JP2013250379A - Voice recognition device, voice recognition method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a voice recognition device, a voice recognition method and a program for improving recognition accuracy of a proper noun included in voice by correcting an error in extraction of acoustic characteristics of the proper noun to contrast with a user dictionary to the proper noun.SOLUTION: A voice recognition device 100 in an information terminal collects a proper name on the basis of a use situation of a user to the terminal to create a user dictionary. Input voice is made into a text by voice recognition on the basis of its voice waveform, and a text part concerning a proper noun is extracted from the voice. A specific character at the extracted text part is replaced, and one or more words are generated. The words are retrieved by the user dictionary, and when the words are included in the user dictionary, the text part concerning the proper noun is replaced with the words.

Description

  The present invention relates to a voice recognition device, a voice recognition method, and a program for improving recognition accuracy for input voice.

  Utilization of information input by voice recognition is increasing in the use situation of computers in environments where manual input using a keyboard or the like is difficult, for example, operation control of a navigation device in a running vehicle, character input on a portable information terminal, etc. Yes. In speech recognition technology, the input speech is analyzed for its acoustic features by comparison with data using a speech corpus called an acoustic model, and the phoneme sequence called a language model for its linguistic features. Conversion to text data through analysis based on linguistic constraints has been established as a basic technology.

  Since the voice input to the information equipment is affected by the voice quality of the speaker, the performance of the input equipment, the surrounding environment, etc., the greatest interest in voice recognition technology is how to improve the recognition accuracy. . In general, by increasing the sample data in the acoustic model and language model, that is, the speech corpus and registered words, the recognition accuracy can be improved, but there is a concern that the processing speed decreases due to the increase in data. Is done.

  Patent Document 1 exists as a technique that can solve such a problem. The technique disclosed in Patent Document 1 extracts a specific section from the speech recognized by the first speech recognition unit, and performs analysis based on a more restrictive language model for the section. And improving the recognition accuracy. And, as a method of extracting this specific section, by using the well-known “proprietary noun extraction technique”, the proper noun is extracted as the specific section, and this is compared with a dedicated proper noun dictionary to improve accuracy. Is disclosed.

  Further, Patent Literature 2 exists as a technique for improving recognition accuracy in a voice recognition device for in-vehicle navigation. The technique disclosed in Patent Document 2 is to improve the recognition accuracy by grouping voice dictionaries for each region and using different dictionaries to be used for voice recognition based on the current position of the vehicle.

JP 2011-242613 A Japanese Patent Laid-Open No. 7-64480

  Although the technique disclosed in Patent Document 1 is expected to improve the recognition accuracy for the specific expression included in the speech to a certain degree, it is considered that the degree is limited for the following reason. That is, the technique of Patent Document 1 performs re-recognition by changing the language model with respect to the expression of the extracted specific section. Therefore, when the recognition in the acoustic model originally includes an error. Even if the section is re-recognized, it cannot be expected to improve the recognition accuracy.

  This problem also applies to the technique disclosed in Patent Document 2. If an error is included in the recognition in the acoustic model, improvement in accuracy cannot be expected.

  The present invention has been made to solve these problems, and corrects errors in the extraction of acoustic features of proper nouns contained in speech and compares them with a user dictionary. The recognition accuracy can be improved.

  The present invention is a speech recognition apparatus in an information terminal, which collects a unique name based on a user's usage status for the terminal, creates a user dictionary, and inputs speech based on the speech waveform. A means for recognizing and converting to text, a means for extracting a text part related to a proper noun from the textized voice, and replacing one or more words by replacing specific characters in the extracted text part. Means for generating, and when the user dictionary includes the one or more words, means for replacing a text part related to the proper noun with the word.

  Preferably, the means for converting the speech into text includes means for determining each phoneme included in the speech waveform based on a probability based on an acoustic model, and the means for generating one or more words from the extracted text portion, A specific character in the extracted text portion is replaced with reference to probability data based on the acoustic model for the speech waveform.

  Preferably, when the means for generating one or a plurality of words from the extracted text part converts the speech into a text, the phoneme that is the next candidate for each phoneme is used as the corresponding phoneme in the extracted text part. Replace with

  Preferably, the voice recognition device is a voice recognition device in an information terminal having a navigation function for a moving body, and the means for creating the user dictionary uses a unique name related to the navigation based on the movement history of the user. Collect and create a user dictionary.

  Preferably, the means for creating the user dictionary collects a unique name including a place name and a facility name acquired from the map data based on a current position of the mobile object, a destination or a route from the current place to the destination value. It is composed.

  Preferably, the user dictionary includes, for each unique name, information on its reading, position coordinates, and registration date and time, and a priority is given based on the information, and the proper noun When there are a plurality of words related to the replacement, the means for replacing the text portion according to determines a word related to the replacement according to the priority order in the user dictionary.

  The present invention relates to a speech recognition method in an information terminal, wherein a unique name is collected based on a user's usage history for the terminal, a user dictionary is created, and an input speech is based on the speech waveform. Recognizing and converting to text, extracting a text part related to a proper noun from the textized voice, and replacing one or more words by replacing specific characters in the extracted text part And a step of replacing the text portion related to the proper noun with the word when the user dictionary includes the one or more words.

  Preferably, the step of converting the speech into text includes determining each phoneme included in the speech waveform according to a probability based on an acoustic model, and generating one or more words from the extracted text portion, A specific character in the extracted text portion is replaced with reference to probability data based on the acoustic model for the speech waveform.

  The present invention is a speech recognition program in an information terminal, which collects unique names based on a user's usage history for the terminal, creates a user dictionary, and inputs speech based on the speech waveform. Recognizing and converting to text, extracting a text part related to a proper noun from the textized voice, and replacing one or more words by replacing specific characters in the extracted text part And a step of replacing the text portion related to the proper noun with the word when the user dictionary includes the one or more words.

  According to the present invention, the text portion related to the extracted proper noun is compared with the user dictionary after the characters are recombined. As a result, even if there is an error in recognition by the acoustic model for the text part, the possibility of correction is increased, and as a result, it can be expected that the recognition accuracy is improved. In particular, when converting speech into text, by making the phoneme that is the next candidate for each phoneme the target of replacement, the number of replacements can be minimized, and the influence on the recognition speed can be minimized.

It is a block diagram which shows the structural example of the speech recognition apparatus in the information terminal which concerns on the Example of this invention. It is a figure showing the probability for every phoneme obtained by an acoustic analysis. It is a figure for demonstrating the function in a replacement word production | generation part. It is an example of the data structure of the specific name registered into a user dictionary. It is a flowchart of the speech recognition process in a speech recognition apparatus. It is the figure which represented typically the change of the output text in each analysis in a speech recognition apparatus.

  Next, embodiments of the present invention will be described in detail with reference to the drawings. Hereinafter, the details of the present invention will be described by taking a voice recognition device in an information terminal having a navigation function for a moving body as an example. In the speech recognition apparatus according to the embodiment, the accuracy of speech recognition is improved by utilizing information on the current position of the moving body and the travel history. As a form of this type of information terminal, an in-vehicle navigation device installed in a console of a vehicle, a portable information terminal that can be connected to an in-vehicle navigation device and can acquire information on a moving body, or has a navigation function by itself. A portable information terminal provided may be considered.

  FIG. 1 is a block diagram illustrating a configuration example of a voice recognition device in an information terminal according to an embodiment of the present invention. As shown in the figure, the speech recognition apparatus 100 includes a speech input unit 102, a speech recognition unit 104, a recognized text correction unit 106, a generated text output unit 108, a storage unit 110, and a control unit 112. Each of these functions can be realized by executing a program related to the voice recognition on a general-purpose computer having a CPU, a memory, a communication function, and the like.

  The voice input unit 102 inputs a user's voice from a microphone or the like and converts it into a voice waveform. The speech recognition unit 104 analyzes an input speech waveform and outputs text corresponding to the speech. The speech recognition unit 104 performs an analysis based on the acoustic model 114, and a language that performs an analysis based on the language model 118. Speech is recognized by analysis in two stages with the analysis unit 120. The acoustic analysis unit 116 separates the input speech waveform into small units such as phonemes, syllables, triphones (triple phonemes), and compares them with a large amount of speech waveform data (speech corpus) accumulated in the acoustic model 114. And the voice with the highest degree of approximation is output as the recognition result. For example, the acoustic analysis unit 116 can obtain an intermediate output as schematically shown in FIG. 2 for an input “kosuge”. The vertical arrangement in the figure represents the degree of likelihood of the corresponding candidate for each phoneme. For example, for the first phoneme of the input speech, “ko”, “ho”, “ka” are selected as candidates by comparison with the acoustic model, and the degree of their probability is 90%, 80%, 60% and so on. The analysis data in the acoustic analysis unit 116 is recorded in the storage unit 110 and is also used in the recognized text correction unit 106.

  The language analysis unit 120 linguistically arranges the text sequence obtained by the acoustic analysis unit 116 from a large amount of word data stored in the language model 118 and phrase / sentence data defining restrictions on word sequence and part of speech. To analyze. In order to express the restriction on the arrangement of words, the probability of the arrangement is expressed probabilistically by using an n.gram grammar that tabulates the appearance frequency in the arrangement of N words. In the speech recognition unit 104, the speech analysis result of the small unit in the acoustic analysis unit 116 is subjected to semantic correction for the input speech by linguistic analysis in the language analysis unit 120, and obtained as speech recognition output.

  The recognized text correction unit 106 acquires the output of the speech recognition unit 104 and further corrects it. The proper noun extraction unit 122, the replacement word generation unit 124, the replacement determination unit 126, the user dictionary generation unit 128, the user A dictionary 130 and a user dictionary management unit 132 are provided. The proper noun extraction unit 122 extracts a text part related to the proper noun in the output related to the speech from the speech recognition unit 104. In the extraction of proper nouns, the result of part-of-speech analysis by the language analysis unit 120 is used. The replacement word generation unit 124 generates one or a plurality of words by replacing characters constituting the word with respect to the proper noun extracted by the proper noun extraction unit 122. At this time, the replacement word generation unit 124 uses the probability value of the phoneme candidate obtained by the acoustic analysis unit 116 to limit the number of combinations. For example, the target candidates are limited by setting boundary values such as those having a probability value of 80% or more for the target phoneme and only the next candidate for the target phoneme. When only the next candidate phoneme, that is, “ho”, “u”, “ge” is targeted for the input speech “kosuge” shown in FIG. 2, as shown in FIG. The number of word combinations generated by the unit 124 is eight.

  The replacement determination unit 126 searches each word generated by the replacement word generation unit 124 to determine whether or not it is included in the user dictionary 130, and when the corresponding word exists in the dictionary, determines this as a replacement character string. To do. For example, in the case where the word “Kosuge” is registered in the user dictionary 130 and the word “Kosugi” recognized by the voice recognition unit 104 is not included. , Instead of “Kosugi”, “Kosuge” is adopted as the recognition character. As will be described later, each word in the user dictionary 130 has a priority based on a predetermined criterion. When a plurality of words are found in the dictionary among the combined words, a replacement word is determined according to the priority. To do.

  The user dictionary 130 is a database in which unique names acquired based on the usage status of the user in the navigation device are accumulated. The user dictionary generation unit 128 is connected to the navigation device 134 and acquires the usage status of the user therefrom. The unique name is extracted. Here, the current and past travel information of the vehicle is used to grasp the usage status of the user in the navigation device. Specifically, information on areas and facilities around the current position of the vehicle calculated by the current position calculation unit 134a, information on areas and facilities around the guidance route and destination constructed by the guidance route guide unit 134b, and Information on the facility searched by the facility search unit 134c is acquired from the map data 134d. In the preferred embodiment, the information acquired includes “reading”, “spelling”, “position coordinates”, and “registration date / time” information of the area or facility. FIG. 4 shows an example of the data structure of unique names registered in the user dictionary 130.

  The registered unique names are given priorities based on predetermined criteria. For example, the priority order is determined according to any of the appearance frequency, registration date and time, the distance from the current vehicle position in the original data, or a composite standard thereof, and the data is rearranged at the time of registration or reading. The user dictionary management unit 132 performs data management related to such data rearrangement and periodically executes a process of deleting old information from the user dictionary 130.

  The generated text output unit 108 replaces the word recognized by the replacement determination unit 126 with respect to the text recognized by the speech recognition unit 104, and passes this to the next processing as a speech recognition result. For example, when this speech recognition is used when using a social text posting service, the speech recognition result is passed to the application interface program of such a service, via a communication function provided in the information terminal, etc. Posting becomes possible. Further, when used for an operation such as facility search for the navigation device, the voice recognition result is input to the navigation device side, and the facility guidance program can be executed.

  The storage unit 110 stores temporary data generated at each processing stage in the speech recognition apparatus 100. Such data includes a speech waveform from the speech input unit 102, a phoneme extracted from the acoustic analysis unit 116 and its probability value, attribute information such as phrase / sentence data and its part of speech from the language analysis unit 120, and proper noun extraction. The proper noun extracted by the unit 122, the word group generated by the replacement word generation unit 124, the replacement word adopted by the replacement determination unit 126, and the generated text output by the generated text output unit 108 are included. The control unit 112 controls each function of the voice recognition device 100.

  Next, the process of speech recognition processing in the speech recognition apparatus will be described with reference to the flowchart of FIG. In the figure, the voice recognition process is started when a user speaks to a voice input device such as a microphone (step S502). The end of voice input is detected by detecting an explicit instruction by a user's button operation or the like, or detecting silence for a certain time (step S504), and the voice input unit 102 converts the input voice into voice waveform data. (Step S506). The acquired speech waveform is input to the speech recognition unit 104, and first subjected to acoustic analysis based on the acoustic model (step S508). In the acoustic analysis unit 116, the speech waveform is separated into small units such as phonemes, each is compared with the speech corpus, and the analysis text is determined according to the probability value. This analysis data is stored in the storage unit for later use. Next, the analysis text is subjected to language analysis based on the language model (step S510). The language analysis unit 120 analyzes the sequence of analysis texts from the word data and the phrase / sentence data in the language model, and expresses the probability of the sequence as a probability value. Then, a sequence with a high probability value is determined as the speech text.

  Next, the output text from the language analysis unit 120 is passed to the proper noun extraction unit 122, and words related to the proper nouns in the text are extracted (step S512). For the extraction of proper nouns, the result of part-of-speech analysis in the language analysis unit 120 is used. If there is no proper noun in the text, the process proceeds from step S514 to step S524, and the output text from the language analysis unit 120 is used as the final generated text and is output to the next process.

  If it is determined in step S514 that the word related to the proper noun is included in the text, next, one or more replacement words for the word are generated (step S516). As described above, the replacement criterion for each character in the word is based on referring to the probability value for each phoneme obtained by the acoustic analysis unit 116. Then, the user dictionary 130 is searched for each word generated here (step S518), and if the corresponding word exists in the dictionary, it is replaced with the proper noun extracted from the original text. The recognition text is completed, and this is output to the next processing as the final recognition text (steps S520 to 524). When a plurality of generated words are found in the user dictionary 130, a word having a higher priority is adopted as a replacement word in accordance with the priority criteria of the proper noun described above. On the other hand, if none of the plurality of generated words exists in the user dictionary 130, or if only the proper noun before replacement exists in the user dictionary 130, the process proceeds from step S520 to step S524, and language analysis is performed. The output text from the unit 120 is output as the final recognized text. As described above, the input speech from the user is recognized and output as text characters through each process of analysis using an acoustic model, analysis using a language model, and replacement of proper nouns using a user dictionary. is there.

  FIG. 6 schematically shows changes in the output text in each analysis in the speech recognition apparatus. Here, as an example, the user utters “Kobuchi Junction Arrival” in a situation where the user posts by voice to the social text posting service. The speech waveform according to this input is separated into individual phonemes, and recognition is performed by comparing with the speech corpus for each character by acoustic analysis. In this example, it is assumed that the phoneme of “ge” in the utterance of “kosuge” is recognized as “gi” having a higher probability value. Other characters shall be analyzed as uttered.

  Each phoneme “ko”, “su”, “gi”, “ja”, “n” ... “ku” of the speech waveform recognized by the acoustic analysis is converted to “kosugi”, “junction” by language analysis. And “arrival”. A character string of “kosugi”, which is a proper noun, is extracted from this text, and “kosugi”, “kosuge”, “hoso”, etc. are generated as combinations of the characters. In the user dictionary, only the proper noun “Kosuge” is hit against these candidates and selected as a replacement word. As a result, the recognition text “arrival of Kosugi junction” is converted into the text “arrival of Kosuge junction” and output as the final result in speech recognition.

  The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific embodiments, and various modifications and changes can be made within the scope of the gist of the invention described in the claims. It can be changed. In the above embodiment, the unique names managed by the user dictionary are collected from the data related to navigation, but other information stored in the information terminal or the navigation device, for example, personal information such as a person's name and address stored in the terminal May also be collected from address data including, information stored for viewing and video and music information, and other information relating to the user.

DESCRIPTION OF SYMBOLS 100: Speech recognition apparatus 102: Speech input part 104: Speech recognition part 106: Recognition text correction part 108: Generated text output part 110: Storage part 112: Control part 114: Acoustic model 116: Acoustic analysis part 118: Language model 120: Language analysis unit 122: proper noun extraction unit 124: replacement word generation unit 126: replacement determination unit 128: user dictionary generation unit 130: user dictionary 132: user dictionary management unit

Claims (9)

A speech recognition device in an information terminal,
Means for collecting a unique name based on a user's usage status for the terminal and creating a user dictionary;
Means for recognizing and text-inputting the input speech based on the speech waveform;
Means for extracting a text portion related to a proper noun from the text-formed speech;
Means for generating one or more words by replacing specific characters in the extracted text portion;
Means for replacing the text part of the proper noun with the word when the user dictionary includes the one or more words;
A speech recognition apparatus. The means for converting the speech into text includes means for determining each phoneme included in the speech waveform with a probability based on an acoustic model;
Means for generating one or more words from the extracted text portion replaces specific characters in the extracted text portion with reference to probability data based on the acoustic model for the speech waveform;
The speech recognition apparatus according to claim 1. The means for generating one or a plurality of words from the extracted text portion replaces the phoneme that is the next candidate for each phoneme with the corresponding phoneme in the extracted text portion when the speech is converted into text.
The speech recognition apparatus according to claim 2. The voice recognition device is a voice recognition device in an information terminal having a navigation function for a moving body,
The means for creating the user dictionary collects unique names related to the navigation based on the movement history of the user, and creates a user dictionary.
The speech recognition apparatus according to claim 1. The means for creating the user dictionary is configured by collecting a unique name including a place name and a facility name acquired from map data based on the current position of the moving body, the destination or the route from the current place to the destination value. Is,
The speech recognition apparatus according to claim 4. The user dictionary, for each unique name, is provided with each information of its reading, position coordinates, registration date and time, the priority is given based on the information,
The means for replacing the text portion related to the proper noun determines the word related to replacement according to the priority order in the user dictionary when there are a plurality of words related to the replacement.
The speech recognition apparatus according to claim 5. A speech recognition method in an information terminal,
Collecting a unique name based on a user's usage history for the terminal and creating a user dictionary;
Recognizing the input speech based on the speech waveform and converting it into text;
Extracting a text portion related to the proper noun from the textified speech;
Generating one or more words by replacing particular characters in the extracted text portion;
If the user dictionary contains the one or more words, replacing the text portion of the proper noun with the words;
A speech recognition method comprising: Converting the speech into text comprises determining each phoneme included in the speech waveform with a probability based on an acoustic model;
Generating one or more words from the extracted text portion refers to probability data based on the acoustic model for the speech waveform and replaces specific characters in the extracted text portion;
The speech recognition method according to claim 7. A speech recognition program for an information terminal,
Collecting a unique name based on a user's usage history for the terminal and creating a user dictionary;
Recognizing the input speech based on the speech waveform and converting it into text;
Extracting a text portion related to the proper noun from the textified speech;
Generating one or more words by replacing particular characters in the extracted text portion;
If the user dictionary contains the one or more words, replacing the text portion of the proper noun with the words;
A speech recognition program.

Images