fixed order.

In addition, a recording and playback device (2
), this recording/playback device (2) can be carried alone, by uttering the voice corresponding to the entry word displayed on the display screen and recording it on the recording/playback device (2), the standard pattern can be reproduced. It becomes possible to create.

As mentioned above, when recording the registered voice for creating a standard pattern into the recording/playback device (2), when creating the standard pattern from the recorded registered voice, the recorded voice may be affected by noise etc. There is a possibility that the corresponding headword may be misaligned, and for the sake of explanation, a case will be described below based on FIG. 9 in which a table fooder is used as the recording and reproducing device. Figure 9(a) shows the state in which the registered voice for standard pattern creation is recorded on the table fooder.
Registered voice “a” ~1 corresponding to headword 1a”~1ka”
This shows the state of the tape between "ka" and "noise" is recorded between "e" and "o".
As shown in Figure 9(a), when voice registration is performed using a tape in which [noise] is completely recorded between the registered voices, the first recorded injury is “a” and the second recorded voice is “a”. If you determine which syllable the input registered voice corresponds to simply by the order of the sounds recorded on the tape, such as ``the first sound is the first one'', [Noiseco will not consider it as registered voice]. Since the headwords are made to correspond to each other, the actual registered voice input and the headwords are misaligned.

Here, FIG. 9(b) incorrectly recognizes [noise] as voice,
This is a prisoner who entered the syllable "e" in place of the entry word "e" and entered the syllable "e" in place of the entry word "o".

In this way, when creating a standard pattern from registered voices, the recorded voice and the headword may deviate due to the influence of noise, etc., so as shown in Figure 9(c), The character code sound indicated by ξ is recorded in the recording/playback device (2) in correspondence with the registered sound ξ. With this method, even if [noise] is recorded between "u" and "e", the gap between the input sound and the headword is prevented as described above.

A method of recording a character-code sound of a specific frequency to prevent this deviation will be explained separately for the case where the tape recorder of the recording/reproducing device (2) is a single track and the case where the tape recorder is a multi-track.

First, with reference to FIG. 10, a case will be described in which a recording/playback device with multi-track is used as the recording method.

When a multi-track recording/playback device is used as a recording method, the character food corresponding to the headword is recorded on a track on which audio is not recorded, as shown in FIG. 2(a). The voice recognition unit (1) knows the headword of the input voice from this character-code sound, and also determines whether the character-code sound is recorded in the section t1 of the sounds recorded on the audio track. Among the sounds, only those that satisfy the condition of being equal to or higher than the sound pressure threshold are considered to be sounds and analyzed.

Alternatively, as shown in FIG. 2(b), a character code corresponding to the headword is recorded at the beginning and end of the audio, and the character code that indicates the beginning of the audio is selected from among the sounds recorded on the audio track. Among the sounds recorded in the interval t2 between the character sound and the character sound indicating the end of the sound,
Only those that satisfy the condition of being equal to or higher than the sound pressure threshold are regarded as voices and analyzed.

Or same 1! ! As shown in J(c), the character code corresponding to the headword is recorded at the beginning of the audio.

Among the sounds recorded on the audio track, among the sounds recorded in the section t3 from the character-code sound indicating the bridge of this sound to the character-code sound corresponding to the next headword, the sound pressure threshold Only those that meet the above conditions are considered to be audio and analyzed.

The second method, in the case of a single-track recording/playback device (2), is to represent the character code corresponding to the headword as a sound outside the audio analysis frequency band, and then add the audio to the track where the audio is recorded. record with. The method for recording character chord sounds in this case is the same as in the multi-track case described above. In other words, the above tl, t
2. Among the sounds recorded in the interval t3, only those that satisfy the same conditions as described above are regarded as voices and analyzed.

However, except for the example shown in Figure 1 (1) where the voice and the character code sound overlap, the character code sound overlaps with the character code sound.
It is not necessary to use a sound outside the voice analysis frequency band as a chord sound.

Next, alphabets, numbers, parentheses, punctuation marks, etc. are preliminarily stored in the word dictionary (13) of the word recognition unit (13) as shown in Figure 6.
The word standard pattern corresponding to the word whose character has been registered in d) is stored in the word standard pattern memory (13) in the same figure.
c) Register.

First, by performing a predetermined operation, the 61st! I parameter buffer (13a) and word standard pattern memory (13d)
is connected by a switch (13sl) to put it into word registration mode.

Next, alphabets, numbers, parentheses, punctuation marks, etc. are displayed on the display device f(6) of the main body (100) of the apparatus, and the operator inputs the corresponding pronunciation by voice.

The speech recognition section (1) analyzes this speech and registers the word standard pattern in the word standard pattern memory (13c).

Voice recognition becomes possible through the operations described above.

However, if you want to recognize a word that is not in the independent word/attached avJ dictionary (14e) or the word dictionary (13d), either register the word you want to recognize in the independent word/attached word dictionary (14e), or register the word you want to recognize in the independent word/attached word dictionary (13d). ) It is necessary to register the word to be recognized by the word standard pattern memory (13c). However, independent word/attached word dictionary (1
Whether to register words in 4e) or to register words and word standard patterns in the word dictionary (13d) and word standard pattern memory (13c) depends on whether the user wants the word to be recognized as a clause utterance, Determine whether you want it to be recognized as a vocalization.

In addition, if the word is in the independent word/adjunct word dictionary (14e) but not in the word dictionary (13d) and you still want to recognize it with word recognition, you can add the word to the word dictionary (13d) and word standard pattern memory (13e). , it is necessary to register words and word standard patterns.

The method for registering arbitrary words will be described below.

To register salivary words, use the independent word/adjunct word dictionary (14e).
There are two methods: registering a character string in a word standard pattern memory (13c), and registering a character string in a word dictionary (13d).

When registering a word in the independent word/adjunct word dictionary (14e), the word to be registered is uttered and input into the device.

At this time, this device recognizes this voice using the voice recognition section (1) and displays the recognition result on the display device (6). If the result is correct, the user presses a predetermined key on the keyboard (7) and registers the uttered voice as a character string displayed on the display device (6) in the independent word/adjunct word dictionary (14e). if,
If the recognition result displayed on the display device (6) is incorrect, use the syllable correction function of this device to correct the recognition result displayed on the display device (6>), or re-speak the word you want to register. If the result is wrong, correct it again using the syllable correction function of this device.The above-mentioned operation is repeated until the character string displayed on the display device (6) matches the word to be registered.

When registering a word in the word standard pattern memory (13c) and the word dictionary (13d), first the display device (
6) Correctly display the character string you want to register in 6) Next, display the correctly recognized character string and word standard pattern in the word dictionary (
13d) and word standard pattern memory (13c).

In addition, recognizing a voice that has been input in a natural manner is
Considering the current level of speech recognition technology, this is unreasonable. Since the current level of speech recognition technology is limited to continuous syllable utterance input, an example of continuous syllable utterance input will be described below.

In the case of continuous syllable vocalization input, the above procedure is the same, but in the case of continuous syllable vocalization input, the word standard pattern is also a continuous syllable vocalization pattern, so the word you want to register can be re-uttered naturally. , a word standard pattern is created by natural utterance, and the word standard pattern and character string are registered in a word standard pattern memory (13C) and a word dictionary (13d), respectively.

With the above operations, the data necessary for creating sentences using voice recognition has been registered.

(i) Text creation An example of text creation will be described below.

First, when performing a recognition operation, the switch (13sl) of the word recognition unit (13) is set to the parameter buffer (13a).
) and the word determination section (13b), and the switch (14sl) of the phrase recognition section (14>) is set so as to connect the parameter buffer (14m) and the syllable recognition section (14b).

There are two ways to create sentences.

The first method is an online recognition method in which the text to be created is directly input into the main body of the device by voice from the microphone (3).

The second method is to use a recording/playback device that records the text (2).
This is an offline recognition method in which the device is connected to the device, the recorded text is played back, and the text is recognized.

First, an example of online recognition will be described.

In the case of online recognition, sentences uttered in units of phrases or words are directly input into this device from the microphone (3), so by performing a predetermined operation, the microphone (3) is connected to the voice recognition unit by the input switching unit (4). Connect (1).

Also, if you want to record the audio input from the microphone (3) on the recording/playback device (2),
) to the main unit, and connect the input switching section (4) to the microphone (3).
) and the recording terminal of the recording/playback device (2).

At the same time, as described later, the silence detection signal is transmitted to the feature extraction unit (1
2) If it is input, it functions to record a beep sound indicating a phrase or a win break.

During speech recognition, the word recognition unit (13) and phrase recognition unit (1
4) is running.

The audio input from the microphone (3) is processed by the preprocessing unit (11).
The input audio is processed to have characteristics suitable for audio analysis (for example, when the sound pressure of the input audio is low, the sound pressure is amplified using an amplifier), and the feature extraction unit (
12).

In the feature extraction unit (12), as shown in FIG.
).

At the same time, the analysis unit determination unit (llb) of the feature extraction unit (12)
), based on the analysis results of the analysis unit (l1m), the silent interval after uttering in syllables or phrases, and the beep sound recorded after uttering in phrases or words (details will be explained later in offline recognition). ) shown in the embodiment, and when a silent section is detected, a silent section detection signal (b) is generated.

The parameter buffer (12c) which receives the silent section detection signal (b) sends the stored feature parameters to the word recognition section (13) and phrase recognition section (14), and erases the stored contents.

The feature parameters input to the word recognition section (13) are stored in the parameter buffer (13m) shown in FIG. The word determination unit (13b) compares the feature parameters stored in the parameter buffer (13a) with the @word standard pattern memory (13c), and selects the feature parameters stored in the parameter buffer (13a) and those with a large likelihood. Words with word standard patterns are added to the word dictionary (13d).
A plurality of words are selected, and the character strings of the selected words and their likelihood values are sent to the candidate creation section (15).

On the other hand, the feature parameters input to the syllable recognition unit (14) are stored in the parameter buffer (14m).
The characteristic parameters stored in the syllable standard pattern memory (14d) are compared, and the characteristic parameters stored in the parameter buffer (13a) are compared.
A phrase determination unit (14c) converts the feature parameters stored in
In 4c), the input syllable string and the independent word/adjunct word dictionary (1
Compare the words registered in 4e), combine independent words and attached words to create multiple sets of phrases with a high likelihood, and send the character strings of the created phrases and their likelihood values to the candidate creation unit (15).
send to

The candidate selection section (15) selects a plurality of strings with a high likelihood from the input character strings, and selects the likelihood value and the data sent from the word recognition section (13) or the phrase recognition section (14). Add a hood to indicate the data that has been added and store it. at the same time,
A signal is sent to the control unit (5) to display the character string with the greatest likelihood on the display device. When the control unit (5) receives this signal, it adds a delimiter mark "1°" after the character string with the greatest likelihood, so that, for example, the input sentence in FIG. 9(a) is changed to the one in FIG. 9(b). At the same time, the candidate selection section (15) causes the control section (5) to display the content stored in the candidate selection section (15) on the storage device (8).
) sends a signal to be stored. The control unit (5) receives this signal and stores the character string stored in the candidate selection unit (15) in the storage device (8) with a code representing a delimiter added after the character string.
to be memorized. Since the character strings stored in this external storage device are used as the primary manuscript in a word processor, a floppy disk drive is generally used;
) files should match the file format of the word processor.

Further, the signal generating section (42) of the input switching section (4) shown in FIG. (41).The switch (41) controls the circuit so that the audio input from the microphone (3) and the beep sound input from the signal generator (42) are recorded in the recording/playback device (2). is connected to the recording and reproducing device (2), and a beep sound is recorded in the silent section considered to be a break between phrases or words of the recorded sentence.

Next, an example of offline recognition will be described.

For offline recognition, this device is equipped with a recording/playback device (2).
Since the text is created by inputting and reproducing the recorded voice, the text is first recorded on the recording/playback device (2).

In addition, in order to input audio from the recording and playback device (2),
The input switching section (4) connects the recording/reproducing device (2) and the speech recognition section (1).

When recording sentences, vocalize in phrase units or *g units to create silent intervals between phrases and words. In addition, when using a recording/playback device t(2) dedicated to this device as shown in Fig. 1, in order to clearly mark the divisions between phrases and words, the recording/playback device (2) or The separator key (
71) to record.

In addition, the registered word will be uttered in a single aS, but if the recording/playback device (2) has a character sound generation function and has a character corresponding to the word you want to input, the voice will be uttered instead. You may also record the character sounds.

In addition, signals indicating the beginning and end of a sentence are recorded together with the voice in order to prevent noise recorded between sentences at the beginning of each sentence from being mistakenly recognized as voice.

However, the method of recording this signal changes depending on whether the recording/reproducing device (2) is a multi-track system or not, as described in the audio registration section. FIG. 11 is a diagram of a multi-track system and a single-track system in which sounds outside the audio band are recorded together with audio. Figure 12 shows a single-track system in which sounds such as DTME signals outside the audio band are recorded before the beginning of a sentence, and then recorded again when the sentence ends.
This method assumes that a sentence is recorded between these two signals.

In addition, when recognizing a sentence, the sections t4 and t5 before and after the recording of the signal are sampled, and in order to determine whether or not it is voice, the beginning of the sentence and the beginning of the signal, and the end of the sentence and the end of the signal are not necessarily the same. It is not necessary that they match; therefore, recognition is possible even if the timing at which the sentence is uttered and the timing at which the key is pressed are slightly different.

Next, the recording/playback device (2) is connected to the main body of this device and the recorded voice is played back to perform recognition processing, but before recognizing this recorded voice, the recognition speed mode is set to a high playback speed of the recorded voice. Fast recognition mode #0 that shortens recognition time.
If you have time and require a high recognition rate, set the recognition mode to the double playback recognition mode.

First, an example of the fast listening recognition mode will be described.

In fast listening recognition mode, the playback speed of the recorded audio is increased, so the characteristics of the input audio are different from the standard pattern created from the registered audio played back at the normal playback speed. Even if you input speeded up audio,
Speech recognition cannot be performed accurately.

Therefore, in order to accurately recognize audio that has been played back at a faster speed, the sampling frequency is changed. Examples of such methods are described below.

The sampling frequency control unit (12d) of the feature extraction unit (12) in FIG. speed) to sample and analyze the audio. Feature extraction part (12
) The subsequent processing is the same as in the example for line recognition, except that the recording and playback device (2) adds a beep sound to indicate the break to clarify the break between phrases and words. When the feature extraction unit (12) detects the beep a, the feature extraction unit 12) generates the beep sound detection signal (mouth l) instead of the silent section detection signal (b).
occurs. When the received signal is not a silent section detection No. 13 (b) but a beep sound detection signal (mouth), the signal generating section (42) of the input switching section (4) detects the break between clauses or words of the sentence. No beep sound is generated.

In addition, the voice r2 recognition section (1) displays a character indicating a word.
When a sound is recognized, a word corresponding to the character-sound is output as a recognition result.

Next, an example of the twice playback recognition mode will be described.

In this mode, the recorded f1# voice is first played back and input to the device. At this time, the preprocessing section (11) of the speech recognition section (1) reads all the sound pressure fluctuations of the recorded voice, and stores this data in the back pressure fluctuation storage memory (llb) shown in FIG. Play the recorded audio and input it to this device. At this time, the preprocessing unit <11) uses the data stored in the sound pressure fluctuation memory (llb) to determine the input sound pressure to the feature extraction unit (12) as shown in FIG. Importantly adjust the amplification of the AGC circuit (lla) to match the appropriate level! That is, the gain G is set to a fixed gain A by the control voltage Va(
(variably adjusted).

Further, as another example of the twice playback recognition mode, a multiple playback recognition mode can also be considered. This involves manually replaying a recorded sentence many times, changing the recognition method in the speech recognition unit (1) each time, comparing the recognition results, and selecting the one with the greatest likelihood of certainty. This is the way to do it.

In addition, if the recording/playback device (2) does not record the audio for registration, and depending on the recording/playback device (2), the frequency characteristics when the playback speed is increased are different from those when the playback speed is normal. or to recognize sentences recorded on a recording/playback device (2) that has different frequency characteristics from the recording/playback device (2) used to create a standard voice pattern, or to create a standard voice pattern. It has the same frequency characteristics according to the standard as the Q sound reproduction device I (2), but due to the influence of errors in the parts used, the actual frequency characteristics are different from the recording and reproduction device (2) used to create the standard audio pattern. When the recorded text is recognized by the Dt playback device (2), the following function for correcting the influence of frequency characteristics is used.

First, a reference sine wave signal, which is a reference when measuring the frequency characteristics of the recording/playback device (2), is generated by the reference signal generator (42).
and record it on the recording/playback device (2). Thereafter, the recorded reference sine wave signal is manually reproduced into the apparatus. The input reference sine wave No. 13 is sent to the voice recognition unit (1).
analyzes and determines the difference in frequency characteristics between the recorded reference sine wave signal and the reference sine wave signal generated by the reference signal generator (42), and compares the recorded reference sine wave signal with the reference signal generator (42). Correction is applied to reduce the difference in frequency characteristics from the reference sine wave signal generated in (42). There are many ways to apply the correction, depending on the feature extraction method of the feature extraction unit (12) of the speech recognition unit (1)6. For example, as shown in FIG.
If an analog filter bank system is used, which consists of a parallel connection of an amplifier (PF) and an amplifier II (AMP), the reference signal generator (42) can be The output from the filter is u41! so as to reduce the difference in frequency characteristics from the generated reference sine wave signal. do. Also, if a digital filter is used as the feature extraction method of the feature extraction section (12),
In addition to changing the parameters that determine the characteristics of the digital filter, any other method can be considered depending on the feature extraction method of the feature extraction section (12) of the speech recognition section (1).

By the above operations, the voice input text has been converted into a string other than the text. A correction method when the kana string-converted text differs from the input text will be described below, using FIG. 14 and classifying each case into error. Correct it using the following steps.

Figure 14 (a) is the input text, Figure 14 (b) is the input voice,
FIG. 4(c) shows the recognition result, FIG. 2(d) to (h) show the correction process, and FIG. 11(1) shows the correction result.

First, what was uttered as a word was incorrectly uttered as a phrase i:2a
11, the correction method will be described below. As shown in Figure (C), the word ``C'' uttered is phrase 1.
If the cursor (X) is recognized as ``,'', first move the cursor (X) to the part of the incorrect word [Figure (d) il, then press the word next candidate key (72) to display the next candidate for the word. [Figure (d) il. If this result is correct, proceed to the next correction part. If this result is incorrect, press the word next candidate key (72) again to display the next word candidate. Repeat until the correct answer is displayed.

Next, we will discuss how to correct when something uttered as a phrase is misrecognized as a word. If the word uttered as “Two and” is recognized as the word “E”,
First, move the cursor (X) to the incorrect clause/\0
Next, press the phrase next candidate key (73) to display the next phrase candidate. If this result is correct, proceed to the next modification section.

If this result is incorrect, the phrase next candidate key (73)
Press to display the next phrase option. Repeat this operation until the correct answer is displayed.

It is also possible to display the previous candidate for a word by pressing the pre-word candidate key 74), and the previous candidate for each phrase by pressing the pre-phrase candidate key (75).

If the correct answer cannot be obtained using the above two correction methods, correction may be performed in units of syllables, or words, phrases, or syllables may be re-inputted.

In addition, in order to avoid recognizing phrases as words or recognizing words as phrases again during re-voice input, the candidate generation section (15) is configured to only recognize the recognition results sent from the word recognition section (13). As a recognition result, phrase recognition unit (14)
The recognition results sent from the computer can be controlled from the outside to be ignored.

In addition, the candidate generation unit (15) is configured to receive an external signal so that only the recognition results sent from the phrase recognition unit (14) are regarded as recognition results, and the recognition results sent from the word recognition unit (13) are ignored. Can be controlled.

The above-mentioned next candidate key is a key having the function described below, and will be explained using FIG. 15.

In the speech recognition unit (1) of this device, word recognition and phrase recognition run in parallel, and as mentioned above, both word and phrase recognition results are obtained. The phrase next candidate key (73) is the key for displaying the recognition results on the display device (6) in order from the highest likelihood to the recognition result, and displays the recognition results in the order from highest to lowest likelihood. The key to display the recognition result on the display device is the word next candidate key (72), which displays the recognition result with one higher likelihood on the display device (6) than the recognition results currently displayed on the display device. The keys to be displayed are the pre-word candidate key and the pre-phrase candidate key.

FIG. 15 shows the candidate buffer (15 m
In this figure, the recognition result for the - position is 'tango', which is the recognition result sent from the -a word recognition unit (13), which is represented by a (word). . Similarly, the recognition result for the third place is "l, r tango", and this is the recognition result sent from the phrase recognition unit (14).
The recognition result of the third place is "Tangoni", and this is the recognition result sent from the phrase recognition unit (14). but,
``Tanko'' and this is the recognition result sent from the word recognition unit (13) (represented by set 1).

Assume that "tango" is currently displayed on the display device (6). In this state, when the phrase next candidate key (73) is pressed, r tango is displayed on the display device (6). Further, when the next word candidate key (72) is pressed, "Tanko" is displayed on the display device (6).

In addition, when ``tango'' is displayed on the display device (6), if you press the pre-word candidate key (74), ``tango'' will be displayed on the display device 6), and the phrase next candidate will be displayed. key (7
3) When the button is pressed, "Tangoni" is displayed completely on the display device (6).

Next, we will discuss how to modify an entire passage at once.

The example in FIG. 14(e) is an example in which the word r 7 is incorrectly recognized as r A . First, move the cursor to the word you want to correct [FIG. 4(e) il.

Next, press the word next candidate key (72) to display the next word candidate [Same 1! J(e)il, if this result is correct, proceed to the next corrected part; if this result is incorrect, press the word next candidate key (72) to display the next word candidate.

Repeat this operation until the correct answer is displayed. If the correct answer is not displayed, re-speak and re-enter. Press the previous word candidate key (74) to display the candidate for the previous word displayed. You can also do it.

Next, we will discuss how to correct an entire word at once.

The example in FIG. 14(f) is an example in which the phrase ``j'' of ``gamen'' is incorrectly recognized as ``gainen no''. First, move the cursor to the phrase you want to modify [Figure (r) il.

Next, press the phrase next candidate key (73) to display the next phrase candidate [same IC(r)il, if this result is correct, proceed to the next correction part, if this result is wrong, proceed to the next phrase candidate. Press the key (73) to display the next candidate for the phrase. Repeat this operation until the correct answer is displayed. If the correct answer is not displayed, re-speak and re-enter. Press the previous clause candidate key (75) to select the previously displayed clause. can also be displayed.

Next, a method for correcting syllables will be described.

The example in FIG. 14(h) is an example in which the phrase ``Onseide'' is incorrectly recognized as ``Onseide''. In this example, the syllable ``ke'' is to be corrected to ``1''. First, move the cursor (X) to the syllable ``ke'' that you want to correct, then press the il, syllable candidate key (76) in (h) By pressing the syllable next candidate key (76), the syllable closest to the syllable of the part you want to correct will be displayed.
Move to the next modified part. If this result is incorrect, press the syllable next candidate key again to display the next syllable candidate.

Repeat this operation until the correct answer is displayed. If the correct answer is not displayed, re-enter by speaking again. If the result of re-input is wrong, correct it again using the above procedure. Repeat this operation until the correct answer is displayed.

Also, by pressing the previous word candidate key 77), the previous candidate for the syllable can be displayed.

When you want to delete a syllable [Fig. 14 (g) +], move the cursor (X) to the syllable you want to modify and press the delete key (78) to delete it [Fig. 14 (g) il].

[If you want to insert i5, move the cursor to the syllable you want to modify and press the insert key (79) to insert it.

Next, using FIG. 16, the several syllable correction method will be described.

In this example, the input sentence "kaijo" shown in FIG. In this case, first move the cursor (X) to the syllable you want to correct [
(C) of the same figure], recursively enters as “kai”. The re-voiced input speech is recognized by the speech recognition section (1), and the recognition result is displayed on the display device (6). If the 0 sense result is correct,
Proceed to the next modification section, if as shown in Figure (d),
If you misrecognize "kai" as "1 change", if it is a word,
Press the word next candidate key (72). If there are 0 phrases, press the phrase next candidate key (73). Since Figure 16 is an example of a word, the method for correcting *g will be described below. d
), when the next word candidate key (72) is pressed, the control unit (5) first selects the recognition result of (b) in the same figure before correction from the word dictionary (13d). The control unit (5) compares the recognition result "r ga egaejiyo" in the same figure (d) after re-voicing and finds the same part "jiyo".Next, the control unit (5) searches the word dictionary (13d) for Choose words that have the same part "jiyo". Figure (f) is single 11! The memory contents of the I dictionary (13d) are shown, and (g) of the same figure shows the zero-order control unit (
5) calculates the luminosity between the word written in (g) of the same figure and the recognition result r ga yō after re-uttering, and displays the word with the largest likelihood value [(e) of the same figure. ].

Next, we will discuss the case of correcting recognition boundary errors of phrases or words.

In the example in Figure 14 (g), it is mistakenly recognized that there is a silent interval indicated by a mako symbol between the phrase ``bun yo wo'' and ``n'' and ``1 shi'', This is an example of a misrecognition that is divided into two parts. In this case, the recognition boundary error must be corrected, but if you want to delete the recognition boundary delimiter, move the cursor (X) to the recognition boundary delimiter you want to delete [1(g)i] and press the delete key ( 7
8) Press [(6) i in the same figure. If you want to insert a recognition boundary delimiter, move the cursor (
X) and press the insert key (79).

However, as will be described later, the delimiter beep of the recording/playback device (2) and the delimiter added to the recognition result stored in the storage device (8) are This will serve as a landmark for synchronization, so you must take precautions.
) indicates that a delimiter has been inserted or deleted in the storage device (8).
Remember it.

By the following correction procedure, as shown in Fig. 14 (1),
Correct the text.

After the recognition boundary error has been corrected, the recognition unit for which the recognition boundary error has been corrected is corrected according to the correction procedure. In the case of correction by re-voicing, the voice of anyone who has registered the standard pattern can be recognized, so corrections can be made even if one is not the person who recorded the text.

The method for correcting kana string sentences has been described above, and the following functions are provided to assist in correction.

By moving the cursor on the character string displayed on the display device (6) and scrolling the display screen, it is possible to sequentially display memorized sentences from the storage device (8) on the display screen, but at this time, the memorized sentences can be displayed on the display screen in sequence. The audio corresponding to the part is recorded by the recording and playback device f
It is played from (2).

It also has the opposite function to the above-mentioned function, and has a recording/playback device (
The character string corresponding to the portion being played back from 2) can be displayed on the display device (6).

In addition, in both of the above methods, the time before the delimiter recorded in the recorded text and the delimiter recorded on the display side are used as timing signals for synchronization, and the playback of the recording and playback device (2) is performed. and the display are controlled so that they operate in synchronization with each other. Also, the keyboard (7
), or when a signal to stop the playback is input from the recording/playback device (2), the playback is stopped and the scrolling of the display or the movement of the cursor is also stopped.

By synchronizing the playback and display of the O sound playback device (2) as described above, character strings can be checked while listening to the playback sound, making it easy to find corrections.

As for the method of synchronization described here, the character string in the storage device (8) corresponding to the part being played back is transferred to the display device (6).
), and the display device (
6) is a display method.

In this case, when the display is stopped for correction, the corrected part of the recorded voice has already been played back, so in order to play the corrected part again, it is necessary to locate the beginning of the part to be corrected from the reproduced text. Therefore, when this method is adopted, the recording/reproducing device (2) is provided with a function of automatically backtracking to the previous delimiter when the display is stopped.

In addition, when a tape recorder is used as the recording/playback device (2), the playback section is controlled by the rotation of a motor, and due to tape slack, etc., it may not be possible to accurately locate the beginning of the section corresponding to the correction section. be.

In such a case, record the input audio for a certain length of time as PCMfi sound or ADPCM recording, and if you want to listen back to the input audio, use PCM recording or ADPCM recording.
Adds a function to listen back to PCM recorded audio.

FIG. 17 shows an embodiment of the above-mentioned functions, and numbers 01 to 05 in FIG. 0, which is a diagram of a PCM data memory for storing PCM recording data, indicate addresses. The input voice is “I am it, -1” shown in Fig. 14.
It is a sentence that says, ``I made 1 sum of 1 te-1 kamen by 1 yūsei to make 1 whole.''

When the above audio is input, the PCM data memory (
DM), the voice "Washiwa" up to the first silent section is stored at address 01. The voice "ten" up to the second silent section is stored at address 02, and the voice "te-" up to the fifth silent section is stored at address 05. At this time, P
The CM address pointer (AP) stores the address of the first data stored in the PCM data memory. In this example, 01 is stored.

At this stage, the PCM data memory is full.

Next, when audio is input, the PCM data memory (
0 to store the input voice at the address of the data stored first among the data stored in DM).
In this example, "Kamen no" is stored at address 01 where "Washiwa" was stored. At this time, the PCM address pointer (AP) stores the address of the first data stored in the PCM data memory (DM). In this example, 02 is stored. Ru.

In this state, when reproducing the contents of the PCM data memory (DM), the contents are reproduced from the address pointed to by the PCM address pointer (AP). In this example, 02, 03,
It will be played back in the order of 04 and 05.01.

With this method, it is possible to listen back to the audio accurately and quickly as many times as desired.

Also, move the cursor (X) to the recognition unit delimiter on the screen.
) and press the voice start key (70) to search the recorded text for the blue part of the delimiter on the recording/playback device (2) that corresponds to the recognition unit indicated by the cursor, and then (It has a function to play sentences.

In addition, in order to confirm the recognition result and the corrected text, the data stored in the storage device (8) can be displayed on the display device (6).
It is very tiring for the eyes because the user has to visually follow and read the string displayed on the display screen.

In view of this, this device has a function to read out the character strings stored in the storage device (8) in which the recognition results are stored using a voice synthesis function. You can do it by listening to the sound.

In this case as well, the delimiter is used as a timing signal for synchronizing the speech synthesis section (9), the storage device (8), the recording and reproducing device (2), and the display device (6).

In other words, a character string corresponding to the part read out from the storage device (8) by the speech synthesis unit (9) is displayed on the display device (6), and at the same time, the recording part is cued up by the recording/playback device (2). . With this method, when an error is discovered by the voice synthesis function and the voice synthesis function is stopped to correct it, the display on the display device (6) and the recorded portion on the recording and playback device (2) will also show the error. , and corrections can be made immediately.

The method of synchronization described here is to display on the display device (6) the kana column of the storage device corresponding to the part being read out by the speech synthesis function, and at the same time display the text being recorded on the recording/playback device (2). A method of playing back the syllable part that corresponds to 1' more, and playing back the part of the sentence that is not recorded on the recording and playback device (2) that is one delimiter later than the part that corresponds to the part that is raised by the speech synthesis function. There is a way to do this. In the latter case, when voice synthesis is stopped for correction, the recording and reproducing device (2) has stopped before the part to be corrected, so if it is played back in this state, the corrected part of the audio can be immediately played back. In the former case, when voice synthesis is stopped for correction, the corrected part of the recorded voice has already been played back, so it is necessary to backtrack in order to play the corrected part again. Therefore, when adopting the former method, when the display is stopped, the recording and playback device R (2
) preferably has the ability to backtrack to the previous delimiter.

(G) Effects of the Invention As described above, the speech recognition system of the present invention plays back all the sentences that have been recorded once, reads the fluctuations in the overall recording level, and then uses the data based on this level fluctuation data. A
Since GC is performed, it is possible to perform gain control that takes global level fluctuations into consideration, rather than being influenced by local level fluctuations. Therefore, the optimum sound pressure of input speech to the speech recognition device can be obtained, contributing to improvement of the speech recognition rate.

,4. D[i's ll#L Description Figure 1 is an external view of a dictionary-cheating machine that is an embodiment of the present invention, Figure 2 is a configuration diagram of the dictionary-cheating machine, and Figure 3 is a speech recognition unit (1). 4 is a block diagram of the preprocessing section (11), FIG. 5 is a block diagram of the feature extraction section (12), FIG. 6 is a block diagram of the word recognition section (13), and FIG. 7 is a block diagram of the word recognition section (13). Fig. 8 is a block diagram of the phrase recognition unit (14), Fig. 8 is a block diagram of the input switching unit (4), Fig. 9 is a diagram of the relationship between headwords, recording methods, and characters and sounds, and Fig. 10 is a recording of characters and sounds. 11 is a diagram showing the recording method when the recording/playback device is a multi-track system; FIG. 12 is a diagram showing the recording method when the recording/playback device is a single-track system; FIG. 13 is a diagram showing the relationship between the method and the audio section. The figure shows an example of a frequency correction circuit, FIG. 14 is a correction diagram for incorrect recognition, and FIG.
Candidate buffer <15a) in 5), FIG. 16 is a diagram showing an example of correcting several syllables at the time of erroneous recognition, FIG. 17 is an explanatory diagram of the PCM recording method, and FIG. 18 is an explanatory diagram of the AGC operation.

(1)...Speech recognition device, (2)...Recording/playback device, (3)...Microphone, (6)...Display device, (7)
... Keyboard, (8) ... Storage device, (11).
・・Preprocessing unit, (12) ・・Feature extraction unit, (13)・
...Word recognition unit, (14)...Syllable recognition unit, (lla
)...Variable gain amplifier, (llb)...Sound pressure variation memory.

Claims (1)

[Claims] (1) Comprising a recording and reproducing device for recording and reproducing audio, a storage device for storing the sound pressure of the audio reproduced by the recording and reproducing device, a speech recognition device, and an automatic gain control device for audio sound pressure, A voice recognition system that adjusts the sound pressure of a voice replayed from a recording/playback device to an optimal level for input to a voice recognition device based on the contents of a storage device that stores the sound pressure of the voice. .