JP2005084071A - Speech recognizing apparatus - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a speech recognizing apparatus capable of highly accurately recognizing a speech even if a speech not to be inputted to the recognizing apparatus, such as the voice of a person behind and a mutter is mixed when the inputted speech is recognized. <P>SOLUTION: A speech collation part 1 collates the speech inputted through a portable telephone with standard patterns for recognition and sends a recognition result out. As the standard patterns for recognition, standard patterns 2 for recognition including recognition vocabulary standard patterns 2-1 generated by modeling speeches to be recognized and background person standard patterns 2-1 added before and after the recognition vocabulary standard patterns 2-1 are used. The background person speech standard patterns are a background person's speech standard pattern group 2-3 obtained by learning based upon an input speech of low level which is inputted through a portable telephone transmission system and distorted more than a normal input speech to the portable telephone and a background person's speech garbage standard patterns 2-4 obtained by learning based upon the speech of the person behind which is included in the standard telephone speech. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a voice recognition device, and in particular, when a voice inputted via a mobile phone is recognized, even if voice other than the input purpose to the voice recognition device such as a human voice or a monologue in the background is mixed. The present invention relates to a speech recognition apparatus that can recognize speech with high accuracy.

  FIG. 8 is a block diagram showing a conventional speech recognition apparatus. A speech input through a mobile phone is collated with a recognition standard pattern 2 in a speech collation unit 1. The recognition standard pattern 2 is stored in advance for collation, and includes a recognition vocabulary standard pattern 2-1 or a combination of a recognition vocabulary standard pattern 2-1 and a speech standard pattern 2-2.

  FIG. 9 shows a recognition standard pattern 2 consisting only of the recognition vocabulary standard pattern 2-1, and FIG. 10 shows a recognition standard pattern 2 consisting of a combination of the recognition vocabulary standard pattern 2-1 and the speech standard pattern 2-2. Indicates. The voice collation unit 1 collates the input voice with the standard pattern for recognition 2, and sends a recognition vocabulary having the highest collation score (similarity) as a recognition result.

  When the recognition standard pattern 2 of FIG. 9 is used, the input speech whose both ends are separated by a silent section is recognized by the speech collating unit 1 as a recognition standard pattern composed of recognition vocabulary standard patterns (recognition vocabulary 1 to N) 2-1. 2 and the recognition vocabulary with the highest collation score is sent as a recognition result.

Further, when the recognition standard pattern 2 in FIG. 10 is used, the input speech whose both ends are separated by a silent section is an arbitrary Japanese speech composed of consonants and vowels before and after the recognition vocabulary standard pattern 2-1. It is collated with the recognition standard pattern 2 to which the received voice standard pattern 2-2 is added.
Tatsuya Kawahara, Toshihiko Sotsugi, Kiyoichi Miki, Shuji Doshita “Examination of language model for word spotting in conversational speech” IEICE Technical Report SP94-28 1994 PP.41-48

  However, when the recognition standard pattern 2 consisting only of the recognition vocabulary standard pattern 2-1 in FIG. 9 is used, if a human voice behind the input voice is mixed in the input voice, the human voice is collated with the recognition vocabulary. Since the recognition result may be transmitted, there is a problem that misrecognition increases.

  When the recognition standard pattern 2 including the combination of the recognition vocabulary standard pattern 2-1 and the speech standard pattern 2-2 in FIG. 10 is used, erroneous recognition due to mixing of human voices and the like in the background can be improved. There is a problem that a new misrecognition occurs when the vocabulary or part of the vocabulary is collated with the preceding and following speech standard patterns 2-2.

  An object of the present invention is to provide a voice recognition device capable of recognizing voice with high accuracy even when voice other than the input purpose to the voice recognition device such as a human voice or monologue in the background is mixed when the inputted voice is recognized. It is to provide.

  In order to solve the above-mentioned problem, the present invention provides a speech recognition apparatus having a speech collation unit that collates speech input via a mobile phone with a recognition standard pattern and sends a recognition result. The pattern is a recognition standard pattern including a recognition vocabulary standard pattern that models speech to be recognized and a background human voice standard pattern added before and after the recognition vocabulary standard pattern, and the background human voice standard pattern is A background human voice standard pattern group obtained by learning based on low-level input speech that is input via a mobile phone transmission system and receives large distortion compared to normal input speech to a mobile phone, or mobile phone It is a background human voice garbage standard pattern obtained by learning based on the human voice behind the telephone voice.

  Further, the present invention provides a speech recognition apparatus having a plurality of speech collating units and a judging unit that collate speech inputted via a mobile phone with a standard pattern for recognition and send a recognition result, and the plurality of speeches The recognition standard pattern in at least one of the collating units includes a recognition vocabulary standard pattern that models the speech to be recognized, and a background human voice standard pattern that is added before and after the recognition vocabulary standard pattern. The background human voice standard pattern is learned on the basis of low-level input speech that is input via a mobile phone transmission system and receives large distortion compared to normal input speech to the mobile phone. Background human voice standard patterns or background human voice garbage standard patterns obtained by learning based on the human voice behind the mobile phone voice, and the determination unit includes the plurality of voice verifications. Add a predetermined value to each of the collation scores sent out, compare the obtained values with each other, and send the recognition vocabulary at the voice collation unit that sent the highest value as the recognition result It is characterized by doing.

  Here, in the recognition standard pattern, a boundary standard pattern representing the boundary between the recognition vocabulary standard pattern and the background human voice standard pattern, for example, a standard pattern corresponding to silence can be sandwiched.

  According to the present invention, a recognition standard pattern consisting of a combination of a recognition vocabulary standard pattern and a background human voice standard pattern is used, input as a background human voice standard pattern via a mobile phone transmission system, and sent to the mobile phone. Learning based on background human voice standard patterns obtained by learning based on low-level input voice that receives large distortion compared to normal input voice, or on the background voice contained in mobile phone voice Since the obtained background human voice garbage standard pattern is used, even if voice other than the input purpose to the voice recognition device such as the human voice behind you or monologue mixed in the voice input via the mobile phone is mixed with high accuracy Can recognize voice.

  In addition, by using the background human voice standard pattern learned based on the voice that has undergone large distortion compared to the normal input voice, the recognition vocabulary or part of it is collated with the background human voice standard pattern before and after, and misrecognition Can be suppressed.

  In addition, by inserting a standard pattern that represents the boundary between the recognized vocabulary standard pattern and the background human voice standard pattern, the recognition vocabulary part is mistakenly recognized as the background human voice, and speech is recognized with higher accuracy. it can.

  The present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing an embodiment of a speech recognition apparatus according to the present invention, where the same or equivalent parts as in FIG.

  In FIG. 1, a voice input via a mobile phone is collated with a recognition standard pattern 2 in a voice collation unit 1. The recognition standard pattern 2 is stored in advance for collation, and includes a combination of the recognition vocabulary standard pattern 2-1 and the background human voice standard pattern group 2-3 or the background human voice garbage standard pattern 2-4. . Since the background human voice standard pattern group 2-3 and the background human voice garbage standard pattern 2-4 are both standard patterns for the background human voice, they may be referred to as background human voice standard patterns below.

  The background human voice standard pattern group 2-3 is learned based on low-level input speech that is input through a transmission system of a mobile phone and receives a large distortion compared to normal input speech to the mobile phone. Created. Here, the low-level input voice that receives a large distortion compared to the normal input voice to the mobile phone is sufficient compared to the voice when the person who is going to make a call normally speaks to the mobile phone. In particular, the sound is low in level and is subject to large distortion (such as encoding distortion) in a mobile phone transmission system.

  For example, in the speech coding system used in mobile phones, the document “ARIB Standard: STD-T64-C.S0014-0 Enhanced Variable Rate Codec, Speech Service Option 3 for Wideband Spread Spectrum Digital Systems (page 4-23, 4.3. As described in section 1.2), the code length for encoding speech is determined by an equation that takes into account the speech input level. The shorter the code length, the greater the distortion added to the speech due to encoding.For example, according to the judgment formula used when selecting the code length, the speech of the level encoded with the short code length is used for background human voice standard pattern learning. To do.

  The background human voice garbage standard pattern is created by learning based on the human voice behind the mobile phone voice. FIG. 2 shows the distribution of the input level of the voice input to the mobile phone and the background human voice. As can be seen from the difference in the input level distribution between the voice input to the mobile phone and the background human voice, the background human voice has a lower input level than the voice uttered to the mobile phone. The transmission system is subjected to large distortion (such as encoding distortion).

  3 and 4 show examples of the standard pattern for recognition 2 used in the present invention. FIG. 3 shows a recognition vocabulary standard pattern 2-1 and a background human voice standard pattern (background human voice standard pattern group) 2- 4 is an example of a recognition standard pattern 2 composed of a combination of 3 and FIG. 4 shows a recognition composed of a combination of a recognition vocabulary standard pattern 2-1 and a background human voice standard pattern (background human voice garbage standard pattern) 2-4. This is an example of the standard pattern 2 for use. The voice collation unit 1 collates the input voice with the standard pattern for recognition, and sends the recognition vocabulary having the highest collation score (similarity) as a recognition result.

  When the recognition standard pattern 2 of FIG. 3 is used, the input speech whose both ends are separated by a silent section is a background human voice standard pattern 2-3, a recognition vocabulary standard pattern (recognition vocabulary 1 to N) 2-1, a background. It is collated with a recognition standard pattern 2 consisting of a human voice standard pattern 2-3.

  When the recognition standard pattern 2 in FIG. 4 is used, the speech whose both ends are separated by a silent section is a background human voice standard pattern (background human voice garbage standard pattern) 2-4, a recognition vocabulary standard pattern ( It is collated with a recognition standard pattern consisting of recognition vocabulary 1 to N) 2-1, background human voice standard pattern (background human voice garbage standard pattern) 2-4.

  Usually, the background human voice has a lower input level than the voice uttered to the mobile phone. For this reason, the background human voice is subjected to different coding distortion from the voice uttered toward the mobile phone in the voice code processing of the mobile phone. By using the recognition standard pattern including the background human voice standard pattern group 2-3 or the background human voice garbage standard pattern 2-4 created in consideration of the coding distortion that the background human voice receives, the background human voice is obtained. Even if is mixed, the voice uttered toward the mobile phone can be recognized well.

  FIG. 5 is a block diagram showing another embodiment of the speech recognition apparatus according to the present invention. In the present embodiment, a plurality of voice collation units 1 (1-1, 1-2,..., 1-n) are provided, and the plurality of voice collation units 1 (1-1, 1-2,. , 1-n), input speech is input in parallel. Each voice collation unit 1 collates the input voice with each recognition standard pattern 2 (2-1, 2-2,..., 2-n), and compares the collation results 1, 2,. It is sent to the determination unit 3.

  Here, at least one of the speech collating units 1 (1-1, 1-2,..., 1-n) is similar to the embodiment of FIG. A standard pattern for recognition consisting of a pattern (background human voice standard pattern group or background human voice garbage standard pattern) is used, but the other voice collating unit is a standard pattern for recognition in a conventional voice recognition device, for example, FIG. Alternatively, the recognition standard pattern shown in FIG. 10 may be used.

  Each speech collation unit 1 (1-1, 1-2,..., 1-n) recognizes a recognition vocabulary having the highest similarity as a collation result between the input speech and the standard pattern for recognition and a collation score Pi (i = 1-n).

  As shown in the following equation, the determination unit 3 preliminarily applies the verification score Pi (i = 1 to n) transmitted from each speech verification unit 1 (1-1, 1-2,..., 1-n). The set value Wi is added, the values Si obtained as a result are compared with each other, and the recognition vocabulary having the highest value is transmitted as the recognition result m.

Here, the value Wi is a weight for each collation score Pi (i = 1 to n). For example, the value Wi is used in an environment where there is a lot of voices other than the input purpose to the voice recognition device, such as a human voice and monologue. When the apparatus is predicted to be used, the value Wi for the collation result of the speech collation unit using the recognition standard pattern as shown in FIG. 3 and FIG. The value Wi for the collation result of the speech collation unit using the recognition standard pattern such as 10 is reduced.

Si = logPi + Wi

  According to the embodiment of FIG. 5, by using various different standard patterns for recognition and weighting each matching result, an environment where there is a lot of voice mixing other than the purpose of input to the voice recognition device, an environment where there is little, etc. The speech recognition device can be adapted so that speech recognition is performed well under the environment in which it is used.

  In the above embodiment, the recognition standard pattern in which the background human voice standard pattern and the recognition vocabulary standard pattern are continuous is used, but the boundary representing the boundary between the recognition vocabulary standard pattern and the background human voice standard pattern is used. By using a standard pattern for recognition with a standard pattern in between, the phenomenon that the recognized vocabulary part is erroneously recognized as the background human voice can be further reduced.

  FIG. 6 shows an example of a recognition standard pattern 2 in which a boundary standard pattern 2-5 that represents the boundary between the background human voice standard pattern 2-3 and the recognition vocabulary standard pattern 2-1 is sandwiched. In the figure, GB indicates a background human voice garbage standard pattern. The input speech with both ends separated by a silent section is a background human voice standard pattern 2-3, a boundary standard pattern 2-5, a recognition vocabulary standard pattern 2-1, a boundary standard pattern 2-5, and a background human voice standard pattern 2- 3 is collated with a recognition standard pattern consisting of three. As a result of this collation, the recognition vocabulary with the highest collation score is sent as the recognition result.

  FIG. 7 shows an example in which a standard pattern corresponding to silence is used as a boundary standard pattern 2-5 representing the boundary between the background human voice standard pattern 2-3 and the recognized vocabulary standard pattern 2-1, and both ends are silent sections. The divided input speech includes a background human voice standard pattern 2-3, a silence pattern (boundary standard pattern) 2-5, a recognition vocabulary standard pattern 2-1, a silence pattern (boundary standard pattern) 2-5, and a background human voice standard. It is collated with a recognition standard pattern consisting of pattern 2-3. As a result of this collation, the recognition vocabulary with the highest collation score is sent as the recognition result.

  In the above, the embodiment in which the recognition standard pattern of FIG. 3 or FIG. 4 is used alone has been described. However, the background human voice configured by combining the background human voice standard pattern group and the background human voice garbage standard pattern in parallel. Standard patterns can also be used before and after the recognized vocabulary standard pattern 2-1. 6 and 7 are examples using the background human voice garbage standard pattern as the background human voice standard pattern. As the background human voice standard pattern, the background human voice standard pattern group or the background human voice standard pattern group is used. A combination of background human voice garbage standard patterns can also be used.

  INDUSTRIAL APPLICABILITY The present invention is useful for a voice conversation service for a call center, a voice conversation system for a car navigation system, and the like in addition to a voice portal service for a mobile phone.

It is a block diagram which shows embodiment of the speech recognition apparatus which concerns on this invention. It is a characteristic view which shows distribution of the input level of the input voice and background human voice to a mobile phone. It is a figure which shows the example of the standard pattern for recognition used by this invention. It is a figure which shows the other example of the standard pattern for recognition used by this invention. It is a block diagram which shows other embodiment of the speech recognition apparatus which concerns on this invention. It is a figure which shows the further another example of the standard pattern for recognition used by this invention. It is a figure which shows the further another example of the standard pattern for recognition used by this invention. It is a block diagram which shows the conventional speech recognition apparatus. It is a figure which shows the standard pattern for recognition used with the conventional speech recognition apparatus. It is a figure which shows the other standard pattern for recognition currently used with the conventional speech recognition apparatus.

Explanation of symbols

1, 1-1 to 1-n ... voice collation unit, 2, 2-1 to 2-n ... recognition standard pattern, 2-1 ... recognition vocabulary standard pattern, 2-2. Voice standard pattern, 2-3 ... background human voice standard pattern, 2-4 ... background human voice garbage standard pattern, 2-5 ... boundary standard (silence) pattern, 3 ... judgment unit

Claims (4)

In a speech recognition apparatus having a speech collation unit that collates speech input via a mobile phone with a recognition standard pattern and sends a recognition result,
The recognition standard pattern is a recognition standard pattern including a recognition vocabulary standard pattern that models speech to be recognized and a background human voice standard pattern added before and after the recognition vocabulary standard pattern,
The background human voice standard pattern is obtained through learning based on low-level input speech that is input via a mobile phone transmission system and receives large distortion compared to normal input speech to the mobile phone. A voice recognition device characterized by being a voice standard pattern group or a background human voice garbage standard pattern obtained by learning based on a human voice behind a mobile phone voice. In a speech recognition apparatus having a plurality of speech collation units and a determination unit that collate speech input via a mobile phone with a standard pattern for recognition and send a recognition result,
The recognition standard pattern in at least one of the plurality of speech collating units includes a recognition vocabulary standard pattern that models speech to be recognized, and a background human voice standard pattern that is added before and after the recognition vocabulary standard pattern. A standard pattern for recognition including
The background human voice standard pattern is input through a mobile phone transmission system, and a background person obtained by learning based on low-level input voice that receives large distortion compared to normal input voice to a mobile phone. Voice standard pattern group, or background human voice garbage standard pattern obtained by learning based on the human voice behind mobile phone voice,
The determination unit adds a predetermined value to each of the collation scores sent from the plurality of voice collation units, compares the obtained values with each other, and sends the highest value. A speech recognition apparatus characterized in that a recognition vocabulary in a collation unit is transmitted as a recognition result. The boundary standard pattern representing the boundary between the recognition vocabulary standard pattern and the background human voice standard pattern is sandwiched between the recognition standard pattern and the background human voice standard pattern. Voice recognition device. The speech recognition apparatus according to claim 3, wherein the boundary standard pattern is a standard pattern corresponding to silence.

Images