JP2010072020A - Voice recognition device, and transportation apparatus loaded therewith - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a voice recognition device dispensing with prior investigation for creation of a word dictionary. <P>SOLUTION: This voice recognition device 10 is equipped with an utterance section detection part 11 for detecting an utterance section including a voice signal from an acoustic signal; a characteristic quantity extraction part 12 for extracting an acoustic characteristic quantity from the voice signal by analyzing the voice signal in the utterance section detected by the utterance section detection part 11; a word dictionary 13 wherein a plurality of kinds of words are sorted by the mora number and registered; an acoustic model 14 wherein the acoustic characteristic quantity of each word registered in the word dictionary 13 is registered; and a collation part 15 for calculating likelihood of each word by collating the acoustic characteristic quantity extracted by the characteristic quantity extraction part 12 with the acoustic characteristic quantity of each word registered in the word dictionary 13 on reference to the word dictionary 13 and the acoustic model 14, and using a word having the highest likelihood as a recognition result. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a speech recognition device, and more particularly to a speech recognition device that recognizes words from speech signals.

  Transport equipment such as motorcycles, automobiles, ships, airplanes, and helicopters are equipped with electronic information equipment such as navigation systems, mobile phones, and audio equipment. Recently, voice recognition has been made so that these equipment can be operated by voice. The device is also installed.

  As one of speech recognition methods, a hidden Markov model (hereinafter abbreviated as “HMM”) is widely known (see Non-Patent Document 1). HMM can be obtained with high recognition accuracy by statistically learning fluctuations of speech such as a speaker and utterance fluctuations, and has become established as a speech recognition method in the present age.

  In Japanese Patent Laid-Open No. 2000-099077 (Patent Document 1), speech recognition is performed at a higher speed without performing collation with a dictionary that can be estimated from the length of input speech as having no possibility as a recognition result. A voice recognition device capable of performing the above is described. The speech recognition apparatus includes an utterance length measurement unit that measures the length of utterance of input speech, and a dictionary selection unit that selects a dictionary having a certain length range from a speech recognition dictionary prepared in advance. By estimating and narrowing down the dictionary of possible length as a recognition result from the utterance length, it is possible to avoid unnecessary collation processing of collating the dictionary with no possibility as a recognition result with the input. Is. In the speech recognition dictionary used here, a dictionary length is registered for each dictionary item (see FIG. 4). For example, 650 ms is registered as the dictionary head of “Shinjuku”. When a dictionary having a length of 60% to 160% of the utterance length L is allowed, assuming that the utterance length L is 500 ms, the dictionary selection unit selects a dictionary having a dictionary length of 300 ms to 800 ms.

However, in order to create such a speech recognition dictionary, the dictionary length must be examined in advance for all dictionary items. Therefore, this pre-investigation takes enormous effort.
JP 2000-099077 A JP 2007-206239 A Japanese Patent Application No. 2008-199713 (unpublished prior application of the same applicant) Lawrence Rabiner, Biing-Hwang Juang (co-author), Sadaaki Furui (supervised), "Basics of Speech Recognition", Chapter 6, NTT Advanced Technology, 1995

  An object of the present invention is to provide a speech recognition apparatus that does not require prior investigation for creation of a word dictionary.

Means for Solving the Problems and Effects of the Invention

  The speech recognition apparatus according to the present invention comprises storage means, feature quantity extraction means, utterance length identification means, word dictionary range designation means, and collation means. The storage means stores a word dictionary and an acoustic model. In the word dictionary, a plurality of types of words are registered together with an utterance length known in advance as an index of length when each word is uttered. In the acoustic model, an acoustic feature amount of a word registered in the word dictionary is registered. The feature quantity extraction unit analyzes the audio signal and extracts an acoustic feature quantity. The utterance length specifying means specifies the utterance length based on the voice signal. The word dictionary range designation means designates a range including words of the utterance length identified by the utterance length identification means in the word dictionary. The matching unit refers to the word dictionary and the acoustic model, and compares each acoustic feature amount extracted by the feature amount extracting unit with the acoustic feature amount of each word included in the range specified by the word dictionary range specifying unit. The likelihood of the word is calculated, and the word with the highest likelihood is used as the recognition result.

  According to the present invention, since words are registered in the word dictionary together with previously known utterance lengths (number of mora, number of vowels, vowel length, etc.), prior investigation is not required for creation of the word dictionary.

  Preferably, the words in the word dictionary are sorted by the number of mora. The storage means further stores a mora number address table in which start and / or end addresses of each mora number are registered. The word dictionary range designation means refers to the mora number address table, and designates a range including words of the utterance length specified by the utterance length specification means by the start and / or end addresses.

  In this case, it is only necessary to sort the words by the number of mora previously known as the utterance length, so that it is easy to create a word dictionary.

  Preferably, the speech recognition apparatus further includes an utterance section detection unit that detects an utterance section including the speech signal from the acoustic signal. The utterance length specifying means includes an utterance time measuring means for measuring the time of the utterance section detected by the utterance section detecting means as the utterance length.

  In this case, the range of the word dictionary can be narrowed down according to the measured time of the utterance section.

  Preferably, the utterance length specifying means includes vowel estimation means for estimating the number of vowels or the vowel length as the utterance length based on the acoustic feature quantity extracted by the feature quantity extraction means.

  In this case, the range of the word dictionary can be narrowed down according to the estimated number of vowels or vowel length.

  Preferably, the word dictionary is divided into a plurality of sub-word dictionaries for each number of mora. The word dictionary range designating unit designates a sub-word dictionary including words of the utterance length specified by the utterance length specifying unit.

  In this case, it is not necessary to register words in the order of addresses.

  Preferably, the word dictionary range designating unit designates a range including not only the utterance length word specified by the utterance length specifying unit but also an utterance length word within a predetermined range close to the specified utterance length.

  In this case, the correct word can be recognized even if the specified utterance length has some errors.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. In the drawings, the same or corresponding parts are denoted by the same reference numerals and description thereof will not be repeated.

[First Embodiment]
Referring to FIG. 1, a speech recognition apparatus 10 according to the first embodiment of the present invention includes an utterance section detection unit 11, a feature amount extraction unit 12, a word dictionary 13, an acoustic model 14, and a matching unit 15. Prepare. The word dictionary 13 and the acoustic model 14 are stored in a storage device such as a hard disk.

  The utterance period detection unit 11 detects an utterance period including an audio signal from an acoustic signal. More specifically, the utterance section detecting unit 11 divides the acoustic signal into frames every predetermined time (for example, 10 ms), performs fast Fourier transform on the divided acoustic signals for each frame, and converts the Fourier-transformed acoustic signal. A differential coefficient is calculated by differentiation, and a speech frame including speech is determined as an utterance section based on the frequency distribution of the calculated differential coefficient. For details, the description of Japanese Patent Application No. 2008-199713 (Patent Document 3) is incorporated herein.

  The feature amount extraction unit 12 extracts an acoustic feature amount from the speech signal by analyzing the speech signal in the speech segment detected by the speech segment detection unit 11.

  Referring to FIG. 2, a plurality of types of words are sorted and registered in word dictionary 13 according to the number of mora k. In the case of Japanese, the mora number k is approximately equal to the number of characters when a word is expressed in kana. For example, the word “tsu” with the number of mora k = 1 is registered at the address s1. Further, the word “Aki” with the mora number k = 2 is registered at the address s2, and the word “Yui” with the mora number k = 2 is registered at the address e2. Further, the word “Atami” with the mora number k = 3 is registered at the address s3, and the word “Seda” with the mora number k = 3 is registered at the address e3. Further, the word “Ariake” with the number of mora k = 4 is registered at the address s4. Further, the word “wakura onsen” with the number of mora k = 7 is registered at the address e7. Further, the word “Yukigaya Otsuka” with the number of mora k = 8 is registered at the address s8.

  In addition to the word dictionary 13 and the acoustic model 14, the storage device stores a mora number address table 16 as shown in FIG. In the mora number address table 16, a mora number k, a start address sk, and an end address ek are registered in association with each other. Each start address sk is the first address in the word dictionary 13 in which words corresponding to the number of mora k are stored. Each end address ek is the last address in the word dictionary 13 in which words corresponding to the number of mora k are stored. For example, since the word with the mora number k = 1 is stored at the start address s1 of the word dictionary 13, the start address s1 is registered in the mora number address table 16 corresponding to the mora number k = 1. Since there is only one word with the number of mora k = 1, the start address s1 = the end address e1. Therefore, the same end address e1 as the start address s1 is registered in the mora number address table 16 corresponding to the mora number k = 1. Further, since the word with the mora number k = 2 is stored from the start address s2 to the end address e2 of the word dictionary 13, the mora number address table 16 has a start address s2 and an end corresponding to the mora number k = 2. Address e2 is registered. The same applies to the other words with the mora number. In the present embodiment, as described above, the number of mora is registered as an index of time length when a word is uttered (hereinafter referred to as “speech length”).

  Referring to FIG. 1 again, the acoustic model 14 is registered with the acoustic feature quantities of words registered in the word dictionary 13 (for example, feature vectors such as spectrum and cepstrum). The acoustic model 14 models the acoustic features of the standard speech pattern for each word, and is reference information for evaluating the acoustic similarity with the input speech pattern. The acoustic parameters learned for each minute unit such as mora are registered.

  The word dictionary 13 and the acoustic model 14 are created by learning using learning data recorded in advance. The learning data is, for example, voice data that records utterances of various speakers in various situations.

  The collation unit 15 refers to the word dictionary 13 and the acoustic model 14, collates the acoustic feature amount extracted by the feature amount extraction unit 12 with the acoustic feature amount of each word registered in the word dictionary 13, and uses the HMM algorithm to The likelihood representing the likelihood of the word is calculated, and the word with the highest likelihood among the calculated likelihoods of each word is used as the recognition result. For details, the description of JP 2007-206239 A (Patent Document 2) is incorporated herein.

  The speech recognition apparatus 10 further includes an utterance time measuring unit 17 and a word dictionary range specifying unit 18. The utterance time measurement unit 17 measures the time of the utterance section detected by the utterance section detection unit 11 (hereinafter referred to as “speech time”) as the utterance length in order to specify the utterance length based on the audio signal. The word dictionary range specifying unit 18 includes a word including words having an utterance time measured by the utterance time measuring unit 17 in the word dictionary 13 and words having an utterance time within a predetermined range close to the utterance time (hereinafter referred to as “recognition target range”). ") 19 is designated. The words that the collation unit 15 collates with acoustic features are not all the words registered in the word dictionary 13 but only words included in the recognition target range 19 specified by the word dictionary range specification unit 18. Details will be described later.

  Next, the operation of the speech recognition apparatus 10 will be described with reference to the flowchart shown in FIG.

  For example, an acoustic signal as shown in FIG. 5A is input to the speech recognition apparatus 10. When the voice recognition device 10 is built in the navigation system, a voice (command) for instructing an operation is detected by a microphone, and a voice signal is input to the voice recognition device 10 from the microphone. The command may be composed of only one word, or may be composed of a short phrase of several words.

  First, the utterance section detector 11 detects the utterance section V from the start time ts to the end time te of the audio signal, as shown in FIG. 5B, from the input acoustic signal (S1). .

  Subsequently, the utterance time measurement unit 17 measures the utterance time T (= te−ts) of the utterance section V detected by the utterance section detection unit 11 as shown in FIG. 5C (S2).

Subsequently, the word dictionary range specification unit 18 estimates the number of mora L of words to be recognized based on the utterance time T measured by the utterance time measurement unit 17 (S3). Specifically, the mora number L (natural number) is determined within the range of the following equation (1).
αT ≦ L ≦ βT (1)

  Here, α and β are predetermined coefficients. For example, when 2 <αT ≦ 3 and 7 ≦ βT <8, the number of mora L is estimated to be 3-7.

  Subsequently, the word dictionary range specifying unit 18 specifies the recognition target range 19 in the word dictionary 13 based on the estimated number of mora L (S4). Specifically, referring to the mora number address table 16 shown in FIG. 3, the start address sk corresponding to the estimated minimum mora number Lmin is read, and the end address ek corresponding to the estimated maximum mora number Lmax is read. Is read. Then, in the word dictionary 13 shown in FIG. 2, a range from the start address sk to the end address ek is designated as the recognition target range 19. For example, when it is estimated that the number of mora L = 3 to 7, a start address s3 corresponding to L = 3 is read, and an end address e7 corresponding to L = 7 is read, and the end address is determined from the start address s3. The range up to e7 is designated as the recognition target range 19.

  On the other hand, the feature quantity extraction unit 12 extracts an acoustic feature quantity such as a mel cepstrum from the speech signal by analyzing the speech signal in the speech section detected by the speech section detection unit 11 (S5).

  Subsequently, the matching unit 15 refers to the word dictionary 13 and the acoustic model 14 and uses the acoustic feature amount extracted by the feature amount extraction unit 12 as the acoustic feature amount of each word designated by the word dictionary range designation unit 18. The likelihood of each word is calculated using the HMM algorithm, and the word with the highest likelihood among the calculated likelihoods of each word is used as the recognition result (S6).

  As described above, according to the first embodiment of the present invention, it is only necessary to sort and register words in the word dictionary 13 according to the number of mora known in advance as the utterance length. The word dictionary 13 can be easily created without prior investigation. Further, since the number of mora is estimated by measuring the utterance time and the range close to the number of mora is designated as the recognition target range 19, it is not necessary to collate with all the words in the word dictionary 13, and the voice Recognition can be performed in a short time. Further, by referring to the mora number address table 16, the recognition target range 19 can be quickly specified by the start address corresponding to the estimated minimum mora number and the end address corresponding to the estimated maximum address. In addition, since the recognition target range 19 is specified including not only the estimated number of mora but also the number of mora close thereto, a correct word can be recognized even if there is some error in the estimated number of mora. .

  In the present embodiment, both the start address and the end address are registered in the mora number address table 16. However, if all the words are registered in the order of addresses, only one of the start address and the end address needs to be registered.

[Second Embodiment]
Referring to FIG. 6, the speech recognition apparatus 20 according to the second embodiment of the present invention includes a vowel estimation unit 21 instead of the utterance time measurement unit 17 in the first embodiment. The vowel estimation unit 21 estimates the number of vowels or the vowel length as the utterance length based on the acoustic feature amount extracted by the feature amount extraction unit 12. The utterance section detection unit 11 in the first embodiment is not provided.

  In Japanese, the number of mora is almost equal to the number of vowels. The phoneme notation of the word shown in FIG. 2 is as shown in FIG. For example, the phoneme notation of the word “tsu” is / ts u / and the number of vowels n = 1. The phoneme notation of the word “aki” is / a k i / and the number of vowels n = 2. The phoneme notation of the word “Atami” is / a t a m i / and the number of vowels n = 3. Also, the phoneme notation of the word “Ariake” is / a r i a k e / and the number of vowels n = 4. All these words have the same number of mora as the number of vowels. On the other hand, a phoneme notation of a word including the repellent sound “n”, for example, “wakura onsen” is / wa kurao N se N /, and the number of vowels is n = 5. However, in this example, the phoneme / N / is also counted as a vowel, and is regarded as a vowel number n = 7 and a mora number k = 7. In addition, the prompt sound is regarded as one vowel, and the long sound is regarded as two vowels.

  Next, the operation of the speech recognition apparatus 20 will be described with reference to the flowchart shown in FIG.

  Unlike the first embodiment, the vowel estimation unit 21 estimates the vowel number K or the vowel length based on the acoustic feature amount extracted by the feature amount extraction unit 12 (S7).

  It is relatively easy to detect only a vowel from a speech signal as compared with general speech recognition. For example, attention is paid to the zero crossing number zc of each frame obtained by dividing the audio signal at predetermined time intervals. An appropriate minimum value min and maximum value max may be set in advance, the number of frames fn where min <zc <max is counted, and the vowel length may be estimated using fn. Alternatively, the average vowel length mean may be calculated and set in advance, and fn / mean may be estimated as the number of vowels. Alternatively, a vowel may be detected by creating an HMM or GMM (Gaussian Mixture Model) dedicated to vowel discrimination by learning. At this time, a likelihood threshold Lt may be set, and the number of frames satisfying Lt <Lf may be counted with respect to the likelihood Lf output from the HMM or GMM to estimate the vowel length. In addition, various methods known in the technical field of phoneme segmentation can be used.

Subsequently, the word dictionary range specifying unit 18 estimates the number of mora L of the word to be recognized based on the vowel number K or the vowel length estimated by the vowel estimation unit 21 (S3). Specifically, the mora number L (natural number) is determined within the range of the following equation (2).
Km ≦ L ≦ Kn (2)

  However, m and n are predetermined integers of 0 or more. For example, when k1-1 <Km ≦ k1 and k2 ≦ Kn <k2 + 1, it is estimated that the number of mora L = k1 to k2.

  Subsequently, the word dictionary range specifying unit 18 specifies the recognition target range 19 in the word dictionary 13 based on the estimated number of mora L (S4). When the number of mora L = k1 to k2, the range from the start address sk1 to the end address ek2 is designated as the recognition target range 19 in the word dictionary 13 shown in FIG.

  As described above, according to the second embodiment of the present invention, the number of vowels or vowel length is estimated based on the extracted acoustic feature amount, and further the number of mora is estimated. The target range 19 can be specified. Therefore, it is not necessary to provide the utterance section detection unit 11 as in the first embodiment.

[Third Embodiment]
Unlike the above embodiment, the word dictionary 13 in the third embodiment of the present invention is divided into N sub-word dictionaries 131 to 13N for each number of mora as shown in FIG. For example, the word “tsu” with the mora number k = 1 is registered in the sub-word dictionary 131, the word “aki” with the mora number k = 2 is registered in the sub-word dictionary 132, and the word “a” with the mora number k = 3. Are registered in the sub-word dictionary 133, the word “Higashi Funabashi” with the mora number k = 7 is registered in the sub-word dictionary 137, and the word with the mora number k = N is registered in the sub-word dictionary 13N. Has been.

  Further, instead of the mora number address table 16 shown in FIG. 3, a mora number dictionary table 30 is provided in the present embodiment as shown in FIG. In the mora number dictionary table 30, the mora number k and the sub-word dictionaries 131 to 13N are registered in association with each other.

  In the present embodiment, as shown in FIG. 11, the word dictionary range designation unit 18 selects one or more subword dictionaries from the subword dictionaries 131 to 13N based on the number of mora L estimated in step S3. Select (S8). When the number of mora L = k1 to k2, the sub word dictionaries 13k1 to 13k2 are selected from the word dictionary 13 and designated as the recognition target range 19.

  As described above, according to the third embodiment of the present invention, the word dictionary 13 is divided into a plurality of sub-word dictionaries 131 to 13N for each number of mora, so that the first and second embodiments There is no need to register words in the order of addresses.

[Usage]
The speech recognition apparatus according to the above embodiment is typically mounted on a motorcycle. For example, as shown in FIG. 12, a motorcycle 50 is worn by a passenger on a motorcycle 50, a vehicle-mounted information device 52, a vehicle-mounted information device 52, a helmet-side wireless communication device 54 </ b> A equipped in a helmet 53 </ b> A worn by a driver. A helmet-side wireless communication device 54B equipped on the helmet 53B is mounted. The voice recognition device is built in the in-vehicle information device 52.

  The motorcycle 50 includes a vehicle body frame 55, a power unit 56 attached to the vehicle body frame 55 so as to be swingable up and down, a rear wheel 57 that rotates by obtaining driving force from the power unit 56, a vehicle body A front wheel 59 as a steering wheel attached to the front portion of the frame 55 via a front fork 58 and a handle 60 that rotates integrally with the front fork 58 are provided. The handle 60 is provided with a main power switch 61.

  The power unit 56 is swingably connected to a lower portion near the center of the vehicle body frame 55 and is elastically coupled to the rear portion of the vehicle body frame 55 via a rear suspension unit 62. A driver's seat 63 is disposed at an upper portion near the center of the vehicle body frame 55, and a passenger's seat 64 is disposed behind the seat 63. In the vehicle body frame 55, a driver step 65 where the driver puts his / her foot is provided at a position between the seat 63 and the handle 60. Further, below the driver's seat 63, steps 66 are provided on both sides of the vehicle body frame 55 for the passenger to place his / her feet. In order to detect the riding state of the driver and the passenger, the seats 63 and 64 are provided with a driver seat seating sensor 67 and a passenger seat seating sensor 68, respectively.

  The in-vehicle communication device 51 is fixed to the vehicle body frame 55 at a position below the passenger seat 64. The in-vehicle communication device 51 is connected to an antenna 69 fixed to the vehicle body frame 55 behind the passenger seat 64, and performs wireless communication with the helmet-side wireless communication devices 54A and 54B. The in-vehicle information device 52 is fixed to the handle 60 and is further connected to the in-vehicle communication device 51 by wiring. Examples of the in-vehicle information device 52 include a navigation system that provides voice guidance of a travel route, a music player, a radio, and a telephone voice relay device that relays call voice of a mobile phone. The in-vehicle communication device 51 and the in-vehicle information device 52 operate by receiving power from the in-vehicle battery 70.

  On the inner surfaces of the helmets 53A and 53B, a pair of speakers 71 are fixed at positions facing the left and right ears of the occupant, and a microphone 72 is fixed at a position facing the occupant's mouth. On the other hand, helmet side wireless communication devices 54A and 54B are fixed to the back of the cap body. The helmet-side wireless communication devices 54A and 54B include an antenna 73 and are connected to a speaker 71 and a microphone 72.

  While the embodiments of the present invention have been described above, the above-described embodiments are merely examples for carrying out the present invention. Therefore, the present invention is not limited to the above-described embodiment, and can be implemented by appropriately modifying the above-described embodiment without departing from the spirit thereof.

It is a functional block diagram which shows the whole structure of the speech recognition apparatus by the 1st Embodiment of this invention. It is a figure which shows the structure of the word dictionary used for the speech recognition apparatus shown in FIG. It is a figure which shows the structure of the mora number address table used for the speech recognition apparatus shown in FIG. It is a flowchart which shows operation | movement of the speech recognition apparatus shown in FIG. It is a wave form diagram of the audio | voice signal for demonstrating the processing content by the utterance area detection part in FIG. 1, and the utterance time measurement part. It is a functional block diagram which shows the whole structure of the speech recognition apparatus by the 2nd Embodiment of this invention. It is a figure which shows the phonetic notation of a word in order to demonstrate the relationship between the number of mora and vowel number of the word dictionary used for the speech recognition apparatus shown in FIG. It is a flowchart which shows operation | movement of the speech recognition apparatus shown in FIG. It is a figure which shows the structure of the word dictionary used for the speech recognition apparatus by the 3rd Embodiment of this invention. It is a figure which shows the structure of the mora number dictionary table used for the speech recognition apparatus by the 3rd Embodiment of this invention. It is a flowchart which shows operation | movement of the speech recognition apparatus by the 3rd Embodiment of this invention. 1 is a side view showing an external configuration of a motorcycle equipped with a voice recognition device according to an embodiment of the present invention.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 Speech recognition apparatus 11 Speech section detection part 12 Feature-value extraction part 13 Word dictionary 14 Acoustic model 15 Collation part 16 Mora number address table 17 Speech time measurement part 18 Word dictionary range designation part 19 Recognition target range 20 Speech recognition apparatus 21 Vowel recognition Part 30 Mora number dictionary table 131-13N Sub-word dictionary

Claims (7)

A word dictionary in which a plurality of types of words are registered together with an utterance length known in advance as an index of length when each word is uttered, and an acoustic model in which acoustic features of words registered in the word dictionary are registered Storage means for storing
Feature quantity extraction means for analyzing the audio signal and extracting the acoustic feature quantity;
An utterance length identifying means for identifying an utterance length based on an audio signal;
A word dictionary range designating unit for designating a range including the word of the utterance length specified by the utterance length specifying unit in the word dictionary;
By referring to the word dictionary and the acoustic model, each acoustic feature amount extracted by the feature amount extracting unit is compared with an acoustic feature amount of each word included in the range specified by the word dictionary range specifying unit. A speech recognition apparatus comprising: a matching unit that calculates the likelihood of a word and uses the word with the highest likelihood as a recognition result. The speech recognition device according to claim 1,
The words in the word dictionary are sorted by the number of mora,
The storage means further stores a mora number address table in which start and / or end addresses of each mora number are registered,
The speech recognition apparatus, wherein the word dictionary range specifying means refers to the mora number address table, and specifies a range including words of the utterance length specified by the utterance length specifying means by the start and / or end addresses. The speech recognition device according to claim 1, further comprising:
Comprising an utterance section detecting means for detecting an utterance section including an audio signal from an acoustic signal;
The utterance length specifying means includes:
A speech recognition apparatus, comprising speech time measuring means for measuring the duration of the speech section detected by the speech section detecting means as the speech length. The speech recognition device according to claim 1,
The utterance length specifying means includes:
A speech recognition apparatus comprising: a vowel estimation unit that estimates a vowel number or a vowel length as the utterance length based on the acoustic feature amount extracted by the feature amount extraction unit. The speech recognition device according to claim 1,
The word dictionary is divided into a plurality of sub-word dictionaries for each number of mora,
The speech recognition apparatus, wherein the word dictionary range designation means designates a sub-word dictionary including words having an utterance length specified by the utterance length specifying means. The speech recognition device according to claim 1,
The word dictionary range designating unit designates a range including not only the utterance length word specified by the utterance length specifying unit but also a utterance length word within a predetermined range close to the specified utterance length. apparatus.   Transportation equipment carrying the voice recognition device according to any one of claims 1 to 6.

Images