JP2019035935A - Voice recognition apparatus - Google Patents

Abstract

To provide a voice recognition apparatus having a recognition rate thereof improved under noises.SOLUTION: A voice recognition apparatus includes: an acoustic model-processing part for obtaining a voice input and generating an input spectrum, the spectrum of voice input; an acoustic model-storing part in which each spectrum of phonemes is stored in advance as an acoustic model; and a phoneme-matching part for calculating spectral intervals among the input spectrum and each spectrum of phonemes in the acoustic model, and for specifying the phoneme having a minimum spectral interval as a phoneme for voice input, wherein the phoneme-matching part weights values based on a margin among the input spectrum and each spectrum of phonemes in each of a plurality of frequencies, and calculates the spectral interval by calculating the total amount of the weighted values.SELECTED DRAWING: Figure 2

Description

  The present invention relates to a speech recognition apparatus.

  2. Description of the Related Art Conventionally, a speech recognition apparatus is known that compares the characteristics of an input speech with the features of each phoneme of an acoustic model created in advance, and identifies the phoneme determined to have a high degree of similarity as a phoneme label for the input speech.

  In such a speech recognition apparatus, it is required to improve the recognition rate under noise. As a related technique, Patent Document 1 discloses that noise is removed by subtracting a frequency component of sound input during non-speech from sound input during speech. Patent Document 2 calculates a distance between an input speech and an acoustic model by a weighted sum of a distance between each static feature parameter and a distance between dynamic feature parameters, and a phoneme having a small calculated value. Is identified based on the degree of dispersion of noise power over time.

Japanese Patent Laid-Open No. 02-179700 JP 2004-184856 A

  In order to improve the recognition rate under noise, it is conceivable to use an acoustic model created in an environment with the same noise characteristics or to subtract noise from speech input. When creating an acoustic model in a noisy environment, it is necessary to create a model for each environment in order to deal with a plurality of environments. Also, if noise is subtracted from the speech input in the spectral region, spectral distortion occurs, and the recognition rate may be lowered depending on the method of comparison with the acoustic model. Thus, there is room for improvement in improving the recognition rate under noise.

  The present invention has been made in view of the above problems, and an object of the present invention is to provide a speech recognition apparatus that improves the recognition rate under noise.

  In order to solve the above-described problem, according to one aspect of the present invention, an acoustic processing unit that acquires voice input and generates an input spectrum that is a spectrum of the voice input, and a spectrum of each phoneme are stored in advance as an acoustic model. Speech recognition comprising: an acoustic model storage unit; and a phoneme matching unit that calculates an inter-spectrum distance between an input spectrum and a spectrum of each phoneme of the acoustic model and identifies a phoneme having the smallest inter-spectral distance as a phoneme for speech input Device. The phoneme matching unit weights values based on the difference between the input spectrum and the spectrum of each phoneme at each of a plurality of frequencies, and calculates the inter-spectral distance by calculating the sum of the weighted values. By such a method for calculating the inter-spectral distance, it is possible to perform an evaluation that places importance on the frequency band that characterizes human voice and an evaluation that reduces the influence of a frequency band that contains a lot of noise, and the recognition rate can be improved.

  The acoustic processing unit further generates a noise spectrum that is a spectrum of noise included in the voice input, and the voice recognition device includes a weighting function storage unit that stores a plurality of noise spectrum patterns and weighting functions in association with each other. A weighting function determining unit that refers to the weighting function storage unit and determines a weighting function associated with the noise spectrum most similar to the noise spectrum generated by the acoustic processing unit, and the phoneme matching unit determines the weighting function Weighting may be performed based on the weighting function determined by the unit. As a result, it is possible to perform evaluation with a reduced influence according to the characteristics of noise included in the voice input, and to further improve the recognition rate.

  In addition, a language-specific weighting function storage unit that associates and stores a plurality of language types and weighting functions, and acquires a language type of speech included in the speech input, refers to the language-specific weighting function storage unit, and acquires the language A language-specific weighting function determining unit that determines a weighting function associated with the type, and the phoneme matching unit may perform weighting based on the weighting function determined by the language-specific weighting function determining unit. Thereby, according to the language type of the voice included in the voice input, it is possible to make an evaluation with an emphasis on the frequency band characterizing the language, and the recognition rate can be further improved.

  ADVANTAGE OF THE INVENTION According to this invention, the speech recognition apparatus which improved the recognition rate under noise can be provided.

Functional block diagram of the speech recognition apparatus according to the first embodiment of the present invention The flowchart which shows the process of the speech recognition apparatus which concerns on the 1st Embodiment of this invention. The figure explaining the distance between spectra concerning a 1st embodiment of the present invention. The figure which shows the example of the weighting function which concerns on the 1st Embodiment of this invention. Functional block diagram of the speech recognition apparatus according to the second embodiment of the present invention. The flowchart which shows the process of the speech recognition apparatus which concerns on the 2nd Embodiment of this invention. The figure which shows the example of matching with the noise spectrum which concerns on the 2nd Embodiment of this invention, and a weighting function Functional block diagram of a speech recognition apparatus according to a third embodiment of the present invention The flowchart which shows the process of the speech recognition apparatus which concerns on the 3rd Embodiment of this invention. The figure which shows the example of matching with the language classification and weighting function which concern on the 3rd Embodiment of this invention.

(Overview)
In the speech recognition apparatus according to the present invention, when calculating the inter-spectral distance between the spectrum of the speech input and the spectrum of the acoustic model, the difference in the specific frequency component is weighted more than the difference in the other frequency components. By such a method for calculating the inter-spectral distance, it is possible to perform an evaluation that places importance on the frequency band that characterizes human voice and an evaluation that reduces the influence of a frequency band that contains a lot of noise, and the recognition rate can be improved.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

(First embodiment)
<Configuration>
FIG. 1 shows a functional block diagram of the speech recognition apparatus 100 according to the present embodiment. The speech recognition apparatus 100 includes an acoustic processing unit 110, an acoustic model storage unit 120, and a phoneme matching unit 130. The acoustic processing unit 110 includes a high frequency emphasizing unit 111, an acoustic analysis unit 112, and a noise processing unit 113.

  The acoustic processing unit 110 receives an audio input from the outside, and generates an input spectrum that is a spectrum of the audio input. The acoustic model storage unit 120 stores a spectrum of each phoneme created in advance as a reference acoustic model. The phoneme matching unit 130 calculates an inter-spectrum distance that is a distance between the input spectrum generated by the acoustic processing unit 110 and the spectrum of each phoneme of the acoustic model stored in the acoustic model storage unit 120. Details of the inter-spectrum distance in this embodiment will be described later. Then, the phoneme having the shortest inter-spectral distance is identified as a phoneme with respect to the voice input, labeled, and a phoneme label is output.

<Operation>
Processing performed by the speech recognition apparatus 100 will be described. FIG. 2 shows a flowchart of processing performed by the speech recognition apparatus 100.

  Step S101: The acoustic processing unit 110 receives a voice input, and generates an input spectrum that is a spectrum based on the voice input. The voice input is framed using, for example, a Hamming window, and the following processing is performed in units of frames. In speech input, a high frequency region that tends to be a characteristic portion is generally emphasized by the high frequency emphasizing unit 111. Next, FFT (Fast Fourier Transform) is performed by the acoustic analysis unit 112, and a spectrum of the voice input is generated. The generated spectrum is subjected to noise reduction processing by the noise processing unit 113. For example, noise reduction is performed by subtracting the noise spectrum from the spectrum generated by the acoustic analysis unit 112. The spectrum of noise may be generated based on, for example, a low power frame or a frame designated as a non-voicing period in the voice input, or may be stored in advance as a noise model. The noise-reduced spectrum is output from the acoustic processing unit 110 as an input spectrum that is a voice input spectrum. In this embodiment, the acoustic processing unit 110 may not include the noise processing unit 113 and may output the spectrum generated by the acoustic analysis unit 112 as an input spectrum.

  Step S102: The phoneme matching unit 130 acquires the input spectrum generated by the acoustic processing unit 110, acquires each phoneme spectrum from the acoustic model storage unit 120, and sequentially determines the inter-spectrum distance between the input spectrum and the spectrum of each phoneme. calculate.

Here, the inter-spectrum distance will be described with reference to FIG. In the present embodiment, the inter-spectral distance is the difference between two spectra d _i (i = 1, 2,..., L) at a plurality of frequencies (angular frequencies) ω _i (i = 1, 2,..., L). Defined by the weighted sum of The weighting coefficient is defined by a function value F (ω _i ) (i = 1, 2,..., L) at each ω _i , defining the function F (ω) of the frequency ω as a weighting function. FIG. 4 shows a graph of the weighting function F (ω) according to an example. The inter-spectrum distance D _W is expressed by the following mathematical formula.

Here, the weighting coefficient F (ω _i ) (i = 1, 2,..., L) is determined in advance so that any value is different from the other values. That is, in calculating the inter-spectrum distance between the input spectrum and the spectrum of each phoneme, the influence of the difference at a specific frequency is made larger than the influence of the difference at another frequency. This makes it possible to hear a specific frequency in the same way as an auditory characteristic that listens mainly to a specific frequency band that characterizes the human voice consciously. Evaluation with emphasis on bands is possible. For example, if the weighting coefficient in a band including a frequency of 150 Hz and a band including a frequency in the range of 1 kHz to 2 kHz is larger than the weighting coefficient in other bands, the recognition rate of Japanese speech under vehicle noise is improved. be able to. The weighting coefficient is not limited to this, and may be designed as appropriate.

The phoneme matching unit 130 calculates the inter-spectral distance D _W between the input spectrum and each of the m phoneme spectra included in the acoustic model by the above-described equation. _Let D _Wj (j = 1, 2,..., _M ) be the inter-spectral distance between the input spectrum and the spectrum of the j-th phoneme.

Step S103: The phoneme matching unit 130 specifies a phoneme that gives the minimum value of the inter-spectral distance D _Wj (j = 1, 2,..., _M ). The speech recognition apparatus 100 outputs the phoneme specified in this way as a phoneme label for the speech input frame. The processing for one frame is completed as described above. However, when there is a frame to be processed next, the processing of steps S101 to S103 is executed for the frame.

<Effect>
In this embodiment, when comparing an input spectrum including noise with a phoneme spectrum, it is possible to improve the recognition rate by performing an evaluation focusing on a frequency band characterizing a human voice.

The inter-spectrum distance _DW is a weighted linear sum of the differences d _i , but is not limited as long as the weighting for each frequency can be performed, and may be a weighted square sum.

(Second Embodiment)
<Configuration>
FIG. 5 shows a functional block diagram of the speech recognition apparatus 200 according to the present embodiment. The speech recognition apparatus 200 further includes a weighting function determination unit 140 and a weighting function storage unit 150 in the speech recognition apparatus 100 according to the first embodiment. The speech recognition apparatus 200 is the first in that the weighting function determination unit 140 refers to the weighting function storage unit 150 to determine the weighting function F (ω) based on the noise spectrum pattern included in the speech input. This is different from the speech recognition apparatus 100 according to the embodiment. Components that are the same as or correspond to those of the speech recognition apparatus 100 according to the first embodiment are denoted by the same reference numerals.

<Operation>
Processing performed by the speech recognition apparatus 200 will be described. FIG. 6 shows a flowchart of processing performed by the speech recognition apparatus 200.

  Step S201: The acoustic processing unit 110 receives a voice input, and generates an input spectrum that is a spectrum based on the voice input. The processing in this step is the same as that in step S101 in the first embodiment, but in this embodiment, it is preferable to execute the noise reduction processing by the noise processing unit 113. The spectrum of noise used for noise reduction processing is generated based on voice input. For example, a spectrum can be generated based on a frame having a low power or a frame designated as a non-voicing period, among voice inputs. The noise-reduced spectrum is output from the acoustic processing unit 110 as an input spectrum that is a voice input spectrum. A noise spectrum is also output from the acoustic processing unit 110.

Step S202: The weighting function determination unit 140 refers to the weighting function storage unit 150 based on the noise spectrum and determines a weighting function. The weighting function storage unit 150 stores a noise spectrum pattern and a weighting function in association with each other. An example of this association is shown in FIG. Each spectrum of noise A, B, and C shown in FIG. 7 is a noise spectrum in a vehicle, an office, and a tunnel. Weighting functions F _A (ω), F _B (ω), and F _C (ω) are associated with the noises A, B, and C, respectively. The weighting function can be designed, for example, with a policy of increasing the weight of the band characterizing human speech and decreasing the weight of the band having a high noise level. The weighting function determination unit 140 selects a noise spectrum most similar to the noise spectrum acquired from the acoustic processing unit 110 from the noise spectrum stored in the weighting function storage unit 150, and weights associated with the selected noise spectrum Determine the function. The similarity determination of the noise spectrum can be performed based on, for example, an unweighted inter-spectral distance.

Step S203: The phoneme matching unit 130 acquires the input spectrum generated by the acoustic processing unit 110, acquires the weighting function determined by the weighting function determination unit 140, acquires each phoneme spectrum from the acoustic model storage unit 120, The inter-spectrum distance D _Wj (j = 1, 2,..., _M ) between the input spectrum and the spectrum of each phoneme is sequentially calculated.

Step S204: The phoneme matching unit 130 specifies a phoneme that gives the minimum value of the inter-spectral distance D _Wj (j = 1, 2,..., _M ). The speech recognition apparatus 200 outputs the phoneme specified in this way as a phoneme label for the speech input frame. The process for one frame is completed as described above, but if there is a frame to be processed next, the processes of steps S201 to S204 are executed for that frame.

<Effect>
In this embodiment, as in the first embodiment, when an input spectrum including noise is compared with a phoneme spectrum, evaluation is performed with emphasis on the frequency band that characterizes human voice, thereby improving the recognition rate. be able to. In the present embodiment, it is further possible to perform evaluation with a reduced influence according to the characteristics of noise, and to further improve the recognition rate.

(Third embodiment)
<Configuration>
FIG. 8 shows a functional block diagram of the speech recognition apparatus 300 according to the present embodiment. The speech recognition apparatus 300 includes a language-specific weighting function determination unit 141 instead of the weighting function determination unit 140 in the speech recognition apparatus 200 according to the second embodiment, and determines a language-specific weighting function instead of the weighting function storage unit 150. A portion 151 is provided. In the speech recognition apparatus 200 according to the second embodiment, the weighting function determination unit 140 refers to the weighting function storage unit 150 based on the noise spectrum pattern included in the speech input, and calculates the weighting function F (ω). In contrast, in the speech recognition apparatus 300 according to the present embodiment, the language-specific weighting function determining unit 141 refers to the language-specific weighting function storage unit 151 based on the language type of the speech included in the speech input. A weighting function F (ω) is determined. Moreover, the speech recognition apparatus 300 includes a word matching unit 160, a Japanese word storage unit 171, and an English word storage unit 172 as an example. Constituent elements similar or corresponding to those of the speech recognition apparatuses 100 and 200 according to the first and second embodiments are denoted by the same reference numerals.

<Operation>
Processing performed by the speech recognition apparatus 300 will be described. FIG. 9 shows a flowchart of processing performed by the speech recognition apparatus 300.

  Step S301: The acoustic processing unit 110 receives a voice input, and generates an input spectrum that is a spectrum based on the voice input. The processing in this step is the same as that in step S101 in the first embodiment.

  Step S302: The language-specific weighting function determination unit 141 determines a weighting function with reference to the language-specific weighting function storage unit 151 based on the language type of the speech included in the speech input. The method for determining the language type will be described later.

The language-specific weighting function storage unit 151 stores a language type and a weighting function in association with each other. An example of this association is shown in FIG. As shown in FIG. 10, weighting functions F _J (ω) and F _E (ω) are associated with Japanese and English, respectively. The weighting function can be designed, for example, with a policy of increasing the weight of the band that characterizes human speech in each language.

Japanese is composed of a structure of “C + V” in which phonemes are basically a consonant (C) followed by a vowel (V), and the appearance ratio of vowels is relatively high. Therefore, a vowel in which formants of 500 Hz or less are dominant is characteristic. In the present embodiment, as an example, the Japanese weighting function F _J (ω) is set to be generally uniform from low to mid to high, but gradually decreases, and the band that characterizes Japanese speech is used. Increase the weight.

In contrast, English has more types of consonants than Japanese, and phonemes are composed of “C + C + V”, “C”, “C + C”, etc. in addition to “C + V”, and the consonant appearance ratio is high. Therefore, a consonant in which the middle to high frequency components are dominant is characteristic. In the present embodiment, as an example, the weighting function F _E (ω) for English is set to increase from the low range to the middle to high range, and the weight of the band characterizing the English speech is increased.

Step S303: The phoneme matching unit 130 acquires the input spectrum generated by the acoustic processing unit 110, acquires the weighting function determined by the language-specific weighting function determination unit 141, and acquires each phoneme spectrum from the acoustic model storage unit 120. Then, the inter-spectral distance D _Wj (j = 1, 2,..., _M ) between the input spectrum and the spectrum of each phoneme is sequentially calculated.

Step S304: The phoneme matching unit 130 specifies a phoneme that gives the minimum value of the inter-spectrum distance D _Wj (j = 1, 2,..., _M ) as a phoneme label for the speech input frame. The phoneme matching unit 130 outputs the phoneme specified in this way as a phoneme label for the voice input frame.

  Step S305: The word matching unit 160 sequentially receives phoneme labels output from the phoneme matching unit 130, and performs word matching. For example, the word matching is performed by associating a word having the highest similarity, that is, a probability (likelihood) with the sequence of phoneme labels received sequentially. As a candidate word, a Japanese word is stored in the Japanese word storage unit 171, and an English word is stored in the English word storage unit 172. The word matching unit 160 refers to both the Japanese word storage unit 171 and the English word storage unit 172 to identify the word with the highest probability. The word matching unit 160 outputs this word as the output of the speech recognition device 300. The process for one frame is completed as described above. However, when there is a frame to be processed next, the processes of steps S301 to S305 are executed for the frame.

  Here, an example of a method for determining the language type of speech included in speech input will be described. In the process of step S305, the word matching unit 160 selects, from the Japanese word storage unit 171 and the English word storage unit 172, a word having a higher probability, for example, the 10th or lower word, and selects the selected Japanese The average value of the probabilities of words is compared with the average value of the probabilities of selected English words. The word matching unit 160 determines the language with the higher probability average value as the language type of the speech included in the speech input, and notifies the language-specific weighting function determination unit 141 of the determination result. In step S302, the language-specific weighting function determination unit 141 determines a weighting function based on the notified determination result. In step S303, the phoneme matching unit 130 performs this weighting function on the input spectrum to be processed. The process using is performed. The language type determination process and the notification process by the word matching unit 160 may be executed constantly in step S305 or may be executed at predetermined intervals. Moreover, you may notify the weighting function determination part 141 classified by language of the result of having taken the majority of the determination result in multiple times.

  In the above description, since the speech recognition device 300 includes the word matching unit 160, the word matching unit 160 is different from the speech recognition devices 100 and 200 according to the first and second embodiments that output phoneme labels. To output the specified word. However, the speech recognition apparatus 300 may output a phoneme label instead of or in addition to the word. The speech recognition apparatus 300 may not include the word matching unit 160, the Japanese word storage unit 171, and the English word storage unit 172, and may output a phoneme label. In this case, instead of receiving the determination result from the word matching unit 160, the language-specific weighting function determining unit 141 receives the voice input and its spectrum from the acoustic processing unit 110, and determines the language by itself using various language identification methods. Alternatively, the result of the language type determination process performed by an external device based on the phoneme label may be received.

  In addition, although two languages, Japanese and English, are given as examples of languages, in this embodiment, the type and number of languages are not limited, and can be applied to other languages.

<Effect>
In this embodiment, as in the first embodiment, when an input spectrum including noise is compared with a phoneme spectrum, evaluation is performed with emphasis on the frequency band that characterizes human voice, thereby improving the recognition rate. be able to. Furthermore, according to the present embodiment, according to the language type of the speech included in the speech input, it is possible to perform an evaluation that particularly emphasizes the frequency band that characterizes the language, and the recognition rate can be further improved.

  In addition, this invention can be grasped | ascertained not only as a structure of the functional block of a speech recognition apparatus but the speech recognition method and program which a computer provided with a processor performs.

  The present invention is useful for speech recognition devices and the like.

100, 200, 300 Speech recognition device 110 Acoustic processing unit 111 High frequency enhancement unit 112 Acoustic analysis unit 113 Noise processing unit 120 Acoustic model storage unit 130 Phoneme matching unit 140 Weighting function determination unit 141 Language-specific weighting function determination unit 150 Weighting function storage 151 Language-specific weighting function storage unit 160 Word matching unit 171 Japanese word storage unit 172 English word storage unit

Claims (3)

An acoustic processing unit that acquires a voice input and generates an input spectrum that is a spectrum of the voice input;
An acoustic model storage unit that stores in advance the spectrum of each phoneme as an acoustic model;
A speech recognition apparatus comprising: a phoneme matching unit that calculates a distance between spectra of the input spectrum and a spectrum of each phoneme of the acoustic model, and specifies a phoneme having a minimum distance between the spectra as a phoneme for the speech input. There,
The phoneme matching unit weights a value based on a difference between the input spectrum and the spectrum of each phoneme at each of a plurality of frequencies, and calculates the inter-spectral distance by calculating a sum of the weighted values. A speech recognition device. The acoustic processing unit further generates a noise spectrum that is a spectrum of noise included in the voice input,
A weighting function storage unit that stores a plurality of noise spectrum patterns and weighting functions in association with each other;
A weighting function determining unit that refers to the weighting function storage unit and determines a weighting function associated with a noise spectrum that is most similar to the noise spectrum generated by the acoustic processing unit;
The speech recognition apparatus according to claim 1, wherein the phoneme matching unit performs the weighting based on a weighting function determined by the weighting function determination unit. A language-specific weighting function storage unit that associates and stores a plurality of language types and weighting functions;
A language-specific weighting function determining unit that acquires a language type of speech included in the speech input, refers to the language-specific weighting function storage unit, and determines a weighting function associated with the acquired language type; ,
The speech recognition apparatus according to claim 1, wherein the phoneme matching unit performs the weighting based on a weighting function determined by the language-specific weighting function determination unit.

Images