JP3259734B2 - Voice recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition device.

[0002]

2. Description of the Related Art A speech recognition apparatus is an apparatus for recognizing speech input data as character data. Since this voice recognition device can input data without performing key operations, anyone can easily operate the voice recognition device without knowing the key layout and key operation method of the key input device. The configuration of such a conventional speech recognition apparatus is shown in the block diagram of FIG. As shown in FIG.
The input voice is analyzed by the voice input unit 110, converted into a digital signal and output. The voice recognition unit 120 recognizes the input voice by performing matching with a basic pattern registered in a voice recognition dictionary based on output data from the voice input unit 110. This recognition result is displayed on the recognition result display unit 130 so that the operator can confirm it. The recognition result is also sent to the storage unit 140 and stored as feature extraction result data obtained by analyzing the input speech.

[0003]

However, in the conventional voice recognition apparatus, even if the operator pronounces and inputs a sound correctly, the voice recognition apparatus may be erroneously recognized. Possible causes of such misrecognition include background noise in the environment in which the conventional device is used, and the fact that input speech data in units of sentences is not processed using grammatical knowledge or semantic knowledge of sentences. Is mentioned. In particular, since particles are rarely pronounced clearly, the recognition rate of the entire input voice data may vary greatly depending on whether particles can be accurately recognized.

[0004] It is an object of the present invention to provide a speech recognition apparatus capable of resolving such causes and correctly recognizing input speech data.

[0005]

In order to solve the above-mentioned problems, a speech recognition apparatus according to the present invention comprises an utterance unit extraction unit for extracting a proper utterance unit of input speech data, and an utterance unit extraction unit. An input instructing unit for instructing an input in the uttered unit and a plurality of dictionaries for each uttered unit, and the input speech data is recognized using the utterance unit dictionary extracted by the uttered unit extracting unit. A voice recognition unit.

[0006]

According to the speech recognition apparatus of the present invention, an appropriate speech unit of the input speech data is extracted by the speech unit extraction unit. The input instruction unit having received the information of the utterance unit gives an instruction to input a voice in the utterance unit. Further, the speech recognition unit selects a dictionary of speech units extracted by the speech unit extraction unit from a plurality of dictionaries provided for each speech unit. Therefore, if the voice data is input in units of utterances in accordance with the instruction in the input instruction unit, the voice recognition unit uses the dictionary of the same utterance unit to perform voice recognition.

[0007]

An embodiment of the present invention will be described below with reference to the accompanying drawings. FIG. 1 is a block diagram illustrating an outline of the speech recognition apparatus according to the present embodiment. As shown in FIG.
The voice data input to 0 is provided to the voice recognition unit 20, and voice recognition processing is performed. This recognition result is displayed on the recognition result display unit 30 and is also given to the utterance unit extraction unit 40. The utterance unit extraction unit 40 receives background noise data and correction input data in addition to the recognition result data, and extracts an appropriate utterance unit of the voice data input to the voice input unit 10. The result of the extraction is displayed on the utterance unit display unit 50, which is an input instruction unit, and an instruction is given to input audio data in appropriate utterance units.
The extraction result is also provided to the speech recognition unit 20, and an appropriate speech recognition dictionary is selected from the extraction result, and input speech data is recognized using the selected speech recognition dictionary.

Before giving a detailed description of this embodiment, the features of the processing of this embodiment will be described. The grammatical units of speech that can be uttered by humans include “phonemes”, “words”, “phrases”, and “sentences”. In the present embodiment, attention is paid to such utterance units of the voice input to the voice recognition device, and the utterance unit is divided into any utterance units from information such as the difficulty level of the conversation content and the level of the background noise. Is determined to be optimal, and a voice input in an optimal utterance unit is instructed. In other words, if the difficulty level of the conversation content is low or the level of the background noise is low, there is a low possibility of misrecognition in the recognition process. When the difficulty level of the conversation content is high or the level of the background noise is high, there is a high possibility that the recognition process will cause erroneous recognition. Therefore, a voice input divided for each "phoneme" or "word" is instructed. The instruction of the optimal utterance unit in the voice input is the first feature of the present embodiment, and is performed by the utterance unit extraction unit 40. Such utterance unit extraction unit 4
In response to the instruction at 0, the speech data input by the operator in units of utterances as instructed is speech-recognized using a speech recognition dictionary most suitable for the instructed utterance unit. The selection of an optimal dictionary in the voice recognition processing is the second feature of the present embodiment, and is performed by the voice recognition unit 20.

The flow of processing in this embodiment is very similar to the case where a person calls in an environment with a high noise level. For example, if you say "I will arrive at Narita at 3 o'clock tomorrow" and the other party cannot hear you, "tomorrow", "3 o'clock",
Say "to Narita" and "to arrive" in units of phrases. If you still could n’t hear it, simply say “tomorrow”, “3
"Time", "Narita", and if the other party mistakenly hears "Narita" and "Haneda", "Na", "Li",
Say “ta” with phonemes. In this way, humans proceed with the conversation by determining the most appropriate utterance unit at that time. In this embodiment, an appropriate utterance unit is not a human,
By making a determination on the device side and transmitting the information to the operator, smoother voice recognition processing is performed.

Next, the detailed configuration of the speech recognition apparatus of the present embodiment will be described with reference to the block diagram of FIG. From the figure,
The audio input unit 10 includes a microphone 11 for inputting audio, a high-frequency emphasis filter 12 for sampling input audio data as a digital signal, and an AD converter 13. The speech recognition unit 20 includes a feature extraction circuit 21 that performs frequency conversion on a speech signal converted into a digital signal and extracts a feature parameter sequence in a frequency domain, and a feature parameter sequence that stores the extracted feature parameter sequence. A storage circuit 22 for performing a speech recognition process on the characteristic parameter sequence;
A P matching circuit 23 and a recognition dictionary group 24 having a plurality of speech recognition dictionaries for each utterance unit are provided.
Further, the recognition result display unit 30 includes a display device 31 for displaying a speech recognition result, and a display control circuit 32 for controlling the display device 31. Also,
The utterance unit extraction unit 40 analyzes the speech recognition result stored in the speech recognition storage circuit 41 and the speech recognition storage circuit 41 that stores the recognition result of the speech recognition unit 20, and determines the difficulty of the utterance content. A difficulty level determination unit 42, a correction input unit 43 for inputting correction data when it is determined from the speech recognition result displayed on the display device 31 that there is an error in the recognition result, and speech recognition based on the correction data. , A microphone 45 for measuring the background sound, a high-frequency emphasis filter 46 as a noise detecting unit for analyzing background noise data from the measured background sound, and a difficulty level determining unit The difficulty utterance data determined in step S42, the recognition rate data detected by the recognition rate detection unit, and the background noise data analyzed by the high-frequency emphasis filter 46 are input to determine the appropriate utterance unit to determine the appropriate utterance unit A determination unit 47, a speech unit memory circuit 48 for storing the determination result is provided. Further, the utterance unit display unit 50 includes a display device 51 for displaying the recognition rate detected by the recognition rate detection unit 44 and the proper utterance unit stored in the utterance unit storage circuit 48, and a display control for controlling the display device 51. A circuit 52 is provided.

Next, the flow of the processing of this embodiment will be described. The audio data input from the microphone 11 by the operator is converted into a digital signal through the high-frequency emphasis filter 12 and the AD converter 13. This audio signal is frequency-converted by the characteristic extraction circuit 21 to extract a characteristic parameter sequence in the frequency domain. The extracted characteristic parameter sequence is stored in the characteristic parameter sequence storage circuit 22. DP
In the matching circuit, the speech signal stored in the feature parameter string storage circuit 22 is read, and DP matching processing is performed using one of a plurality of speech recognition dictionaries registered in the recognition dictionary group 24. Done. In this process, the speech signal is recognized as a code string, and transferred to the recognition result display unit 30 and the speech recognition storage circuit 41. The code string transferred to the recognition result display unit 30 is displayed as a recognition result on the display device 31 so that the operator can confirm it. If there is an error in the recognition result, the operator can input correction data to the correction input unit 43.
The corrected data is sent to the recognition rate detection unit 44, where the recognition rate is detected. The recognition rate data detected by the recognition rate detecting unit 44 is sent to the appropriate utterance unit determining unit 47. Also, from the DP matching circuit 23 to the speech recognition storage circuit 41
Are transferred to the difficulty level determination unit 42. The difficulty determination unit 42 calculates the difficulty information calculated from the frequency of words, the politeness information of a sentence that is one of the features of the conversation format, the length information of one sentence, and the information related to the sentence. As an index, the degree of difficulty of the input code string is comprehensively determined. This determination result is sent to the appropriate utterance unit determination unit 47 as difficulty level data. Further, the background sound measured by the microphone 45 is analyzed by the high-frequency emphasis filter 46, and information on the influence of the background noise level and type, such as the level of the background noise, on the speech recognition rate is determined as background noise data. 4
7

The appropriate utterance unit determination unit 47 determines an appropriate utterance unit based on the recognition rate data, difficulty level data, and background noise data thus input, and stores the determination result in the utterance unit storage circuit 48. Is done. In addition, the result of this determination is displayed on the display device 51 of the utterance unit display unit 50, and the operator is given an instruction on an appropriate input unit when the next voice input is performed. That is, the appropriate utterance unit determination unit 4
If it is determined in step 7 that the state is most suitable for speech recognition, information indicating an instruction to input speech in units of sentences or a plurality of sentences is displayed on the display device 51. If the recognition rate cannot be sufficiently obtained unless the input is performed in units of phrases and words, the information instructing the voice input in units of phrases and words is displayed on the display device 51. Further, in a poor environment, information instructing voice input in units of phonemes is displayed on the display device 51.

The utterance unit information stored in the utterance unit storage circuit 48 is given to the DP matching circuit 23. Then, D with the voice data input in the specified utterance unit
The most suitable speech recognition dictionary for P matching is the recognition dictionary group 2
4 is selected from a plurality of speech recognition dictionaries registered in. That is, when the specified utterance unit is “phoneme”, the dictionary of “phoneme information” registered in the recognition dictionary group 24 is selected. When the specified utterance unit is “word”, a dictionary of “word information” registered in the recognition dictionary group 24 is selected. Further, when the designated utterance unit is “phrase”, a dictionary related to “phrase” of “grammar information” and “semantic information” registered in the recognition dictionary group 24 is selected. If the specified utterance unit is “sentence”,
A dictionary relating to “sentence” of “grammar information” and “semantic information” registered in the recognition dictionary group 24 is selected. Speech recognition processing is performed using the dictionary selected in this way.

Regarding the process of selecting the speech recognition dictionary,
This will be described using a specific example. First, the operator who has received the instruction information that the appropriate utterance unit is “word” performs the next voice input in units of words. For example, when inputting "Hanagasa Ondo", it is performed by dividing into "Hanagasa" and "Ondo". From the input speech, the feature of the speech is extracted by the feature extraction circuit 21 and stored in the feature parameter string storage circuit 22. Then, the information of the appropriate utterance unit stored in the utterance unit storage circuit 48 is transferred to the DP matching circuit 23, and a dictionary of “word information” is selected from the recognition dictionary group 24. The DP matching circuit 23 uses the dictionary of “word information” to perform DP matching with a feature parameter sequence, which is a speech signal read from the feature parameter sequence storage circuit 22. At this time, the speech recognition process can proceed without using any information such as "grammar information" or "semantic information" that is valid only when the speech input unit is "bunsetsu" or "sentence". Can be processed efficiently. And D
The code string recognized by the P matching circuit 23 is displayed on the display device 31.

On the other hand, if the speech recognition processing in the DP matching circuit 23 has failed (if the distance from all the data registered in the speech recognition dictionary used is larger than the threshold), the speech recognition processing is stopped. The display indicating the failure is performed on the display device 31 to prompt the operator to re-input the voice.
The appropriate utterance unit at the time of re-input is the difficulty level data in the difficulty level determination unit 42, the recognition rate data in the recognition rate detection unit 44,
In consideration of the background noise data input from the microphone 45, the appropriate utterance unit determination unit 47 determines the appropriate input unit by reevaluation.

When the operator who sees the recognition rate data displayed on the display device 51 determines that the expected speech recognition rate has not been obtained, the operator inputs the information to the correction input section 43. can do. Information from the operator that the speech recognition rate is poor is given to the appropriate utterance unit determination unit 47, and an utterance unit shorter than the utterance unit that is currently appropriate is determined as an appropriate utterance unit. In response to the instruction based on the determination result, the next input voice data is a short utterance unit input, so that the recognition rate is improved.

Next, the processing of the difficulty level determination section 42 will be described.
This will be described in detail. The difficulty level determination unit 42 analyzes the input code string, and determines the frequentness information calculated from the frequentness of words, the politeness information of the text, the length information of the text, and the dependency information of the text. Comprehensively determine the difficulty level. The frequency information calculated from the frequency of words is information obtained by adding the frequency points of the input code string. this is,
It is based on the fact that the less frequently the word is, the higher the difficulty of the sentence is. The politeness information of a sentence is information obtained by extracting respectful words, humble words, and polite words contained in the input code string. This is based on the difficulty of sentences using polite expressions. The sentence length information is information obtained by adding one point to a sentence consisting of 20 or more words and subtracting one point from a sentence consisting of 10 or less words. This is based on an empirical rule that a sentence with a large number of words is a sentence with a high degree of difficulty. The sentence dependency information is information obtained by grammatically defining the dependency and other expressions in the sentence with 12 levels of difficulty, and adding the difficulty of the sentence. This is based on the correlation between the dependency information expressed in the text and the difficulty level. The sum of the points obtained from the information calculated as described above is calculated, and the result is sent to the appropriate utterance unit determination unit 47 as the difficulty of the input code string.

FIG. 3 shows an example in which the speech recognition apparatus of this embodiment is realized by a personal computer. In this example, the recognition result information and the appropriate utterance unit instruction information are displayed on the same screen, thereby facilitating the operator's confirmation work.

In this embodiment, the voice input unit 10
Microphone 11, high-frequency emphasis filter 12, AD converter 1
3, sampling is performed as a digital signal. However, any other configuration may be used as long as the input sound can be sampled quickly. In the feature extraction circuit 21, a method of frequency-converting a voice signal converted into a digital signal, extracting a feature parameter in a frequency domain, and expressing it as a feature parameter sequence of a spoken word is used. Any method may be used as long as the method can accurately extract features. Further, in this embodiment, a method of displaying the voice recognition result on the display device 31 is used as a means for notifying the operator of the voice recognition result, but other methods may be used to promptly notify the operator of the voice recognition result. Any method can be used. Also,
When determining the difficulty level of the utterance content, the difficulty level determination unit 42 used the information on the frequency of the words and the form of the conversation, but any other information that can accurately determine the difficulty level of the utterance content May be used.

[0020]

According to the speech recognition apparatus of the present invention, an input in an appropriate utterance unit is instructed by the input instructing section. And
Voice data input in units of utterances in accordance with this instruction can be promptly subjected to voice recognition processing using a dictionary composed of accurate information that matches the units of utterance. This makes it possible to perform speech recognition processing in a highly flexible manner even when the use environment changes, changes in voice input data, or situations where the voice input unit frequently changes. The application can be greatly expanded.

Further, since the utterance unit at the time of voice input is specified in advance, even a beginner can acquire the voice input operation method of the voice recognition device in a short time. For this reason, it has become possible to provide a speech recognition device that can be used by many people.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating an outline of a speech recognition apparatus according to an embodiment.

FIG. 2 is a block diagram illustrating a detailed configuration of the speech recognition device according to the embodiment.

FIG. 3 is an external view of a voice recognition device of the present embodiment.

FIG. 4 is a block diagram of a conventional speech recognition device.

[Explanation of symbols]

10: voice input unit, 20: voice recognition unit, 30: recognition result display unit, 40: speech unit extraction unit, 50: speech unit display unit.

Continuation of the front page (58) Field surveyed (Int.Cl. ⁷ , DB name) G10L 15/28 G10L 15/20 G10L 15/22 JICST file (JOIS)

Claims (7)

(57) [Claims] 1. An utterance unit extraction unit for extracting an appropriate utterance unit of input speech data; an input instruction unit for instructing input in the utterance unit extracted by the utterance unit extraction unit; A speech recognition unit that has a plurality of dictionaries and recognizes input speech data using the speech unit dictionary extracted by the speech unit extraction unit; A difficulty level determination unit that determines the difficulty level of the utterance content of the voice data; and an utterance unit determination unit that determines an optimal utterance unit for voice recognition from the difficulty level of the utterance content determined by the difficulty level determination unit. A speech recognition device characterized by being used. 2. An utterance unit extraction unit for extracting an appropriate utterance unit of input speech data; an input instruction unit for instructing input in the utterance unit extracted by the utterance unit extraction unit; A speech recognition unit that has a plurality of dictionaries and recognizes input speech data using the speech unit dictionary extracted by the speech unit extraction unit; A recognition rate detection unit that detects a recognition rate of a recognition result of the voice data; and a speech unit determination unit that determines an optimal speech unit for speech recognition based on the recognition rate detected by the recognition rate detection unit. A speech recognition device characterized in that: 3. An utterance unit extraction unit for extracting an appropriate utterance unit of the input speech data; an input instruction unit for instructing input in the utterance unit extracted by the utterance unit extraction unit; A speech recognition unit that has a plurality of dictionaries and recognizes input speech data using the speech unit dictionary extracted by the speech unit extraction unit; Speech recognition characterized by comprising: a noise detection unit for detecting; and a speech unit determination unit for determining an optimal speech unit for speech recognition according to the level of background noise detected by the noise detection unit. apparatus. 4. An utterance unit extraction unit for extracting an appropriate utterance unit of the input speech data; an input instruction unit for instructing input in the utterance unit extracted by the utterance unit extraction unit; A speech recognition unit that has a plurality of dictionaries and recognizes input speech data using the speech unit dictionary extracted by the speech unit extraction unit; A difficulty level determination unit that determines the difficulty level of the utterance content of the voice data; a recognition rate detection unit that detects a recognition rate of the recognition result of the input voice data; and a difficulty level of the utterance content determined by the difficulty level determination unit A speech recognition apparatus, comprising: a speech unit determination unit that determines an optimal speech unit for speech recognition based on a degree and a recognition rate detected by the recognition rate detection unit. 5. An utterance unit extraction unit for extracting an appropriate utterance unit of the input speech data; an input instruction unit for instructing input in the utterance unit extracted by the utterance unit extraction unit; A speech recognition unit that has a plurality of dictionaries and recognizes input speech data using the speech unit dictionary extracted by the speech unit extraction unit; A difficulty level determination unit that determines the difficulty level of the utterance content of the voice data, a noise detection unit that detects background noise, and a difficulty level of the utterance content determined by the difficulty level determination unit and the noise level detected by the noise detection unit. A speech recognition device comprising: a speech unit determination unit that determines a speech unit optimal for speech recognition based on a background noise level. 6. An utterance unit extraction unit for extracting an appropriate utterance unit of the input speech data; an input instruction unit for instructing input in the utterance unit extracted by the utterance unit extraction unit; A speech recognition unit that has a plurality of dictionaries and recognizes input speech data using the speech unit dictionary extracted by the speech unit extraction unit; A recognition rate detection unit that detects a recognition rate of a recognition result of voice data, a noise detection unit that detects background noise, and a recognition rate detected by the recognition rate detection unit and a background noise detected by the noise detection unit. A speech recognition apparatus, comprising: a speech unit determination unit that determines an optimal speech unit for speech recognition based on a level. 7. An utterance unit extraction unit for extracting an appropriate utterance unit of the input speech data; an input instruction unit for instructing input in the utterance unit extracted by the utterance unit extraction unit; A speech recognition unit that has a plurality of dictionaries and recognizes input speech data using the speech unit dictionary extracted by the speech unit extraction unit; A difficulty determination unit that determines the difficulty of the utterance content of the voice data; a recognition rate detection unit that detects a recognition rate of a recognition result of the input voice data; a noise detection unit that detects background noise; Determining an optimum utterance unit for speech recognition based on the difficulty level of the utterance content determined by the determination unit, the recognition rate detected by the recognition rate detection unit, and the level of background noise detected by the noise detection unit; Utterance unit determining unit A speech recognition device characterized in that: