JP2003108188A - Voice recognizing device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a voice recognizing device which can recognize a voice with high precision by discriminating CODEC-dependent non-stationary noise with high precision even if the non-stationary noise is included in an inputted voice. SOLUTION: Viterbi search parts 2-1 to 2-N are provided which execute viterbi search algorithms using voice HMM, voiceless HMM, and noise HMM and noise HMM's of the viterbi search parts 2-1 to 2-N are modeled depending upon CODECs of various systems. A most likelihood search result selection part 4 selects the most likelihood search results out of search results of the viterbi search parts 2-1 to 2-N and the selected search result is outputted as a recognition result by a recognition output part 5.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device, and more particularly to a code / decoder of various systems (hereinafter, CODEC).
That is, the present invention relates to a voice recognition device capable of highly accurately recognizing even a voice that has been subjected to inherent non-linear distortion.

[0002]

2. Description of the Related Art Conventionally, in a mobile phone or the like, a voice recognition device that recognizes a voice by a Viterbi search (maximum likelihood route search) algorithm using a hidden Markov model (hereinafter referred to as HMM) has been adopted. There is.

For example, Japanese Patent Laid-Open No. 9-230886 discloses a voice recognition method for outputting a recognition result based on the similarity between an input voice and a voice HMM, and learning noises recorded for each real environment cause various noises. Noise-resistant voice HM for
It is described that M is created and a recognition result is output from the similarity between the noise resistant speech HMM and the input speech.

Further, Japanese Patent Laid-Open No. 2000-284792 discloses an external telephone by learning using a noise database.
After creating non-voice HMM parameters for each telephone type such as extension telephones, ordinary telephones, digital mobile telephones, etc., and using the non-voice HMM parameters to determine which telephone type the acoustic signal is from Performing voice recognition using the acoustic model for the phone type, and voice recognition is H
It is described that the Viterbi search algorithm using MM can be performed.

Further, various noises that fluctuate in a short time (hereinafter referred to as non-stationary noises), for example, unnecessary sounds emitted by the speaker such as cough, sneeze, and interjection, surrounding human voices, tap sounds, shoe sounds, vehicles. It is also possible to collect non-stationary noise sections as speech sections by collecting and modeling the engine sound and noise generated from disturbance of radio waves and using the obtained model as a noise HMM. Schult
z I. Rogina “Acoustic and Language Modeling of Human
and Nonhuman Noises for Human-to-human Spontaneou
s Speech Recognition, ”Proc.ICASSP 95, pp.293-296
(199) '' and `` T. Yamada et. Al: “Voice Activity Detect
ion using Non-speech Models and HMM composition, ”
Proc. HSC 2001, pp.131-134 (2001) ".

[0006]

By the way, various types of CODECs are adopted in mobile phones and the like according to their types. For example, CELP (Code Excited Linear) is used.
Prediction: CS-ACELP (Conjugate-StructureAlgebra), which is an encoding method based on code-excited linear prediction
ic Code Excited Linear Prediction) is used in some cases, and EVRC (En
Some have adopted the hanced Variable Rate Codec). The CODEC has a non-linear characteristic peculiar to the CODEC because its configuration is different depending on the method.
The passed signal is subjected to CODEC-dependent nonlinear distortion. In particular, CODECs having a narrow bandwidth are used in mobile phones and the like, and since the quantization is performed with a small number of quantizations, the nonlinear distortion in the transmitted voice is large, and the CODEC of the other device is used. There is also a large difference in nonlinear distortion depending on whether or not is adopted.

Noise having CODEC-dependent non-linear distortion causes a problem that the beginning detection of a voice section is particularly accelerated in speech recognition, the voice section is misidentified, and the rate of false recognition of voice is increased. Although it is a factor, the above-mentioned proposal only uses a noise HMM created by modeling various non-stationary noises, and that noise in input speech has CODEC-dependent nonlinear distortion. I have not considered anything.
Therefore, the above-mentioned fixed proposal has a problem that the accuracy of the voice section identification and the voice recognition is reduced due to the noise being CODEC-dependent. In addition, this is not only a problem of recognizing voice from mobile phones,
This is a problem that is common to voice recognition devices to which voice signals that have passed through various types of CODECs are input, although they differ in degree.

According to the present invention, in a voice recognition device for recognizing a voice by a Viterbi search algorithm using an HMM, even if the input voice contains non-stationary noise having CODEC-dependent nonlinear distortion, the noise interval is highly accurate. It is an object of the present invention to provide a voice recognition device capable of identifying a voice section and performing voice recognition with high accuracy.

[0009]

In order to solve the above problems, the present invention provides a maximum likelihood selection of a plurality of Viterbi search units and a search result of the plurality of Viterbi search units having the highest likelihood. A search result selection unit and a recognition result output unit that outputs the selected search result as a recognition result, and each of the plurality of Viterbi search units includes a voice HMM.
And a noise HMM, and the Viterbi search algorithm is executed using them, and the noise HMM of each of the plurality of bitter search units is modeled depending on each CODEC of various methods. One point is the first feature. A second feature of the present invention is that each of the plurality of Viterbi search units further includes a silent HMM. Further, in the present invention, the voice HMM is also a C of various systems.
The third feature is that it is modeled depending on each ODEC. The present invention also provides the noise HMM.
Has a fourth characteristic in that it is obtained by learning by using a noise database obtained by cutting out a voice in a noise section in the voice input through each of various CODECs. Furthermore, the present invention has a fifth feature in that the voice input to the voice input unit is input through a telephone line.

According to the first feature, the input voice has C
Even if non-stationary noise having ODEC-dependent non-linear distortion is included, the noise section can be identified with high accuracy, and the speech section and the speech can be identified with high accuracy. Further, according to the second and third characteristics, it is possible to identify the voice section and recognize the voice with higher accuracy. Further, according to the fourth feature, it is possible to easily create a noise HMM that depends on each of various types of CODECs.
Furthermore, according to the fifth feature, it is possible to identify the voice section and the voice recognition of a mobile phone or the like that is inputted through a telephone line with high accuracy.

[0011]

BEST MODE FOR CARRYING OUT THE INVENTION Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. FIG. 1 is a block diagram showing an embodiment of the present invention. In FIG. 1, the voice input through the voice input unit 1 includes a plurality of Viterbi search units 2.
, 1, 2-2, ..., 2-N and are processed in parallel.

The voice input unit 1 is for capturing the voice transmitted from the other device through the telephone line in a call using a mobile telephone, for example, and the other device adopts the captured voice. CODE according to CODEC
It is subject to C-dependent nonlinear distortion.

Viterbi search units 2-1, 2-2, ...
2-N will be described in more detail below, each of which includes a speech HMM, a silence HMM and a noise HMM,
By executing the Viterbi search algorithm using them, the speech section of the input speech, the silent section, and the noise section are identified, and the phonemes that compose the speech in the speech section are recognized, and the result is a dictionary for speech recognition. Recognize speech by applying grammar 3. Voice HMM, silence H
The MM and the noise HMM are created and stored in advance prior to the recognition processing. It should be noted that the silent section is originally a section having no voice, and if the section does not affect the identification of the voice section and is naturally identified by the identification of the voice section, the silent HMM does not need to be specially provided.

Viterbi search units 2-1, 2-2, ...
Each of 2-N constitutes a voice recognition system corresponding to each of various types of CODECs. These Viterbi search unit 2
Among the search results by -1, 2-2, ..., 2-N, the one with the highest likelihood is selected by the maximum likelihood search result selection unit 4, and the selected search result is recognized as the recognition result. It is output from the result output unit 5.

The voice HMM is a voice model used for identifying the voice section of the input voice and recognizing the phonemes that compose the voice, and is composed of a set of phoneme HMMs for the phonemes to be recognized. . Silent HM
M is a silence model used to identify the silence section of the input voice.

The noise HMM is a noise model used for identifying a noise section in input speech,
It consists of a set of noise HMMs for each of the various non-stationary noises. Further, the noise in the input voice is CODEC-dependent depending on the type of the other party with whom the call is being made, so the noise HMM is modeled depending on CODEC. Therefore, a plurality of Viterbi search units 2-1, 2-2, ... Corresponding to each of the various types of CODECs.
., 2-N are provided. That is, CODEC
, CODEC2, ..., When it is planned to talk to various parties using CODEC N, noise HM modeled depending on CODEC1
Viterbi search unit 2-1 including M, Viterbi search unit 2-including noise HMM modeled depending on CODEC 2
2, ..., Viterbi search unit 2-N including a noise HMM modeled depending on CODEC N is provided. It should be noted that it is not always necessary to provide the Viterbi search unit and the noise HMM for all the CODECs of the other devices, and it may be provided for the CODECs adopted in the typical model. If so, the structure is simplified and there is no practical problem.

Viterbi search units 2-1, 2-2, ...
The Viterbi search algorithm executed in parallel in 2-N is the phoneme HM that constitutes the input voice and the voice HMM.
M, a silent HMM, and a noise HMM of various non-stationary noises are compared to calculate a similarity, and a voice section, a silent section, and a noise section are identified based on the calculated values, and phonemes constituting a voice are recognized. Is. Noise H in this process
Noise H modeled as MM depending on CODEC
By using the MM, the influence of the CODEC-dependent non-linear distortion is removed, non-stationary noise is recognized, and the noise section is identified. Therefore, the speech section and the speech are recognized with high accuracy and erroneous. Recognition is reduced.

It should be noted that the above-described embodiment is particularly effective for noise HMMs.
Is modeled depending on CODEC. In addition to this, the speech HMM can also be modeled depending on CODEC. In this case, the accuracy of speech segment identification and speech recognition can be improved. Can be improved.

FIG. 2 is a waveform diagram showing a specific example of the input voice, and FIG. 3 is an explanatory diagram showing a specific example of the Viterbi search algorithm executed for the voice of FIG. In this example, the silent section-noise section-silent section-speech section-silent section-noise section-silent section is continuous, and either "red", "blue" or "yellow" is input to the speech section. It is assumed that Also, "si
"1" represents a silent HMM, and "noise1" to "no"
“Ise3” represents a noise HMM for each type of non-stationary noise, and “a”, “k”, “o”, and the like represent phoneme HMMs included in the speech HMM.

First, the first silence state of the input voice is identified by the maximum likelihood of the similarity between the voice and the silence HMM "sil", and during the silence interval, the silence HMM is detected.
The identification process using "sil" is continued. That is,
The silent section is identified by continuing the processing through the self-loop of the identification by the silent HMM “sil”.

When non-stationary noise is input, the similarity with the noise HMM becomes maximum likelihood.
The process of exiting the identification loop by il "and entering the noise section identification process through the transition path. This process includes noise and noise HMMs" noise1 "to" n "in the input speech.
This is a process of continuously performing a process through a self-loop based on that the degree of similarity with "noise3" is maximum likelihood to identify a noise section. A loop for recognizing which noise HMM among "noise1" to "noise3" Depending on the type of non-stationary noise, the noise interval is identified by this process in any case, and the following silent segment is identified in the same manner as the first silent segment.

Subsequently, for example, assuming that a voice of "red" is input, in that case, the recognition procedure shown in the upper part of FIG. 3 is executed. That is, when the voice "a" is input, "a" is identified and recognized based on the similarity between the phoneme HMM "a" and the phoneme HMM "a" being the maximum likelihood, and then "ka" is input. Phoneme HMM "k", "a"
"K" and "a" based on the maximum likelihood of similarity with
Are sequentially identified, and "ka" is recognized. The middle part of the voice section in FIG. 3 shows the recognition procedure executed when the "blue" voice is input, and the lower part shows the recognition procedure executed when the "yellow" voice is input. There is.

As described above, the voice recognition is performed together with the voice section identification, and the subsequent silent section, noise section, and silent section are also identified in the same manner as described above, but the noise before the voice section becomes the voice section. The problem that it is identified as included and the start of voice section detection is earlier than the actual
It is reduced by making the noise HMM model dependent on CODE. The search algorithm described above is merely an example, and it goes without saying that how the HMMs used in the Viterbi search are continuous and what phoneme HMM is used depends on the input voice.

FIG. 4 is a flow chart showing a specific example of a method of creating a noise HMM used in the present invention. Steps S1 to S5 are performed for each of the various CODECs, and in FIG.
An EVRC is assumed as ACELP and CODEC2, and an example of creating a noise HMM for each of them is illustrated. Hereinafter, a case of creating a noise HMM for CS-ACELP will be described.

First, through CS-ACELP, on the street, in buildings, offices, parks, roadsides in urban areas, station premises, etc.
Collect speech containing various non-stationary noises (step S
1). The various non-stationary noises contained in these collected voices include the non-linear distortion inherent in CS-ACELP. Next, the non-stationary noise in the noise section in the collected speech is cut out and classified by the type of noise source (step S
2). The non-stationary noise in the noise section can be cut out by cutting out the speech in the section with a high level, and the classification by noise source type is based on the noise characteristics of each noise source such as level, frequency, and waveform. Can be done based on. As described above, a noise database of non-stationary noise including non-linear distortion when CS-ACELP intervenes can be obtained.

Next, using this noise database, H
MM learning is performed (step S3) to obtain HMM parameters. HMM parameters obtained by this are stored in ROM
Alternatively, it is stored in RAM (step S4) and CS-AC
The noise HMM of the Viterbi search unit for ELP is set (step S5).

A noise HMM for EVRC or other CODEC can be created by similar steps. The noise HMM created in this way is
Since it is created by learning using the noise database of non-stationary noise including DEC-dependent nonlinear distortion, it is modeled depending on each CODEC.

The present invention can be used as a voice recognition device for a voice call in which a mobile phone is frequently used, but is not limited to this and can be used as a voice recognition device to which voice signals from various types of CODECs are input.

[0029]

As is apparent from the above description, according to the first aspect of the present invention, even if the input voice includes non-stationary noise that depends on CODEC, the non-stationary noise can be identified with high accuracy. It is possible to identify and identify a voice section and recognize a voice with high accuracy. Further, according to the inventions of claims 2 and 3, it is possible to identify the voice section and recognize the voice with higher accuracy. Further, according to the invention of claim 4, the noise HMM depending on each of the various types of CODECs is used.
Can be created easily. Further, according to the invention of claim 5, it is possible to identify the voice section and the voice recognition of a cellular phone or the like which is inputted through a telephone line with high accuracy.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention.

FIG. 2 is a waveform chart showing a specific example of input voice.

FIG. 3 is an explanatory diagram showing a specific example of a Viterbi search algorithm for the voice of FIG.

FIG. 4 is a flowchart showing a specific example of a method of creating a noise HMM used in the present invention.

[Explanation of symbols]

1 ... Voice input unit, 2-1 to 2-N ... Viterbi search unit,
3 ... Speech recognition dictionary / grammar, 4 ... Maximum likelihood search result selection unit, 5 ... Recognition result output unit

   ─────────────────────────────────────────────────── ─── Continued front page    (72) Inventor Toru Shimizu             2-15-1 Ohara, Kamifukuoka City, Saitama Stock             Company CAD Research Institute F term (reference) 5D015 GG00 HH23 KK02

Claims (5)

[Claims] 1. A voice recognition device based on a Viterbi search algorithm using a Hidden Markov Model (HMM), wherein a voice input unit to which a voice is input, and a plurality of Viterbi search units that process signals from the voice input unit in parallel are provided. A maximum likelihood search result selection unit that selects the most likely one of the search results by the plurality of Viterbi search units, and a recognition that outputs the search result selected by the maximum likelihood search result selection unit as a recognition result. A result output unit, each of the plurality of Viterbi search units includes a speech HMM and a noise HMM, and executes a Viterbi search algorithm using them, and the noise HMM of each of the plurality of Viterbi search units. Is a speech recognition apparatus characterized by being modeled depending on each of various types of CODECs. 2. The voice recognition device according to claim 1, wherein each of the plurality of Viterbi search units further includes a silent HMM. 3. The voice HMM is also a CODE of various types.
The speech recognition apparatus according to claim 1 or 2, wherein the speech recognition apparatus is modeled depending on each C. 4. The noise HMM is a CODE of various types.
4. The voice recognition according to claim 1, wherein the voice recognition is obtained by learning using a noise database obtained by cutting out a voice in a noise section in a voice input through each of C. apparatus. 5. The voice recognition device according to claim 1, wherein the voice input to the voice input unit is input via a telephone line.