JPH04181299A - Speech recognition and correction device - Google Patents

Abstract

PURPOSE:To correct a misrecognized part speedily by specifying and correcting the misrecognized part with a voice indication without indicating the misrecognized part by cursor movement. CONSTITUTION:An input speech in sampled as a 8kHz, 12-bit digital signal by the speech input part 21 consisting of a microphone, a high-frequency-band emphasizing filter, and an AD converter as constituent elements of an acoustic analytic part 11. A feature extraction part 22 as a constituent element of the acoustic analytic part 11 performs the frame-by-frame frequency conversion of the speech signal converted into the digital signal in its 16ms section to extract feature parameters in the frequency range, thereby obtaining a sequence of the feature parameters of a voiced document. The feature parameter sequence of the voiced document which is extracted by the feature parameter extraction part 22 is stored in a feature parameter sequence storage circuit 27. The misrecognized part of the speech is reinputted to specify and correct the misrecognized part speedily.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a continuous speech recognition device that recognizes speech that is continuously uttered without dividing each word.

"Conventional technology" A conventional speech recognition device will be explained using FIG. 6.

In a conventional speech recognition device, an acoustic analysis section 1 analyzes the components of the input speech, and based on the output from the acoustic analysis section 1, a text speech recognition section 2 recognizes the input speech. The results are displayed on the display section 4. The displayed results are checked by the operator, and if there are no errors in the recognition results, the operator immediately moves on to the next voice input operation. However, if the input voice is incorrectly recognized, use the cursor indicator 3 to move the cursor to the beginning of the part where the recognition result was incorrect and to the end of the incorrectly recognized part and mark it. Then, the misrecognized parts are specified. In conventional speech recognition devices, the operator operates the cursor pointing section 3 to indicate the portion where an error has occurred in the speech recognition result. In order to correct the correction point specified by the operation of the cursor instruction section 3, the operator again operates the correction point manually, such as by voice. In this correction operation, in a speech recognition device that can display several recognition results, if there is a correct recognition result among the candidates, the correct recognition result is selected.

``Problem to be solved by the invention'' With voice input, data can be entered without performing key input operations, and there is no need to know the key layout positions or key operation methods on a human-powered device, so anyone can use voice input. This is an easy-to-use input method.

However, unlike manual methods using key operations, voice input methods tend to have a slightly lower probability of accurately recognizing input data. Therefore, regarding the text in the journal of the Acoustical Society,
When we investigated the number of phonemes that make up one sentence, we found that the average number of phonemes per sentence was approximately 47 phonemes. Here, even if a speech recognition device with a phoneme recognition rate of 9596 or higher is used,
Misrecognition of 2 to 3 phonemes in every 1,000 sentences is unavoidable.

Furthermore, even when considering a speech recognition device that has a phoneme recognition rate of 98% or more, it is quite possible that one phoneme in a - sentence may be misrecognized. Considering the above facts,
It can be seen that it is not uncommon for it to be necessary to correct the recognition results of input speech. When considering speech recognition devices, the role of speech recognition error correction devices is considered to be extremely important.

However, in the speech recognition error correction device using the above-mentioned conventional technology, when an error occurs in the speech recognition result, the operator manually operates the cursor indicator 3 in FIG. Must be specified. It is very burdensome for the operator to have to perform such correction operations frequently. In addition, while using an excellent input method called voice input, the conventional technology specifies the correction part by manually moving the cursor.
It is not possible to fully utilize the advantages of voice recognition devices, such as ``easy operation by anyone'' and ``fast data input speed compared to manual input or other methods.'' .

In other words, even if data can be manipulated quickly,
It is difficult for anyone to easily and quickly perform a correction operation using a speech recognition correction device using the conventional technology. As described above, in the speech recognition correction apparatus using the conventional technology, the correction operation takes a very long time, and the work efficiency is extremely low. The conventional technology has the problems described above.

SUMMARY OF THE INVENTION An object of the present invention is to eliminate such drawbacks and to provide a speech recognition and correction device that can specify and correct a misrecognized portion by voice instructions without instructing the misrecognized portion by moving a cursor.

ri! The speech recognition and correction device of the present invention includes an acoustic analysis section that analyzes the components of input speech, and an acoustic analysis section that analyzes the first input speech outputted from the acoustic analysis section based on the first candidate. A speech recognition section for recognizing, a means for displaying the output of the speech recognition section, and a means for displaying the output of the speech recognition section, and storing the first human-powered speech output from the acoustic analysis section, and then re-human-generating only the parts of the misrecognized speech. a sentence/speech spotting unit that determines which part of the first input speech is the erroneously recognized part, and outputs the next candidate to the speech recognition unit for the misrecognized part, and displays it on the display unit. and means for determining recognition of the input voice when the next candidate selected matches the input voice.

"Examples" The present invention will be described in detail below based on Examples.

FIG. 1 is a block diagram of the principle of the speech recognition correction device of the present invention.
FIG. 2 is a block diagram of one embodiment of the present invention. The human-generated voice is sampled as an 8 KHz, 12-bit digital signal by a 21 voice input unit composed of 11 components of the acoustic analysis unit, such as a microphone, a high-frequency emphasis filter, and an AD converter. Furthermore, the acoustic analysis section 1
1, the feature extraction unit 22 converts the frequency of the audio signal converted into a digital signal frame by frame, with a 16m5 section as one frame, extracts feature parameters in the frequency domain, and calculates the frequency of the uttered sentence. It is expressed as a feature parameter sequence. The feature parameter string of the spoken sentence extracted by the feature extractor 22 is stored in the feature parameter string storage circuit 27.

The feature parameter string extracted by the acoustic analysis section 11 is passed through a continuous DP matching circuit 23 that constitutes the text speech recognition section 12.
In the step, the word dictionary 24 and the feature parameter string of the uttered sentence are matched. The word lattice recognized and determined in the continuous DP matching circuit 23 is stored in the word lattice storage circuit 32, and is stored in the display unit control circuit 25.
is displayed on the display section 26 under the control of. If there is an error in the recognition result of the input sentence displayed on the display unit 26, the operator notifies the user of the need to correct the recognition result by touching or pressing the correction key.

The correction key is controlled by a correction key control circuit 28, and the correction key control circuit 28 instructs the user to immediately enter a system for requesting input of the incorrectly recognized word.

The re-entered words will appear in the same way as the originally entered sentences.
The voice recognition section 21 and the feature extraction section 22 convert it into a feature parameter string. The feature parameter string of the misrecognized word is word-spotted with the feature parameter string of the first input voice stored in the feature parameter string storage circuit 27 in the continuous DP matching circuit 29. By re-inputting the erroneously recognized portions that have been erroneously recognized, the corrected portions of the recognition results are reliably spotted. If the correct word exists in the word lattice storage circuit 32, the word lattice replacement circuit 30
The words selected as the second and third candidates are corrected, and the recognition result is confirmed using the confirmation key controlled by the confirmation key control circuit 31. If the correct word does not exist among the second and third candidates, the correct word is input again and word recognition is performed through the same route as when inputting the sentence for the first time.

The present invention will be further explained based on an example using FIGS. 3 and 4.

FIG. 3 is a diagram showing the word lattice structure in the word lattice storage circuit of the present invention.

Now, assume that the operator inputs the sentence "Today's weather is not sunny." The word lattice structure in the word lattice storage circuit 32 when this input is received is shown in FIG. 3, as shown in FIG. be. In addition, the word lattice structure of "hare" has the first candidate "needle" and the second candidate "needle".
The third candidate is ``Hashi''. Similarly, the word lattice structure for "desu" is the first candidate "desu" and the second candidate "fat". In this case, "sunny" is mistakenly recognized as "needle". The operator uses the correction key to notify that the recognition result needs to be corrected. The correction key control circuit 28 recognizes that there is an error in the recognition result,
A system will be established to immediately request the input of the misrecognition results themselves.

FIG. 4 is a diagram illustrating the order of correcting word recognition according to the present invention.

The operator re-enters the word "needle" that was incorrectly recognized. This is to spot erroneously recognized parts. The human-powered "needle" is 21 voice input section, 22
It is converted into a feature parameter string through the feature extraction unit, and the 29D
The P matching circuit performs Dr matching with the feature parameter string stored in the 27 feature parameter storage circuit, and spots misrecognized words. FIG. 4 shows, as a feature parameter sequence, the voice power of the first input and the voice power when the misrecognized word "needle" is input again. As shown here, it is not difficult to spot the word ``needle'' that has been recognized twice. In this way, the misrecognized part "needle" is detected as a part that requires correction, and fortunately the correct word "hare" exists in the second candidate, so the 30-word lattice exchange circuit is controlled using the confirmation key to Select the second candidate "sunny" and confirm the recognition result. Through the above operations, the correction of the incorrect recognition result is completed and the correct sentence recognition result ``Today's weather is sunny.'' is obtained.

The process of the present invention will be explained in detail with reference to FIGS. 5, 2, 3, and 4. FIG. 5 is a flowchart showing an example of processing according to an embodiment of the present invention.

First, the operator inputs data by voice. The audio input unit shown in FIG. 2 is involved in inputting audio data. Features of the input speech are immediately extracted.

The feature extractor shown in FIG. 2 is involved in this feature extraction. The extracted features are stored in the feature parameter string storage circuit shown in FIG. 27. A feature parameter string of some type obtained by feature extraction is subjected to continuous DP matching with a standard feature parameter string of words described in a word storage dictionary. The continuous DP matching circuit 23 in FIG. 2 is involved in this continuous DP matching. Further, the characteristic parameter string used as the standard button at this time is that of the word memory dictionary 24 in FIG. Continuous D
Recognition is determined based on the P matching results, and the results are displayed. The display control section shown in FIG. 25 and the display section 26 shown in FIG. 2 are involved in this display. Examples of displayed results are as shown in FIGS. 3 and 4. If there is an error in the displayed recognition result,
Use the correction key to notify that recognition correction is necessary. The correction key used here (=, is controlled by the correction key control circuit shown in Fig. 2).If correction is necessary, it is necessary to immediately detect the corrected part. In order to spot the corrected part, the operator is asked to input the erroneous recognition result of the erroneously recognized part by voice.The human power of the erroneously recognized part is immediately extracted as a feature and becomes a feature parameter string. In the matching circuit, continuous DP matching is performed with the feature parameters of the first voice input stored in the feature parameter string storage circuit shown in FIG. 27.

Misrecognized portions are spotted based on the results of this continuous Dr matching. The situation is as shown in Fig. 4. The misrecognized parts were spotted and the corrected parts became clear. Here, if the user recognizes a correct recognition result in the word lattice, he or she operates the priority order of the word lattice using the confirmation key to confirm the recognition result. The word lattice exchange circuit shown in FIG. 2, 30, and the confirmation key control circuit, shown in FIG. However, if there is no correct recognition result in the word lattice, a new voice input will be required.

In the embodiment, the audio input section shown in FIG. 2 is composed of a microphone, a high-frequency emphasis filter, and an AD converter,
, 12-bit digital signals are used, but other configurations may be used as long as the input audio can be sampled quickly.

In addition, the feature extraction unit in FIG.
Although we used a method of converting the frequency for each frame, extracting feature parameters in the frequency domain, and expressing them as a feature parameter sequence of the uttered sentence, other methods can also be used to accurately extract features. It doesn't matter if that's the case.

"Effects of the Invention" As described above, the speech recognition correction device of the present invention corrects the erroneously recognized part of the input speech recognition result.
There is no need to specify the erroneously recognized portion by moving the cursor, and by re-operating the erroneously recognized portion by voice, the erroneously recognized portion can be specified and corrected very quickly. Therefore, even in cases where voice recognition errors may occur frequently due to the deterioration of the usage environment of the voice recognition device or the physical condition of the operator who manually operates the voice recognition device, special measures are taken to correct the voice recognition. It does not require advanced operations or knowledge, reducing the burden on the operator and significantly improving work efficiency.

[Brief explanation of drawings]

FIG. 1 is a block diagram of the principle of the speech recognition and correction device of the present invention. FIG. 2 is a block diagram of an embodiment of the present invention. FIG. 3 is a diagram showing a word lattice structure in the word lattice storage circuit of the present invention. FIG. 4 is a diagram illustrating the order of correcting word recognition according to the present invention. FIG. 5 is a flowchart showing an example of processing according to an embodiment of the present invention. FIG. 6 is a block diagram of a conventional speech recognition correction device. Applicant Seiko Epson Co., Ltd. agent Patent attorney Kizobe Suzuki and 1 other person ↓ 5 Figure 1 Figure 2 - Candidate Today's weather is needles
The second candidate is the angel.
Sunny fat third candidate
Bridge Figure 4 Figure 5

Claims (1)

[Claims] an acoustic analysis section that analyzes components of input speech; a speech recognition section that recognizes the first input speech outputted from the acoustic analysis section based on a first candidate; and means for displaying the output of the speech recognition section. , the first input voice outputted from the acoustic analysis unit is memorized, and when only the misrecognized portion of the voice is input again, which part of the first input voice is the next input? and outputs the next candidate to the speech recognition unit for the erroneously recognized part; and if the next candidate displayed on the display unit matches the input speech, A speech recognition correction device comprising means for determining recognition of input speech.