JP2015087544A - Voice recognition device and voice recognition program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a voice recognition device and a program for identifying a phoneme string of a pronounced word even when pronunciation of a speaker is ambiguous.SOLUTION: A voice recognition device comprises: a storage section 5 having candidate phoneme storage means 13 for storing a candidate phoneme group including correct phonemes included in a correct phoneme string of a word, and/or similar phonemes whose pronunciation is similar to that of the correct phonemes, and which are associated with the correct phonemes, and priority of the respective elements of the candidate phoneme group; voice information input means 17 for generating voice information from a voice signal; assumption phoneme detection means 15 for detecting assumption phonemes to be assumed from the voice information in order of the pronunciation to obtain an assumption phoneme string; and pronunciation phoneme string identification means 19 for identifying a pronunciation phoneme string of the pronounced word from the assumption phoneme group, etc., in which the pronunciation phoneme string specification means 19 retrieves whether or not continuous elements in the assumption phoneme string are included in the candidate phoneme group, and when the continuous elements in the assumption phoneme group are included in the candidate phoneme group, an element of the assumption phoneme group with high priority is identified to be pronounced.

Description

  The present invention relates to a speech recognition apparatus and a speech recognition program.

  A speech recognition technique for identifying a word spoken by a speaker from speech information obtained by sampling and quantizing speech signals of words is known. As an application example of this speech recognition technology, Patent Literature 1 describes a pronunciation learning device that points out a speaker's pronunciation error. This pronunciation learning device converts a voice signal output from a microphone into voice information by an analog-to-digital converter, and detects a phoneme string of a word pronounced from the voice information. Then, this phoneme string is collated with a correct phoneme string of the word and a presumed error phoneme string to detect a pronunciation error of the speaker.

  The pronunciation learning device extracts an acoustic feature parameter from the speech information, and specifies a phoneme corresponding to the acoustic feature parameter as a phoneme of a pronounced word. However, it is difficult to specify a phoneme string from an ambiguous pronunciation of a speaker whose native language is not this word. Specifically, the audio information acquired from the ambiguous pronunciation may include an acoustic feature parameter in which a plurality of corresponding phonemes are detected and an acoustic feature parameter in which the corresponding phonemes are not detected. In this case, the phoneme string of the pronounced word could not be specified, and the speaker had to be asked to pronounce it again, or speech recognition had to be interrupted.

Japanese Patent Laid-Open No. 06-110494

  An object of the present invention is to provide a speech recognition apparatus and a speech recognition program that can specify a phoneme string of a pronounced word even if the pronunciation of the speaker is ambiguous.

The speech recognition apparatus according to the present invention includes a candidate phoneme group including a phoneme included in a correct phoneme string of a word and / or a similar phoneme similar in pronunciation to the phoneme, and the candidate A storage unit having candidate phoneme group storage means for storing the priority of each element of the phoneme group, an input unit having voice information input means for generating voice information from the voice signal of the pronounced word, and the voice information Assumed phoneme detection means for detecting assumed phonemes in the order of pronunciation and obtaining an assumed phoneme group; pronunciation phoneme string specifying means for specifying a pronunciation phoneme string of the word pronounced from the assumed phoneme group; A processing unit, and an output unit that outputs the pronunciation phoneme string,
The phoneme phoneme string specifying means searches for whether or not a continuous element in the assumed phoneme group is included in the candidate phoneme group, and when a continuous element in the assumed phoneme group is included in the candidate phoneme group The element of the assumed phoneme group having a high priority is identified as being pronounced.

  In the speech recognition apparatus of the present invention, the storage unit includes a phoneme feature storage unit that stores phoneme features that are feature information of the regular phoneme and the similar phoneme in association with the regular phoneme and the similar phoneme. The speech information input means samples and quantizes the speech signal to generate the speech information, and the assumed phoneme detection means calculates a partial speech feature from a part of the speech information, The regular phoneme or the similar phoneme corresponding to the speech feature that is at least partially in common with the partial speech feature is detected as an assumed phoneme.

  The speech recognition apparatus according to the present invention is characterized in that the candidate phoneme group is stored in association with the word.

  Furthermore, in the speech recognition apparatus of the present invention, the priority is determined from the speech information of the word, the speech signal, and / or the statistics of the assumed phonemes that are pronounced by a speaker who does not use the word as a native language. It is the occurrence frequency of candidate phonemes.

  In the speech recognition apparatus of the present invention, the priority is an approximation degree of the similar phonemes to the regular phonemes.

  Furthermore, in the speech recognition apparatus according to the present invention, the priority is an occurrence frequency of the candidate phonemes and an approximation degree of the similar phonemes with respect to the regular phonemes, and the phoneme phoneme string specifying means Are included in the candidate phoneme group, the element of the assumed phoneme group having the highest occurrence frequency is identified as being pronounced, and a plurality of candidate phonemes having the highest occurrence frequency are further specified. In some cases, an element of the assumed phoneme group having a high degree of approximation is specified as being pronounced.

The speech recognition program of the present invention is a candidate phoneme group including a normal phoneme included in a correct phoneme sequence of a word and / or a similar phoneme similar in pronunciation to the normal phoneme, And candidate phoneme group storage means for storing the priority of each element of the candidate phoneme group, voice information input means for generating voice information from the voice signal of the pronounced word, and an assumed phoneme assumed from the voice information Assumed phoneme detection means for obtaining an assumed phoneme group by detecting the order of pronunciation, a pronunciation phoneme string specifying means for specifying a pronunciation phoneme string of the word pronounced from the assumed phoneme group, and an output means for outputting the pronunciation phoneme string Function as,
The phoneme phoneme string specifying means searches for whether or not a continuous element in the assumed phoneme group is included in the candidate phoneme group, and when a continuous element in the assumed phoneme group is included in the candidate phoneme group The element of the assumed phoneme group having a high priority is identified as being pronounced.

  The speech recognition apparatus and the speech recognition program of the present invention can recognize a word that is pronounced based on the priority even when a speaker who does not speak the word that is the target of speech recognition is vaguely pronounced. It is. That is, in the speech recognition apparatus and speech recognition program of the present invention that have received an ambiguous pronunciation, the assumed phoneme detection means detects an assumed phoneme group that is assumed from the speech information. When elements similar in pronunciation to each other continue in the assumed phoneme group, the elements of the assumed phoneme group having a high priority are specified as the phonemes that are pronounced. For this reason, it does not occur that the speaker is asked to pronounce again or the speech recognition is stopped.

It is a block diagram of the speech recognition apparatus of this invention. It is a flowchart of voice recognition. It is a flowchart of a pronunciation phoneme specific step. It is a block diagram which shows the other Example of the speech recognition apparatus of this invention. (A) It is a figure which shows the other example of a candidate phoneme group memory | storage means, (b) It is an example which shows an assumed phoneme sequence set, (c) It is an example which shows an assumed phoneme group.

  The speech recognition apparatus of the present invention will be described taking speech recognition of the English word “apple” as an example. In the present specification, the same reference numerals shown in the drawings indicate the same or similar elements.

  As shown in FIG. 1, the speech recognition apparatus 1 of the present invention includes an input unit 3, a storage unit 5, a processing unit 7, and an output unit 9. The voice recognition device 1 is, for example, a computer such as a tablet terminal, a personal computer, a mobile phone terminal, or a voice recognition dedicated device.

  The storage unit 5 stores a program necessary for speech recognition, and can store, hold, and retrieve data relating to speech recognition. Typically, the storage unit 5 is provided in a computer, and the computer is Auxiliary storage devices such as a hard disk, a flash memory, and a dynamic random access memory that store programs for functioning as means. The storage unit 5 includes a regular phoneme string storage unit 11 and a candidate phoneme storage unit 13.

  The phoneme string storage means 11 stores a phoneme string that is a correct phoneme string of a word in association with the word. The phoneme sequence is a sequence of a plurality of phonemes arranged in the order of pronunciation. A regular phoneme is one phoneme that constitutes a correct pronunciation. In the present application, the phoneme is not limited to the phoneme symbol shown in each table, but includes information defined as indicating a phoneme.

この正音素列記憶手段11はテーブルであり、一のレコードに一の単語情報"apple"、及びその単語の正音素列（正音素１、正音素２、正音素３、及び正音素４）が格納される。この正音素列は、一のフィールドに対して一の正音素が格納される。 Table 1 shows an example of the regular phoneme string storage means 11.

The phoneme sequence storage means 11 is a table, and one word information “apple” and one phoneme sequence (the phoneme 1, the phoneme 2, the phoneme 3, and the phoneme 4) of the word are stored in one record. Stored. In the regular phoneme string, one regular phoneme is stored for one field.

  The candidate phoneme storage means 13 stores the candidate phoneme group and the priority. The candidate phoneme group is typically a set whose elements are phonemes that can be detected by the assumed phoneme detection means 15 described later with respect to a single phoneme constituting a word, and is similar to a word phoneme group. Consists of phonemes. This similar phoneme is a phoneme whose pronunciation is similar to a regular phoneme. Whether or not the pronunciation is similar to a regular phoneme can be determined based on, for example, the proximity of the phonetic distribution map of the international phonetic symbols. In addition, there is a tendency toward pronunciation by multiple speakers (Japanese) whose words (English) are non-native and / or speakers who do not speak standard words (speakers of so-called dialects) even if the word to be recognized is a native language. You may decide based on. In the present application, the normal phoneme and the similar phoneme stored in the candidate phoneme storage means 13 are referred to as candidate phonemes. The priority is stored in association with each candidate phoneme. This priority indicates the priority for selecting one candidate phoneme from the candidate phoneme group. Further, in order to enhance the learning effect, the priority of the regular phonemes may be set lower than other candidate phonemes.

この候補音素記憶手段13は、テーブルであり、一のレコードに対して、一の候補音素群（候補音素１、候補音素２、候補音素３、及び候補音素４）と、各候補音素の優先度（優先度１、優先度２、優先度３、及び優先度４）が格納される。また、各フィールドには、候補音素又はその優先度が格納される。なお、一の候補音素群を構成する候補音素の数は特に限定されない。また、優先度は候補音素の数に応じて増減するものである。 Table 2 shows an example of the candidate phoneme storage means 13.

This candidate phoneme storage means 13 is a table, and for one record, one candidate phoneme group (candidate phoneme 1, candidate phoneme 2, candidate phoneme 3, and candidate phoneme 4) and the priority of each candidate phoneme (Priority 1, Priority 2, Priority 3, and Priority 4) are stored. Each field stores candidate phonemes or their priorities. The number of candidate phonemes constituting one candidate phoneme group is not particularly limited. The priority is increased or decreased according to the number of candidate phonemes.

  The input unit 3 has voice information input means 17. This voice information input means 17 generates voice information from the voice signal of the pronounced word and inputs it to the processing unit 7. This audio information is a plurality of digital data obtained by sampling and quantizing an analog audio signal. The voice information input means 17 is typically a microphone that receives voice and an analog-digital converter that is electrically connected to the microphone. An amplifier and / or automatic gain control may be provided between the microphone and the analog-digital converter.

  The processing unit 7 identifies the phoneme string sequence of the pronounced word based on the voice information acquired from the input unit 3, and is typically a central processing unit of the terminal. As shown in FIG. 1, the processing unit 7 is connected to an input unit 3, a storage unit 5, and an output unit 9 so as to be communicable, and includes an assumed phoneme detection unit 15 and a pronunciation phoneme string identification unit 19.

The assumed phoneme detection means 15 analyzes the speech information acquired from the input unit 3 and detects an assumed phoneme group. The assumed phoneme group is, for example, a set having assumed phonemes arranged in the order of pronunciation as elements, and the assumed phonemes are phonemes corresponding to phoneme features that are at least partially in common with partial speech features.
Further, the assumed phoneme may be a phoneme corresponding to a phoneme feature having a predetermined common term with a partial speech feature. The partial voice feature is a feature calculated from a part of data included in the voice information, and is, for example, a characteristic obtained by frequency analysis of the part of the data, a frequency component, or the like. The phoneme feature to be compared with the partial speech feature is typically an acoustic model or the like indicating the frequency component of the phoneme, and is stored in the storage unit 5 in association with the phoneme. The detected assumed phonemes are individually stored in variables. The assumed phoneme detection means 15 is, for example, a library installed in an operation system such as iOS (registered trademark), Android (registered trademark), or Windows (registered trademark).

この想定音素群は、話者が発音した単語"apple"から得られたものであり、配列変数Soutei[]に格納される。この例においては、一の部分音声特徴に対する２つの想定音素がSoutei[0]及びSoutei[1]に格納され、他の部分音声特徴に対する想定音素がSoutei[2]に格納され、さらに他の部分音声特徴に対する想定音素がSoutei[3]に格納されている。さらに他の部分音声特徴に対する想定音素がSoutei[4]に格納されている。 Table 3 shows an example of the assumed phoneme group output.

This assumed phoneme group is obtained from the word “apple” pronounced by the speaker, and is stored in the array variable Soutei []. In this example, two assumed phonemes for one partial speech feature are stored in Soutei [0] and Soutei [1], an assumed phoneme for another partial speech feature is stored in Soutei [2], and another part Assumed phonemes for speech features are stored in Soutei [3]. Furthermore, assumed phonemes for other partial speech features are stored in Soutei [4].

  The phoneme string identification means 19 identifies a phoneme string that is a phoneme string of a pronounced word from the assumed phoneme group acquired from the assumed phoneme detection means 15, and is continuous in the assumed phoneme group. When elements (assumed phonemes) are included in the same candidate phoneme group, these consecutive elements (assumed phonemes) are a set of probable phonemes for a certain partial speech feature, and the highest priority among them. The element with high (speech phoneme) is identified as the phoneme that was pronounced. For example, when the assumed phoneme group Soutei [] shown in Table 3 is acquired, the values of consecutive Soutei [0] and Soutei [1] are included in the same candidate phoneme group, and therefore correspond to these values. The priorities are compared, and Soutei [1] having a high priority is identified as one phoneme that is pronounced. On the other hand, since the values of Soutei [2] to Soutei [4] are not included in the same candidate phoneme group, they are specified as other phonemes that are pronounced without comparing priorities. Therefore, Soutei [1], Soutei [2], Soutei [3], and Soutei [4] are specified as phoneme phoneme strings. This phoneme phoneme string is stored in, for example, the array variable Hatsuone [].

  The output unit 9 outputs the phoneme phoneme string acquired from the phoneme phoneme string specifying means 19. The output unit 9 is, for example, display means 21 provided in the terminal, and displays the phoneme phoneme string converted into text information. The output unit 9 is not limited to the display unit 21 and may output to a rating unit (not shown) that evaluates the pronunciation of the speaker based on the phoneme string. Further, the output unit 9 may further include an audio reproducing means for reproducing and outputting the audio. This voice reproducing means is, for example, a voice guide that outputs a voice of a normal phoneme string from a speaker. Thereby, the speaker can know the correct pronunciation sample of the word to be recognized. Further, the sound reproducing means may output the acquired sound information from a speaker. Thereby, the speaker can hear and compare the voice guide and his / her voice.

  Next, a voice recognition method will be described. As shown in FIG. 2, the speech recognition method includes a speech information input step (s10), an assumed phoneme detection step (s20), and a pronunciation phoneme string identification step (s30).

  The voice information input step (s10) is a step in which the voice information input means 17 generates voice information from the voice signal of the pronounced word and inputs this voice information to the processing unit. For example, a microphone receives a pronounced “apple” voice and converts it into an electrical signal, and an analog-digital converter samples and quantizes the electrical signal to generate voice information.

  The assumed phoneme detection step (s20) is a step in which the assumed phoneme detection means 15 acquires the speech information obtained in the speech information input step (s10) and compares the speech information with phoneme features to detect an assumed phoneme group. is there. For example, first, a plurality of partial voice features included in the voice information are detected. The method for detecting each partial voice feature is not particularly limited, but the voice information is divided into a set of data of a predetermined frame to generate a plurality of pieces of divided voice information, and frequency analysis is performed for each piece of the divided voice information. In addition, when continuous partial voice features are common, these may be combined into one partial voice feature. Next, a phoneme corresponding to a phoneme feature having a common term with a partial speech feature is extracted from the storage unit, and this phoneme is stored as an assumed phoneme in each element of the array variable Soutei [] (Table 3).

  The pronunciation phoneme string identification step (s30) is a step in which the pronunciation phoneme string identification means 19 identifies a pronunciation phoneme string that is a phoneme string of a word pronounced from the assumed phoneme group. This is a step of specifying a high-priority element (assumed phoneme) as a pronounced phoneme when the element (assumed phoneme) to be included is included in the same candidate phoneme group. An example of this phoneme phoneme string specifying step (s30) is shown in FIG. The phoneme sequence identification step (s30) includes an assumed phoneme group acquisition step (s301), an initialization step (s302), a candidate phoneme search step (s303), a first element identification step (s304), This step includes a setting step (s305), a priority comparison step (s306), an iterative determination step (s307, s308), and a second element specifying step (s309).

  The assumed phoneme group acquisition step (s301) is a step of acquiring the assumed phoneme group detected in the assumed phoneme detection step (s20). For example, the array variable Soutei [] is acquired.

  The initialization step (s302) is a step of storing an initial value in a variable used in each step described later. The variables are a reference value and an array element variable. The reference value stores one element (presumed phoneme) to be compared when comparing whether or not consecutive elements (assumed phonemes) are included in the same candidate phoneme group and the level of priority. Variable. Soutei [0] is stored as the initial value of the reference value. The array element variables are the element variable i of the array variable Soutei [] that stores the assumed phoneme group and the element variable j of the array variable Hatsuone [] that stores the phonemic phoneme string. 0 is stored as the initial value of the element variable i, and 1 is stored as the initial value of the element variable j.

  The candidate phoneme search step (s303) is a step of searching whether or not consecutive elements (assumed phonemes) in the assumed phoneme group are included in the same candidate phoneme group. In this example, continuous elements (assumed phonemes) are a reference value and a value stored in Soutei [i]. In order to search whether or not these consecutive elements (assumed phonemes) are included in the same candidate phoneme group, a candidate phoneme group (record) including a reference value is extracted from the candidate phoneme storage unit 13 and extracted. It is determined whether or not the value of the other element (assumed phoneme) Soutei [i] to be compared is included in the candidate phoneme group.

  In the above candidate phoneme search step (s303), when the continuous elements (assumed phonemes) are not included in the same candidate phoneme group, the process proceeds to the first element specifying step (s304). The first element specifying step (s304) is a step of specifying the value of the reference value as an element of the phoneme string sequence. In this example, the value of the reference value is stored in the element of the array variable Hatsuone [j] that stores the phonemic phoneme string, and the value of the element variable j is incremented.

  The reference value resetting step (s305) is a step of newly setting the other element (assumed phoneme) Soutei [i] as a reference value.

  On the other hand, in the above candidate phoneme search step (s303), when successive elements (assumed phonemes) are included in the same candidate phoneme group, the process proceeds to a priority comparison step (s306). The priority comparison step (s306) is a step of comparing priorities of consecutive elements (assumed phonemes). For the priority comparison, one priority corresponding to the reference value and another priority corresponding to the other element (assumed phoneme) Soutei [i] are extracted from the candidate phoneme storage means 13, and one priority and another Compares the priority levels of.

  In the priority comparison step (s306), if another priority is high, the process proceeds to the reference value resetting step (s305). By this step, another element (assumed phoneme) having a higher priority is set as a new reference value.

  On the other hand, when one priority is high in the priority comparison step (s306), and when the reference value resetting step (s305) is completed, the process proceeds to a repetition determination step (s307, s308). The repetition determination step (s307, s308) is a step for determining to repeat the above steps (s303) to (s305) for all assumed phonemes. In this example, the element variable i is incremented (s307), and the presence / absence of the value of the array variable Soutei [i] is determined (s308). If there is a value of the array variable Soutei [i], the process proceeds to a candidate phoneme search step (s303).

  On the other hand, if there is no value of the array variable Soutei [i] in the repetition determination step (s307, s308), the repetition process is terminated and the process proceeds to the second element identification step (s309). The second element identification step (s309) is a step of identifying the element (assumed phoneme) set as the reference value as the last element of the phoneme phoneme string, and storing the reference value in the element of the array variable Hatsuone [j] To do.

  Through the above steps, each phonetic element sequence element is stored in each element of Hatsuone [], and the pronunciation phoneme string is specified.

  The speech recognition apparatus 1 of the present invention is capable of speech recognition of a word that is pronounced based on priority even if it is an ambiguous pronunciation of a speaker who does not use a speech recognition target word as a mother tongue. That is, in the speech recognition apparatus 1 of the present invention that has received an ambiguous pronunciation, a probable phoneme group is detected by the assumed phoneme detection means 15. Then, based on the priority, it is determined which of the detected assumed phoneme groups is to be adopted. Thus, even when it is difficult to determine phonemes based on phoneme features such as an acoustic model, phonemes can be specified based on priority. For this reason, it does not occur that the speaker is asked to pronounce again or the speech recognition is stopped.

  This priority may be the frequency of occurrence of candidate phonemes. The frequency of occurrence of candidate phonemes is determined based on statistics of pronunciation information of a plurality of speakers. The pronunciation information is an audio signal, audio information, and / or an assumed phoneme group. The plurality of speakers are a plurality of speakers who do not use a speech recognition target word as their mother tongue. As a result, it is possible to specify a phoneme string sequence considering the pronunciation tendency of a speaker (for example, Japanese) who does not use the word (English) as a speech recognition target language. The plurality of speakers may be a plurality of speakers (so-called dialect speakers) who do not speak the standard language even if the recognition target word is a native language. Further, the generation frequency of candidate phonemes may be determined for each region. As a result, it is possible to perform speech recognition reflecting the pronunciation tendency of each region generated from a dialect. Further, the occurrence frequency of candidate phonemes may be determined based on statistics of pronunciation information accumulated by a specific user. As a result, it is possible to perform voice recognition that reflects a user-specific error tendency.

  Further, the priority may be an approximation degree of similar phonemes with respect to regular phonemes. This degree of approximation is determined by the proximity of each phoneme in the international phonetic symbol distribution map. Alternatively, it may be determined based on the degree of coincidence between the speech signals of the regular phoneme and the similar phoneme. The degree of coincidence is determined, for example, by observing the sound signal waveform of each phoneme with an oscilloscope.

  Furthermore, the priority may be a priority order of each candidate phoneme.

  Further, the priority of the present invention is the above-mentioned occurrence frequency and approximation degree, and the phonemic phoneme string specifying means 19 generates the occurrence when a continuous element (assumed phoneme) in the assumed phoneme group is included in the candidate phoneme group. The element with the highest frequency (assumed phoneme) is identified as being pronounced, and if there are multiple elements with the highest frequency of occurrence (assumed phonemes), the element with the highest degree of approximation (the assumed phoneme) was pronounced. It may be specified as a thing. As a result, the phoneme string sequence can be specified in consideration of the speaker's pronunciation tendency and the approximation of the sound.

  The speech recognition apparatus 1 of the present invention is not limited to the above example.

この候補音素記憶手段13は、複数の単語と、各単語に対応付けられた候補音素群及び候補音素群を構成する各要素（各候補音素）の優先度が、単語毎にレコードに記憶されたものである。 For example, as shown in Table 4, the candidate phoneme storage unit 13 may store a candidate phoneme group and a priority in association with a word. Moreover, you may memorize | store the priority of the element of a some candidate phoneme group and each candidate phoneme group.

The candidate phoneme storage means 13 stores a plurality of words, the candidate phoneme group associated with each word, and the priority of each element (each candidate phoneme) constituting the candidate phoneme group in a record for each word. Is.

  In addition, the input unit 3 of the present invention may include a word input unit 23 for designating a word to be subjected to speech recognition to the pronunciation phoneme string specifying unit 19 as shown in FIG. The word input means 23 is, for example, a character input device such as a keyboard for inputting a word, or a coordinate input device such as a mouse, a touch panel, or a digitizer for selecting a word displayed on the display means 21.

In the speech recognition apparatus 10, a word to be recognized is input by the word input means 23. Then, the phoneme phoneme string specifying means 19 specifies a candidate phoneme group (record) using the input word as a key. Thereby, voice recognition of a plurality of words becomes possible. Moreover, since the candidate phoneme group is defined for each word, similar phonemes that occur only in specific words and that occur frequently can be included in the candidate phoneme group. Thereby, the detection accuracy of the phoneme phoneme sequence is improved.

  Further, the candidate phoneme storage means 13 may be a table provided for each word. That is, one table (apple table (not shown)) that stores the candidate phoneme group and priority for one word “apple”, and another table that stores the candidate phoneme group and priority for another word “very”. A table (very table (not shown)) may be provided. In this case, when a word to be recognized is input by the word input unit 23, the phoneme phoneme string specifying unit 19 searches a table of the input words and specifies a candidate phoneme group.

  Further, as shown in FIG. 5, the candidate phoneme group 33 is a candidate phoneme string set 27 having the candidate phoneme strings 25a and 25b as elements, and the assumed phoneme group detected by the assumed phoneme detection means 15 includes the expected phoneme string 29 as an element. Assumed phoneme string set 31 may be used.

  The candidate phoneme string set 27 is stored in association with a word as shown in FIG. The candidate phoneme sequences 25a and 25b are correct phoneme sequences 25a of words or erroneous phoneme sequences 25b including similar phonemes. The candidate phoneme strings 25a and 25b are stored in the candidate phoneme storage unit 13 in association with the priority 35.

  Further, the assumed phoneme string 29 is formed by arranging assumed phonemes in the order of pronunciation, as shown in FIG. Then, the phoneme sequence identification unit 19 determines whether or not the consecutive elements of the assumed phoneme sequence set 31 (the assumed phoneme sequence (Soutei [0], Soutei [1])) are included in the same candidate phoneme sequence set 27. As a result, when consecutive elements (assumed phoneme string 29) are included in the same candidate phoneme string set 27, elements having higher priority (assumed phonemes) among consecutive elements (assumed phoneme string 29). The column 29) may be identified as a pronunciation phoneme sequence.

  Further, the assumed phoneme group may be a two-dimensional array as shown in FIG. In this assumed phoneme group, assumed phonemes are stored in each column so that one assumed phoneme column is determined in each row. In this case, the phoneme string specifying means 19 sets the consecutive elements as Soutei [0] [0] and Soutei [1] [0], and these consecutive elements are the same candidate phoneme group (Tables 2 and 4). To determine whether it is included. The phonemic phoneme string specifying means 19 may generate the assumed phoneme string set 31 by combining the phonemes of the columns for each row and then specify the phoneme phoneme string described above.

  Moreover, although the present invention has been described by taking English speech recognition as an example, the present invention is not limited to this, and speech recognition in other languages may be used.

  Further, the above speech recognition apparatus 1 of the present invention shows a preferred embodiment of the present invention, and the regular phoneme sequence storage means 11 of the storage unit 5 is not an essential configuration of the present invention.

  The candidate phoneme storage means 13 may be updatable. That is, the speech recognition apparatus of the present invention includes a transmission / reception unit that can be connected to a network, and the processing unit includes registration means for registering candidate phonemes and their priorities acquired via the transmission / reception unit in the candidate phoneme storage means 13, and transmission / reception You may provide the notification means which notifies a phonetic phoneme row | line | column via a part.

  Thus, a server that can communicate with the speech recognition apparatus 1 of the present invention via a network can collect phoneme phoneme strings from the plurality of speech recognition apparatuses 1 and store them in the storage unit of the server. Further, the processing unit of the server can generate candidate phonemes based on the collected phonological phoneme sequence statistics, and can transmit the generated candidate phonemes to each speech recognition apparatus. Then, the speech recognition apparatus 1 can newly register the acquired candidate phonemes and priorities. The speech recognition apparatus 1 can perform speech recognition based on pronunciation tendency of a plurality of geographically distant speakers.

  The speech recognition program of the present invention includes a computer that includes the above-described normal phoneme storage means 11, candidate phoneme storage means 13, assumed phoneme detection means 15, pronunciation phoneme string identification means 19, speech information input means 17, word input means 23, and display. This is to make it function as the means 21.

  Although the speech recognition apparatus and speech recognition program of the present invention have been described above, the present invention can be implemented in variously modified, modified, and modified forms based on the knowledge of those skilled in the art without departing from the spirit of the present invention. These embodiments are all within the scope of the present invention.

  The speech recognition apparatus of the present invention is used as speech recognition of words in a pronunciation training machine for practicing pronunciation of a non-native language. Moreover, you may use as voice recognition in the terminal provided with voice input.

1 ・ 10 ... Voice recognition device
3 Input section
5 ... Memory
7 ... Processing section
9 ... Output section
11 ... Means of storing phonemes
13 ... Candidate phoneme storage means
15 ... Assumed phoneme detection means
17 ... Voice information input means
19 ... Phoneme sequence identification method
21 ... Display means
23 ... Word input means

Claims (7)

Precise phonemes included in the correct phoneme string of a word and / or candidate phoneme group including similar phonemes that are similar in pronunciation to the phoneme and associated with the phoneme, and priority of each element of the candidate phoneme group A storage unit having candidate phoneme group storage means for storing
An input unit having voice information input means for generating voice information from the voice signal of the pronounced word;
Assumed phoneme detection means for detecting assumed phonemes from the speech information in the order of pronunciation and obtaining an assumed phoneme group, and a pronunciation phoneme string that identifies a phoneme string sequence of the word pronounced from the assumed phoneme group A processing unit having identification means;
An output unit for outputting the phonemic phoneme string;
With
The phoneme phoneme string specifying means searches for whether or not a continuous element in the assumed phoneme group is included in the candidate phoneme group, and when a continuous element in the assumed phoneme group is included in the candidate phoneme group A speech recognition device that identifies an element of the assumed phoneme group having a high priority as being pronounced. The storage unit includes phoneme feature storage means for storing phoneme features that are feature information of the regular phonemes and the similar phonemes in association with the regular phonemes and the similar phonemes,
The voice information input means samples and quantizes the voice signal to generate the voice information,
The assumed phoneme detection means calculates a partial speech feature from a part of the speech information and assumes the phoneme or the similar phoneme corresponding to the phoneme feature at least partially in common with the partial speech feature. The speech recognition apparatus according to claim 1, wherein   The speech recognition apparatus according to claim 1, wherein the candidate phoneme group is stored in association with the word.   The priority is an occurrence frequency of the elements of the candidate phoneme group determined from the speech information of the word pronounced by a plurality of speakers, the speech signal, and / or statistics of the assumed phoneme group. The speech recognition apparatus according to claim 3.   The speech recognition apparatus according to claim 1, wherein the priority is an approximation degree of the similar phoneme to the regular phoneme. The priority is the occurrence frequency of the elements of the candidate phoneme group determined from the speech information of the word pronounced by a speaker who does not speak the word, the speech signal, and / or the statistics of the assumed phoneme group, The degree of approximation of the similar phonemes to the regular phonemes,
The pronunciation phoneme string specifying means specifies that the element of the assumed phoneme group having the highest occurrence frequency is pronounced when a continuous element in the assumed phoneme group is included in the candidate phoneme group, Furthermore, when there are a plurality of the candidate phonemes having the highest occurrence frequency, the elements of the assumed phoneme group having the high degree of approximation are specified as being pronounced. Voice recognition device. Computer
Precise phonemes included in the correct phoneme string of a word and / or candidate phoneme group including similar phonemes that are similar in pronunciation to the phoneme and associated with the phoneme, and priority of each element of the candidate phoneme group Candidate phoneme group storage means for storing
Voice information input means for generating voice information from the voice signal of the pronounced word;
Assumed phoneme detection means for obtaining an assumed phoneme group by detecting assumed phonemes assumed from the sound information in the order of the pronunciation;
Pronunciation phoneme string identification means for identifying a pronunciation phoneme string of the word pronounced from the assumed phoneme group;
An output means for outputting the phonemic phoneme string;
Function as
The phoneme phoneme string specifying means searches for whether or not a continuous element in the assumed phoneme group is included in the candidate phoneme group, and when a continuous element in the assumed phoneme group is included in the candidate phoneme group A speech recognition program that identifies an element of the assumed phoneme group having a high priority as being pronounced.

Images