JP2012252026A - Voice recognition device, voice recognition method, and voice recognition program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve accuracy of voice recognition.SOLUTION: A voice recognition device 100 comprises: a voice processor 130 that extracts a word included in voice data and calculates a reliability degree of each word; a speaker degree calculating section 140 that calculates a speaker degree for each word indicating voice likeness of a speaker of the voice data corresponding to the word; a score calculating section 150 that calculates a score of each word, on the basis of the reliability degree of the word calculated by the voice processor 130 and the speaker degree of the word calculated by the speaker degree calculating section 140; a threshold setting section 160 that sets a predetermined threshold; and a word selecting section 170 that selects the word to be employed as a voice recognition result, based on the score of each word and the predetermined threshold.

Description

  The present invention relates to a voice recognition device, a voice recognition method, and a voice recognition program.

  Conventionally, a technique for performing speech recognition by a statistical method using an acoustic model of human speech and a language model has been proposed (for example, see Patent Document 1).

JP 2008-58503 A

  However, in the technique described in the above-mentioned patent document, when the human voice is included in the noise that is not the voice of the speaker, the word extracted from the human voice included in the noise is erroneously adopted as the voice recognition result. May end up. For this reason, there exists a problem that voice recognition accuracy will fall.

  The present invention has been made to solve the above-described problems, and it is an object of the present invention to provide a speech recognition device, a speech recognition method, and a speech recognition program that can improve speech recognition accuracy. .

  The speech recognition apparatus according to the present invention extracts a word included in sound data and outputs a reliability of each word, and a speech processing unit that outputs the reliability of each word, The score of each word based on the speaker degree calculating means for calculating the speaker degree, the reliability of each word output by the speech processing means, and the speaker degree of each word calculated by the speaker degree calculating means And a word selection means for selecting a word to be adopted as a speech recognition result based on the score of each word and a predetermined threshold value.

  The speech recognition method according to the present invention is a speech recognition method executed by a speech recognition apparatus, which extracts a word included in sound data and outputs a reliability of each word; and each word Calculated by the speaker degree calculating step for calculating the speaker degree of each word indicating the speech quality of the speaker corresponding to the sound data, the reliability of each word output by the voice processing step, and the speaker degree calculating step A score calculation step for calculating the score of each word based on the degree of utterance of each word, and a word selection step for selecting a word to be adopted as a speech recognition result based on the score of each word and a predetermined threshold It is characterized by comprising.

  In addition, the speech recognition program according to the present invention allows a computer to extract a word included in sound data and output the reliability of each word, and the voice likeness of a speaker of sound data corresponding to each word. Utterance degree calculating means for calculating the degree of utterance of each word, and the reliability of each word output by the speech processing means and the degree of utterance of each word calculated by the utterance degree calculating means And a score calculating means for calculating the score of each word, and a word selecting means for selecting a word to be adopted as a speech recognition result based on the score of each word and a predetermined threshold value.

  In the speech recognition apparatus, the speech recognition method, and the speech recognition program, a word included in sound data is extracted, the reliability of each word is output, and an utterance that indicates the speech likeness of the speaker of the sound data corresponding to each word The degree is calculated. Then, the score of each word is calculated based on the reliability of each word and the degree of speaker of each word, and a word to be adopted as a speech recognition result is selected based on the score of each word and a predetermined threshold. In this way, by calculating the score of each word based on the degree of speaker of each word in addition to the reliability of each word, the word score extracted from the voice of the speaker and the word extracted from noise The difference from the score can be clarified, and the speech recognition accuracy can be improved.

  Further, the speaker degree calculation means includes a volume of each word, a likelihood of the speech model of each word, a likelihood of a noise model of each word, a spatial transfer characteristic of each word, a fundamental frequency of each word, and a voice quality of each word The degree of utterance of each word may be calculated using at least one of. In this case, at least one of the volume of each word, the likelihood of the speech model of each word, the likelihood of the noise model of each word, the spatial transfer characteristics of each word, the fundamental frequency of each word, and the voice quality of each word Since the difference between the speaker degree of the word extracted from the speech of the speaker and the speaker degree of the word extracted from the noise becomes clearer, the speech recognition accuracy can be further improved.

  Further, the speaker degree calculating means may calculate the speaker degree of each word using the volume of each word in the frequency band of human speech. In this case, the difference in volume between the word extracted from the speech of the speaker and the word extracted from noise that contains a lot of sounds in frequency bands other than human speech becomes clearer, and the difference in the degree of speaker is more pronounced. It becomes clear. For this reason, the voice recognition accuracy can be further improved.

  The speech recognition apparatus further includes threshold setting means for setting a predetermined threshold based on the maximum value of the score calculated by the score calculation means, and the word selection means has the predetermined threshold set by the threshold setting means. You may select the word employ | adopted as a speech recognition result using. In this case, it is possible to select a word to be adopted as a speech recognition result in a flexible manner corresponding to an increase / decrease in the score of all words due to the sound collection environment, and to suppress variations in the speech recognition result due to the sound collection environment.

  The speech processing means outputs the relative reliability of each word with reference to the maximum reliability of each word, and the score calculation means outputs the relative reliability of each word output by the speech processing means. The score of each word may be calculated using the degree. In this case, it is possible to select a word to be adopted as a speech recognition result by calculating a score flexibly corresponding to an increase or decrease in the reliability of all words due to the sound collection environment. Therefore, it is possible to further suppress the variation in the speech recognition result due to the sound collection environment.

  Further, the speaker degree calculating means calculates the relative speaker degree of each word on the basis of the maximum value of the speaker degree of each word, and the score calculating means calculates each speaker calculated by the speaker degree calculating means. You may calculate the score of each word using the relative speaker degree of a word. In this case, it is possible to select a word to be adopted as a speech recognition result by calculating a score flexibly corresponding to increase / decrease in the degree of utterance of all words due to the sound collection environment. For this reason, the dispersion | variation in the speech recognition result by sound collection environment can be suppressed more.

  Further, the word selection means further deletes a word that is discontinuous from the word string including the word having the maximum score after deleting a word that is not adopted as the speech recognition result based on the score of each word and a predetermined threshold. You may delete and employ | adopt the remaining word as a speech recognition result. In this case, erroneous adoption of words extracted from noise can be further reduced, and speech recognition accuracy can be further improved.

  According to the speech recognition apparatus, speech recognition method, and speech recognition program according to the present invention, speech recognition accuracy can be improved.

It is a block diagram which shows the function of one Embodiment of the speech recognition apparatus which employ | adopted the speech recognition method which concerns on this invention. It is a block diagram which shows the hardware constitutions of the speech recognition apparatus of FIG. It is a flowchart which shows the speech recognition procedure by the speech recognition apparatus of FIG. It is a figure which shows an example of the speech recognition result by the speech recognition apparatus of FIG. It is a figure which shows the other example of the speech recognition result by the speech recognition apparatus of FIG. It is a block diagram which shows the other hardware constitutions of the speech recognition apparatus of FIG. It is a block diagram which shows the structure of one Embodiment of the speech recognition program which employ | adopted the speech recognition method which concerns on this invention.

  Hereinafter, embodiments of a speech recognition apparatus and a speech recognition program that employ a speech recognition method according to the present invention will be described.

  FIG. 1 is a block diagram illustrating functions of a speech recognition apparatus 100 that employs a speech recognition method according to the present invention. The voice recognition apparatus 100 is an apparatus used to enable data input to an application by voice, for example.

  As shown in FIG. 1, the speech recognition apparatus 100 includes a sound data input unit 110, a feature amount calculation unit 120, a speech processing unit 130, an acoustic model holding unit 131, a language model holding unit 132, and dictionary data holding. A unit 133, a speaker degree calculating unit 140, a score calculating unit 150, a threshold setting unit 160, a threshold holding unit 161, a word selecting unit 170, and a speech recognition result output unit 180.

  The sound data input unit 110 is a part that acquires sound data using, for example, a microphone.

  The feature amount calculation unit 120 is a portion that calculates feature amount data indicating acoustic features from sound data, for example, every 10 ms time interval (frame). The feature amount data indicating the acoustic features is a speech recognition spectrum, and is data represented by a frequency such as MFCC (Mel Frequency Cepstrum Coefficient), for example.

  The sound processing unit 130 refers to the feature amount data calculated by the feature amount calculation unit 120 and the data stored in the acoustic model holding unit 131, the language model holding unit 132, and the dictionary data holding unit 133. This is a part that extracts words included in the sound data acquired by the data input unit 110 and outputs the reliability of each word.

  The acoustic model holding unit 131 is a part that stores phonemes and their spectra in association with each other. The language model holding unit 132 is a part that stores statistical information indicating the chain probability of words, characters, and the like. The dictionary data holding unit 133 is a part that stores, as word data, for example, phonemes and phonetic symbols that represent words and their pronunciation.

  The speaker degree calculating unit 140 is a part that calculates the speaker degree indicating the speech likeness of the speaker of the sound data corresponding to each word from the sound data or feature amount data corresponding to each word.

  The score calculation unit 150 calculates the score of each word based on the reliability of each word output by the speech processing unit 130 and the speaker level of each word calculated by the speaker level calculation unit 140 It is. In the present embodiment, an example will be described in which a high score is calculated as the reliability level and the speaker level increase. Note that a low score may be calculated as the reliability level and the speaker level increase.

  The threshold setting unit 160 refers to the highest score calculated by the score calculation unit 150 and the data stored in the threshold holding unit 161, and sets a threshold for selecting a word to be adopted as a speech recognition result. It is a part to do. The threshold value holding unit 161 is a part that stores, for example, the difference between the threshold value and the highest score value as the relative threshold value data. The threshold setting unit 160 sets the threshold by adding the highest score calculated by the score calculation unit 150 and the difference stored in the threshold holding unit 161. Thereby, the threshold value for selecting a word to be adopted as the speech recognition result is changed according to the highest score calculated by the score calculation unit 150.

  The word selection unit 170 is a part that selects a word to be adopted as a speech recognition result based on the score of each word calculated by the score calculation unit 150 and the threshold set by the threshold setting unit 160. In the present embodiment, a word having a score higher than the threshold set by the threshold setting unit 160 is selected.

  The speech recognition result output unit 180 is a part that outputs the word selected by the word selection unit 170 and displays it on, for example, a display screen of an application.

  FIG. 2 is a block diagram illustrating a hardware configuration of the speech recognition apparatus 100. The speech recognition apparatus 100 includes, as a hardware configuration, a CPU 11, a RAM 12, a ROM 13, an input device 14, an auxiliary storage device 15, a communication device 16, an output device 17, a reading device 18 of a storage medium 18a, It has. The function of each part of the voice recognition device 100 described above is realized by causing the RAM 12 or the like to read a program or data from the auxiliary storage device 15 or the reading device 18 and causing the CPU 11 to execute the program. The input device 14 is, for example, a microphone that constitutes the sound data input unit 110, and the output device 17 is, for example, a monitor that constitutes the speech recognition result output unit 180.

  FIG. 3 is a flowchart showing a speech recognition procedure executed by the speech recognition apparatus 100. In the speech recognition apparatus 100, first, sound data is acquired by the sound data input unit 110 (step S10), and feature amount data is calculated for each frame from the sound data by the feature amount calculation unit 120 (step S20).

  Subsequently, the voice processing unit 130 performs a process using a statistical method on the feature data, extracts words included in the sound data, and outputs the reliability of each word (step S30). As a specific example, first, a plurality of candidates (N best) of word strings (hereinafter referred to as hypotheses) are listed, and the reliability of the words forming each hypothesis is calculated. Further, the reliability of each hypothesis is calculated from the reliability of each word, and the hypothesis having the highest reliability is selected. The word that makes the selected hypothesis becomes the extraction result of the word included in the sound data. And for the selected hypothesis, together with data such as word breaks, phoneme string of each word, notation of each word, reading of each word, part of speech information of each word, time information of each word, and dependency information of each word , The reliability of each word is output. The time information of each word is represented by, for example, the start frame number and end frame number of the sound data corresponding to the word. The reliability of each word is information indicating the correct answer accuracy of each word, and is calculated based on the likelihood of the acoustic model of each word, the likelihood of the language model of each word, and the like. In the present embodiment, for example, GWPP (Generalized Word Posterior Probability) is calculated as the reliability of each word.

Subsequently, the speaker degree calculation unit 140 calculates the speaker degree of each word extracted by the speech processing unit 130 (step S40). In the present embodiment, the speaker degree of each word is calculated using the volume of each word. As a specific example, sound data corresponding to time information of each word is cut out for each word. Furthermore, sound data in the frequency band of human voice is cut out. The frequency band of the human voice is a frequency band in which the amplitude of the human voice is high, and is, for example, 300 Hz to 3.4 kHz. And the volume is calculated from the cut out sound data,
Speaker degree = volume. For example, the volume may be calculated as a time average value of the amplitude of the sound data, or may be calculated as a maximum value of the amplitude of the sound data. In the present embodiment, the relative speaker level of each word is calculated with reference to the maximum speaker level of each word. As a specific example, the difference between the speaker level of each word and the maximum value of the speaker level of each word is calculated. Note that the sound volume may be calculated using feature data instead of sound data, or both sound data and feature data may be used.

Subsequently, the score of each word is calculated by the score calculation unit 150 (step S50). In the present embodiment, the score is calculated based on the GWPP of each word and the relative speaker level of each word, for example, score = ₁₀ Log ₁₀ (GWPP) + relative speaker level.

Subsequently, the threshold value setting unit 160 sets a threshold value for selecting a word to be adopted as the voice recognition result (step S60). In the present embodiment, the threshold value is calculated by, for example, threshold = maximum score value + relative threshold value.

  Subsequently, the word selection unit 170 selects a word having a score higher than the threshold (step S70), and the selected word is output by the speech recognition result output unit 180 (step S80).

FIG. 4 shows a speech recognition result when a speaker utters “Hiroshima, Okonomiyaki” toward a microphone or the like constituting the sound data input unit 110 in an environment where human voice is included in noise. ing. In this example, −4 is stored in the threshold holding unit 161 as the relative threshold data. In the example shown in FIG. 4, “love”, “wait”, “mas”, “hiroshima”, “okonomiyaki”, “jewelry”, “shobo”, and “shareholder” are extracted. For these words, the GWPP calculation results are 0.008, 0.059, 0.03, 0.554, 0.708, 0.049, 0.014, 0.57. The calculation results of the speaker degree are −10, −7, −20, −5, −8, −2, −7, and −11. The calculation result of the relative speaker level is −8, −5, −18, −3, −6, 0, −5, −9 because the maximum value of the speaker level is −2. The score calculation results are -29, -17.3, -33.2, -5.6, -7.5, -13.1, -23.5, -11.4. Since the maximum value of the score is -5.6, the threshold value is -5.6-4 = -9.6.
Thus, “Hiroshima, Okonomiyaki” having a score higher than this threshold is adopted as a speech recognition result, and other words are not adopted.

  In the example of FIG. 4, the GWPP of “shareholder”, which is a word extracted from noise, is 0.57, and the GWPP of “Hiroshima”, which is a word extracted from the voice of the speaker, is 0.554. That is, the reliability of the word extracted from the noise is higher than the reliability of the word extracted from the speech of the speaker. On the other hand, the score of “shareholder” is −11.4, and the score of “Hiroshima” is −5.6. That is, the score of the word extracted from the noise is lower than the score of the word extracted from the speech of the speaker. As a result, “shareholders” who are more reliable than “Hiroshima” are not adopted as speech recognition results, but only “Hiroshima” and “okonomiyaki” extracted from the speech of the speaker are adopted as speech recognition results. Has been. Thus, according to the speech recognition apparatus 100, by calculating the score of each word based on the degree of speaker of each word in addition to the reliability of each word, the score of the word extracted from the speech of the speaker And the difference between the score of the word extracted from the noise and the speech recognition accuracy can be improved. In particular, since the speaker degree calculation unit 140 calculates the speaker degree using the volume of sound data corresponding to each word, it is extracted from the speaker degree and noise of words extracted from the voice of the speaker. The difference between the spoken word and the speaker level becomes clearer, and the speech recognition accuracy can be further improved.

  Further, the speaker degree calculation unit 140 calculates the speaker degree using the volume of each word in the frequency band of human speech. As a result, the difference in volume between the word extracted from the speech of the speaker and the word extracted from noise that contains a lot of sound in a frequency band other than human speech becomes clearer, and the difference in the degree of speaker is more pronounced. It becomes clear. For this reason, the voice recognition accuracy can be further improved.

  The speech recognition apparatus 100 further includes a threshold setting unit 160 that sets a threshold based on the maximum value of the score calculated by the score calculation unit 150, and the word selection unit 170 includes the threshold set by the threshold setting unit 160. Words are selected using. For this reason, it is possible to select words flexibly corresponding to the increase / decrease in the score of all words due to the sound collection environment, and it is possible to suppress variations in the speech recognition results due to the sound collection environment.

  Further, the speaker degree calculation unit 140 calculates the relative speaker degree of each word based on the maximum value of the speaker degree of each word, and the score calculation unit 150 is calculated by the speaker degree calculation unit 140. The score of each word is calculated using the relative speaker degree of each word. For this reason, it is possible to select a word by calculating a score in a flexible manner corresponding to the increase or decrease in the degree of speaker of all words due to the sound collection environment. Therefore, it is possible to further suppress the variation in the speech recognition result due to the sound collection environment.

  The preferred embodiments of the speech recognition apparatus employing the speech recognition method according to the present invention have been described above. However, the present invention is not limited to the above-described embodiments, and various modifications can be made without departing from the scope of the present invention. Is possible.

The speaker degree calculation unit 140 calculates the speaker degree using the volume of each word, but instead of the volume of each word, the likelihood of the speech model of each word and the likelihood of the noise model of each word , Even if the speaker degree is calculated using any one of the spatial transfer characteristics of each word, the fundamental frequency of each word, and the voice quality of each word, the speaker degree of the word extracted from the speech of the speaker, It is possible to clarify the difference from the degree of speaker of a word extracted from noise. Also, two or more of the volume of each word, the likelihood of the speech model of each word, the likelihood of the noise model of each word, the spatial transfer characteristics of each word, the fundamental frequency of each word, and the voice quality of each word The speaker degree may be calculated using a combination. The speaker degree in this case is, for example, a speaker calculated from the volume, the noise likelihood ratio calculated based on the likelihood of the speech model and the likelihood of the noise model, the spatial transfer characteristic, the fundamental frequency, and the voice quality. With the degree adjustment value,
Speaker degree = volume−α × ₁₀ Log ₁₀ (noise likelihood ratio) + β × (speaker degree adjustment value)
Is calculated by α and β are predetermined weighting factors.

The likelihood of the speech model is calculated for each frame based on a GMM (Gaussian mixture model) learned from speech. The likelihood of the noise model is calculated for each frame based on the GMM learned from the noise. The likelihood of the speech model of each word and the likelihood of the noise model of each word are calculated as the sum of the likelihoods calculated for the frame corresponding to each word. Also, the noise likelihood ratio of each word is
Noise likelihood ratio = noise model likelihood / speech model likelihood. Since the voice of the speaker is emitted near a microphone or the like constituting the sound data input unit 110, the noise likelihood ratio of the sound data including the voice of the speaker is the noise of the sound data not including the voice of the speaker. It becomes lower than the likelihood ratio. For this reason, the difference in the noise likelihood ratio tends to be large between a word extracted from sound data including the speech of the speaker and a word extracted from sound data not including the speech of the speaker.

  The spatial transfer characteristic of each word is calculated as, for example, a linear prediction coefficient obtained by linear prediction analysis or a reverberation time. The reverberation time is the time required for the sound volume to decay to a predetermined value at the end of each word. The spatial transfer characteristic of each word varies depending on the degree of reverberation of the speech. Since the voice of the speaker is emitted near a microphone or the like constituting the sound data input unit 110, the reverberation of the sound data including the voice of the speaker is compared with the reverberation of the sound data not including the voice of the speaker. Few. For this reason, a difference in spatial transfer characteristics tends to be large between a word extracted from sound data including the speech of the speaker and a word extracted from sound data not including the speech of the speaker.

  The fundamental frequency of each word is, for example, a component in which a comb filter that transmits harmonics of the fundamental frequency F0 of speech is applied to the frequency power characteristics obtained by Fourier transform, and the power after passing through the filter is maximum. Is calculated as the frequency of. Alternatively, the fundamental frequency of each word is calculated as a frequency with a period having a maximum autocorrelation in the time domain of the speech waveform. Based on the fundamental frequency of each word, it is possible to grasp the sound quality of each word. Since the voice of the speaker is emitted near a microphone or the like constituting the sound data input unit 110, the basic frequency of the sound data including the voice of the speaker is the basic frequency of the sound data not including the voice of the speaker. Compared to voice-like values. For this reason, a difference in fundamental frequency tends to be large between a word extracted from sound data including the voice of the speaker and a word extracted from sound data not including the voice of the speaker.

  The voice quality of each word is calculated as, for example, a spectral slope coefficient of speech. The voice quality of each word indicates the tendency of the utterance method. For this reason, a difference in voice quality tends to be large between a word extracted from sound data including the voice of the speaker and a word extracted from sound data not including the voice of the speaker.

FIG. 5 shows the volume of each word, the likelihood of the speech model of each word, the likelihood of the noise model of each word, the spatial transfer characteristics of each word, the fundamental frequency of each word, under the same conditions as the example shown in FIG. The voice recognition results are shown when the score is calculated using a combination of all the voice qualities of each word. In this example, the threshold holding unit 161 stores −6 as relative threshold data. The weight coefficient α is set to 0.2, and the coefficient β is set to 0.2. In this case, the score calculation results are −33, −17.5, −33.4, −2.2, −2.5, −13.1, −23.3, and -13. Since the maximum value of the score is -2.2, the threshold value is -2.2-6 = -8.2.
Thus, “Hiroshima, Okonomiyaki” having a score higher than this threshold is adopted as a speech recognition result, and other words are not adopted.

  In the example of FIG. 5, the score of “shareholder” that is a word extracted from noise is −13, and the score of “Hiroshima” that is a word extracted from the voice of the speaker is −2.2. The difference in the scores of the words is −10.8. On the other hand, in the example of FIG. 4, the score of “shareholder” was −11.4, the score of Hiroshima was −5.6, and the difference between the scores of these words was −5.8. That is, in the example of FIG. 5, the difference between the score of the word extracted from the noise and the score extracted from the speech of the speaker is larger. Thus, two of the volume of each word, the likelihood of the speech model of each word, the likelihood of the noise model of each word, the spatial transfer characteristics of each word, the fundamental frequency of each word, and the voice quality of each word By using a combination of the above, it is possible to clarify the difference between the degree of utterance of words extracted from the speech of the speaker and the degree of utterance of words extracted from noise, and the speech recognition accuracy is further improved. Can be improved.

  The speech processing unit 130 outputs the relative reliability of each word with reference to the maximum reliability of each word, and the score calculation unit 150 calculates the relative value of each word output by the speech processing unit 130. The score of each word may be calculated using a certain degree of reliability. In this case, it is possible to select a word by calculating a score flexibly corresponding to an increase or decrease in the reliability of all words due to the sound collection environment. Therefore, it is possible to further suppress the variation in the speech recognition result due to the sound collection environment.

  In addition, the word selection unit 170 further deletes words that are discontinuous from the word string including the word having the maximum score after deleting a word that is not adopted as the speech recognition result based on the score and threshold value of each word. The remaining words may be adopted as the speech recognition result. In this case, in the example of FIG. 4, even if the third highest score “shareholder” exceeds the threshold, “shareholder” has the word string “Hiroshima, Okonomiyaki” including the maximum score “Hiroshima”. Deleted because it is discontinuous. In this way, it is possible to further reduce the erroneous adoption of words extracted from noise, and to further improve the speech recognition accuracy.

  Moreover, although the structure of FIG. 2 was shown as an example of the hardware structure of the speech recognition apparatus 100, it is not restricted to this. For example, the hardware configuration of the voice recognition device 100 may be a configuration in which functions are distributed to the client device 210 and the server device 220 connected via a network, as shown in FIG. For example, by configuring the sound data input unit 110, the feature amount calculation unit 120, the word selection unit 170, and the speech recognition result output unit 180 in the client device 210 and configure the remaining part in the server device 220, the client device 210 is configured. The calculation load can be reduced. In this case, feature amount data, score calculation results, threshold setting results, and the like are transmitted and received between the client device 210 and the server device 220. Therefore, these data can be compressed to reduce the load on the network. The division of functions between the client device 210 and the server device 220 is not limited to the above-described example. Furthermore, the client device 210 or the server device 220 may be further divided into a plurality of devices.

  The invention relating to the speech recognition apparatus 100 can be regarded as an invention relating to a speech recognition program for causing a computer to function as a speech recognition apparatus.

  FIG. 7 is a block diagram showing modules of a speech recognition program P100 for causing a computer to function as the speech recognition apparatus 100. The voice recognition program P100 of FIG. 7 includes a sound data input module P110, a feature amount calculation module P120, a voice processing module P130, a speaker degree calculation module P140, a score calculation module P150, a threshold setting module P160, a word A selection module P170 and a speech recognition result output module P180 are provided. Functions realized by executing the modules P110 to P180 are the same as the functions of the units 110 to 180 in FIG. The voice recognition program P100 is stored in, for example, the storage medium 18a shown in FIG. Examples of the storage medium 18a include storage media such as a flexible disk, a CD, and a DVD. The voice recognition program P100 may be provided to the voice recognition apparatus 100 via a wired network or a wireless network as a computer data signal superimposed on a carrier wave.

  DESCRIPTION OF SYMBOLS 100 ... Voice recognition apparatus, 130 ... Voice processing part, 140 ... Speaker degree calculation part, 150 ... Score calculation part, 160 ... Threshold setting part, 170 ... Word selection part, P100 ... Voice recognition program, P130 ... Voice processing module, P140 ... Speaker degree calculation module, P150 ... Score calculation module, P160 ... Threshold setting module, P170 ... Word selection module.

Claims (9)

Voice processing means for extracting words included in the sound data and outputting the reliability of each word;
A speaker degree calculating means for calculating a speaker degree of each word indicating the speech likeness of a speaker of sound data corresponding to each word;
Score calculating means for calculating the score of each word based on the reliability of each word output by the speech processing means and the speaker degree of each word calculated by the speaker degree calculating means;
Word selection means for selecting a word to be adopted as a speech recognition result based on the score of each word and a predetermined threshold;
A speech recognition apparatus characterized by that. The speaker degree calculating means includes a volume of each word, a likelihood of a speech model of each word, a likelihood of a noise model of each word, a spatial transfer characteristic of each word, a fundamental frequency of each word, and Calculating the degree of speaker of each word using at least one of the voice qualities of each word;
The speech recognition apparatus according to claim 1. The speaker degree calculating means calculates the speaker degree of each word using the volume of each word in the frequency band of human speech.
The speech recognition apparatus according to claim 2. The voice recognition device further includes threshold setting means for setting the predetermined threshold based on the maximum score calculated by the score calculating means,
The word selecting means selects a word to be adopted as a speech recognition result using the predetermined threshold set by the threshold setting means;
The speech recognition apparatus according to claim 1, wherein The speech processing means outputs the relative reliability of each word based on the maximum reliability of each word,
The score calculation means calculates the score of each word using the relative reliability of each word output by the voice processing means.
The speech recognition apparatus according to claim 1, wherein The speaker degree calculating means calculates the relative speaker degree of each word based on the maximum value of the speaker degree of each word,
The score calculating means calculates the score of each word using the relative speaker degree of each word calculated by the speaker degree calculating means;
The speech recognition apparatus according to claim 1, wherein The word selection means, after deleting a word that is not adopted as a speech recognition result based on the score of each word and the predetermined threshold, a word that is discontinuous with a word string including a word having the maximum score Delete further and adopt the remaining words as speech recognition results.
The speech recognition apparatus according to claim 1, wherein A speech recognition method executed by a speech recognition device,
A voice processing step of extracting words included in the sound data and outputting the reliability of each word;
A speaker degree calculating step of calculating a speaker degree of each word indicating the speech quality of the speaker of the sound data corresponding to each word;
A score calculation step of calculating a score of each word based on the reliability of each word output by the speech processing step and the speaker degree of each word calculated by the speaker degree calculation step;
A word selection step of selecting a word to be adopted as a speech recognition result based on the score of each word and a predetermined threshold;
A speech recognition method characterized by the above. Computer
Voice processing means for extracting words included in the sound data and outputting the reliability of each word;
A speaker degree calculating means for calculating a speaker degree of each word indicating the speech likeness of a speaker of sound data corresponding to each word;
Score calculating means for calculating the score of each word based on the reliability of each word output by the speech processing means and the speaker degree of each word calculated by the speaker degree calculating means;
Word selection means for selecting a word to be adopted as a speech recognition result based on the score of each word and a predetermined threshold;
A voice recognition program characterized by functioning as

Images