JPH03233600A - Voice segmenting method and voice recognition device - Google Patents

Abstract

PURPOSE:To accurately detect the time area of a voice at all times by setting a threshold value with an ambient noise level which is different from an acoustic signal and detected by an acoustic input means, segmenting an acoustic signal area according to the threshold value, and extracting the acoustic signal area as a voice area. CONSTITUTION:A 1st segmentation control part 3 detects a 1st acoustic signal area where it is considered that there is a voice without fail from an acoustic signal where a voice is present by using a 1st threshold value varying with ambient noise. Further, a 2nd segmentation control part 7 detects a 2nd acoustic signal area obtained by extending time length about the 1st acoustic si cal area by using a 2nd threshold value which is less than the 1st threshold value and the 2nd acoustic signal area can be regarded as a voice area. Therefore, a feature parameter which represents the features of a voice properly is extracted from the acoustic signal extending to the voice area. Consequently, the voice recognition rate can be improved.

Description

DETAILED DESCRIPTION OF THE INVENTION (A) Field of Industrial Application The present invention relates to a speech recognition device, and more particularly to a method for cutting out speech for detecting the time domain of speech input to the speech recognition device.

(b) In conventional technology voice recognition devices, ambient noise is always input into the microphone for inputting voice in addition to voice.
An important issue is to accurately detect the time domain of speech included in this ambient noise.

For example, even in an office where background music (BGM) is playing, it may be necessary to use voice recognition to perform input into a word processor, for example, and in this case, BGM is used. Since M is mixed with the speaker's voice and input into the microphone for voice recognition,
Speech recognition is impossible unless it is possible to accurately detect from which time position to which time position of this input acoustic signal the speech region is. This is the same even when attempting to operate automobile electrical equipment by voice recognition in a car while music or songs are being played on an in-vehicle audio device such as a car stereo.

Therefore, in conventional devices, the time period when the level of the signal input to the microphone is detected and the level exceeds a specific threshold determined in advance based on the environment and conditions in which the sound is generated is regarded as the time domain of the sound. An audio extraction method was used to extract the audio.

However, with such conventional audio extraction methods, the playback level of GBM (ambient noise) and songs is not constant, so there is an inconvenience that accurate audio extraction cannot be achieved just by using a conventional fixed threshold. .

(c) Problems to be Solved by the Invention The present invention has been made in view of the above-mentioned conventional disadvantages, and it is an object of the present invention to accurately detect the time domain of speech even in an environment with ambient noise whose level fluctuates. The present invention aims to provide a voice extraction method that enables the following, and furthermore, to realize a voice recognition device by employing this voice extraction method.

(d) Means for Solving the Problems The audio extraction method of the present invention is a method of extracting an audio region by detecting the presence of audio in a time region in which the level of an acoustic signal in which audio exists exceeds a specific threshold. The threshold value is set based on the ambient noise level detected by a sound input means different from the sound signal, and the sound signal region is cut out using the threshold value, and the sound signal region is extracted as a voice region.

The voice recognition device of the present invention also includes a microphone for inputting voice;
a voice analysis unit that analyzes an acoustic signal obtained from the microphone to extract a voice characteristic parameter time series; a voice pattern creation unit that creates a voice pattern based on the characteristic parameter time series obtained from the voice analysis unit; a standard voice pattern memory that stores voice patterns of a plurality of standard voices as standard voice patterns; an identification unit that patterns-matches each voice pattern in the memory with the voice pattern to identify the voice pattern; an audio input terminal for inputting audio, a first audio extraction threshold setting unit that sets a first audio extraction threshold based on a noise sound level from an audio device connected to the input terminal that is a source of ambient noise; a first cutout control unit that detects a first acoustic signal region from the acoustic signal obtained from the microphone according to a set first cutout threshold; an acoustic signal in which the first acoustic signal region detected by the control unit is located at the center; a second cutout threshold setting unit that further sets a second threshold lower than the first threshold based on the ambient noise level; a second cut-out control section that detects a second acoustic signal region including the region, and the second cut-out control section regards the second sound signal region detected by the second cut-out control section as a sound region, and from the sound analysis section. Among the obtained feature parameter time series, the speech pattern creation section creates a speech pattern based on the feature parameter time series existing in the speech region.

(e) Effect: According to the audio extraction method of the present invention, the threshold for detecting the time domain of audio at that level from an acoustic signal where audio is present can be dynamically set according to the ambient noise level, so the ambient noise fluctuates. It is possible to effectively detect audio areas even under different environments.

According to the speech recognition device of the present invention, the first cut-out control unit uses the first threshold value that varies depending on the ambient noise to determine the first cut-out control unit that determines whether speech is present from an acoustic signal in which speech is present.
a second acoustic signal area in which the second cutout control unit further extends the time length around the first acoustic signal area using a second threshold value smaller than the first threshold value; By detecting the second acoustic signal region and regarding the second acoustic signal region as the speech region, characteristic parameters that appropriately represent the characteristics of the speech can be extracted from the acoustic signal spanning the speech region,
By creating a speech pattern based on this feature parameter, it is possible to improve the speech recognition rate.

(f) Embodiment FIG. 1 shows the origins of an embodiment of the speech recognition device of the present invention.

In the figure, 1 is a microphone into which voice is input, 2 is a voice analysis unit that analyzes the acoustic signal input from the microphone 1 and extracts a time series of characteristic parameters representing the characteristics of the voice.
For example, frequency analysis provides spectral parameters that preserve acoustic signal level information. Reference numeral 3 denotes a first extraction control unit for extracting a time domain in which speech exists from the time series of feature parameters obtained from the voice analysis unit 2, and controls the first feature parameter and the last feature parameter of the time domain, respectively. A temporary start code and an end code are assigned to output a series of time series of feature parameters (including a sufficient number of time series that precede and follow these coded parameters). Reference numeral 4 denotes a first audio buffer that temporarily stores the feature parameter time series to which a temporary start code and end code have been added from the first extraction control unit 3.

5 is a noise level input terminal different from the above microphone,
This includes a second microphone for inputting ambient noise, an output terminal of an audio reproduction device that is a source of ambient noise, or a terminal for indicating the reproduction level of this audio reproduction device (for example, a level meter consisting of an LED bar display). The signal line is connected. Reference numeral 6 denotes a first threshold value necessary for time-domain extraction of the audio with respect to the feature parameter time series in the first extraction control unit 3 and the acoustic signal from the microphone 1 and the noise level input terminal 3.
This is a first threshold value setting unit that sets the threshold value with reference to the ambient noise level from .

Reference numeral 7 denotes a second cutout control unit for cutting out again strictly a time region in which speech exists from the time series of feature parameters to which temporary start codes and end codes obtained from the first speech buffer 4 are added. , a true start code is given to the feature parameter at a position temporally earlier than the tentative first feature parameter in the time domain (corresponding to the start position of the true speech domain), and the time domain is after position (end position of true audio area)
The end code of the soldier is given to the feature parameters of the model, and a time series of these feature parameters is output. 8 is the second
This is a second audio buffer that temporarily stores the feature parameter time series to which the true start code and end code have been added from the extraction control unit 7. Reference numeral 9 denotes a second threshold setting unit that sets a second threshold necessary for the true time domain extraction of the audio for the feature parameter time series in the second extraction control unit 7 to be smaller than the first threshold; It is empirically set to a value that allows the true time domain to be extracted appropriately, for example, approximately 80% of the first threshold value, although it varies somewhat depending on the environment.

10 generates an input speech pattern based on the feature parameter time series of the true speech region belonging to between these codes from the feature parameter time series to which the true start code and end code are added stored in the second buffer 8; This is an audio pattern creation unit that creates audio patterns that are normalized feature patterns in a specific time series. Reference numeral 11 denotes a noise level buffer that stores the noise level obtained from the noise level input terminal 5 over the true speech region obtained from the second extraction control section 7; 12 indicates a time average of the noise level of the noise level buffer 11; This is a level comparison section that compares the value with a predetermined level set empirically, and prohibits the speech pattern creation process in the speech pattern creation section 10 when the average noise level of the noise level buffer 11 is higher than the predetermined level. do.

Reference numeral 13 indicates a standard speech pattern memory in which the speech patterns of a large number of standard speeches are stored in advance as standard speech patterns; and 14 indicates an input speech pattern obtained from the speech pattern creation section 10, which is stored in each standard speech in the standard speech pattern memory 13. This is an identification unit that performs pattern matching to detect a standard speech pattern that has the smallest error between patterns and that is less than the recognition threshold, which is the allowable limit of this error, and the recognition result that corresponds to the detected standard speech pattern. Output a signal.

Reference numeral 15 denotes a recognition threshold setting unit that variably sets the recognition threshold in the identification unit 14 according to the average noise level of the noise level buffer 11, and when the average noise level is high, the recognition threshold becomes large.

FIG. 2 is a signal waveform diagram showing the speech extraction operation in the speech recognition device of the present invention, and the operation will be described in detail based on the diagram.

First, we will explain how to set the audio time domain extraction threshold.

The first cutout threshold setting unit 6 takes in the noise level from the noise level input terminal 5 that changes stepwise as shown by N in FIG. A first threshold value for extraction is determined. In this case, the noise level input terminal 5 has
A signal line for a level meter consisting of an LED bar display is connected.

That is, the cutout threshold value (denoted as Vtl) setting is a function of the noise level N as shown below.

Vtl = f (N) Specific examples of f(N) are listed below.

1st place alliance f (N) = a X N + b.

Here, a and b each indicate a constant, and in particular, b indicates that in normal steady noise conditions, the noise input from the microphone 1 is not extracted as voice by the first extraction control unit 3. As such, a value greater than the normal steady noise level is given.

An effective method is to set the threshold for clipping based on the input from the microphone of Natsume 2 (about 1 increment).The method for setting the threshold for clipping in this case is shown below.

Here, in the case, condition C is a constant. Furthermore, tl and t2 mean times before the current time, and a is a weight regarding time i. Therefore, the above equation is obtained by adding the above-mentioned constant s to the time average of the acoustic signal level of only the noise from the microphone 1 before voice input.

f (N) shown above assumes that only the known voice is input as noise to microphone 1, but in addition to this, stationary ambient noise can also be input to microphone 1. be. In this case, the above-mentioned threshold setting cannot deal with the problem. Therefore, since ambient noise is constantly input through the microphone 1, this input is input to the first cutout threshold setting unit 6.
while saving data from a certain period of time before the current input (for example, 50
It shows the power of the input from the microphone 1 before (for example, about 50 msec).

10M11 In the fourth function example above, regardless of whether the noise level N is larger or smaller than the constant C, the above equations ■ and equations
The larger value of f(N) can be set as f(N').

By employing the above example of the function of f (N), the first threshold value Vtl, which varies depending on the ambient noise N, is determined, as shown by the solid curve in FIG.

Therefore, the first extraction control section 3 determines the level of the feature parameter time series obtained from the speech analysis section 2 [in this case, as shown by the broken line curve V in FIG. ΣV (=V) ] and the first
A comparison is made with a threshold value Vtl, and a provisional start code is given to the feature parameter at the beginning time Tsl of the continuous time series where ΣV≧Vtl, and a provisional end code is given to the feature parameter at the final time Tel.

In this way, the feature parameter time series to which the temporary start code and end code have been added is stored in the first audio buffer 4. At this time, the buffer 4 sufficiently stores the time series before the feature parameter to which the temporary start code has been assigned, and the time series after the feature parameter to which the temporary end code has been assigned.

Next, audio clipping by the second clipping control section 7 will be explained.

When the noise level from the noise level input terminal 5 is large, the first cutout control section 3 may not be able to accurately cut out the beginning and end of the speech, and therefore the speech region is shorter than the true speech region. This means that it can only be detected. Therefore, the second cutout control section 7 is provided to compensate for this.

That is, the second cutout control section 7 sets a second threshold Vt2 that is smaller than the first threshold Vt1 set by the first cutout threshold setting section 3, and uses this threshold Vt2 to A more appropriate audio region is extracted from the characteristic parameter time series of buffer 4.

Here, the setting of the second threshold value Vt2 will be explained. The first threshold value Vtl set by the first extraction threshold setting section 6 is provided to the second extraction threshold setting section 9 together with time information.

The second cutout threshold setting unit 9 sets the audio level V(Tsl )=Vt at the tentative head time point Tsl determined by the first threshold Vtl set by the first cutout threshold setting unit 6.
A second threshold value Vt2 smaller than the first threshold value V(T61)=Vtl(Tel) at the tentative final time point Tel is determined.

Specifically, the second
The threshold value Vt2 is a function of Vtl (Tsl), and is expressed as follows.

For example, Vt2 = Vtl (Tsl) - d, d is a constant or Vt2 = Vtl (Tsl)7m%m
is a constant.Furthermore, the second
The threshold value Vt2 is a function of Vjl (Te1), and is expressed as follows, as in the case of the true leading time Ts2.

For example, Vt2 = Vtl (Tel) - d, d is a constant, or Vt2 = Vtl (Tel) 7m, m is a constant. Note that the setting of the second threshold Vt2 is the same as the setting of the first threshold Vtl. Similarly, if the minimum value constant C is set, there is no fear that stationary noise will be cut out as speech.

Therefore, using the second threshold value Vt2 set by the second extraction threshold setting unit 9, the second extraction control unit 7 sets V(Ta) before the time Tsl stored in the first audio buffer 4.
2) Detect a time point Ts2 that is considered to be the true beginning point of the voice where = Vt2, and assign a true start code to the feature parameter at this time point. Furthermore, after time Tel, V(T
a2) Detect a time Ts2 that can be regarded as the true final time of the voice at −Vt2, and assign a true end code to the feature parameters at this time.

In this way, the feature parameter time series to which the true start code and end code have been assigned is temporarily stored in the second audio buffer 8, and the feature parameter time series during which the true start code and end code have been assigned is temporarily stored in the second audio buffer 8. is supplied to the audio pattern creation section 10.

When the noise level is extremely high, it may be difficult to accurately detect the voice region even with the above-mentioned voice extraction means, and in this case, safety measures such as not performing voice recognition are required.

Therefore, in the embodiment shown in FIG. 1, the level comparison section 12 is provided to take the above-mentioned safety measures.

That is, the audio region (
Since the noise levels for Ts2 to Te2 in FIG.
(N)*ΣN/(Te2-Ta2) is calculated, and when this value exceeds a certain value, the voice pattern creation section 10 is prohibited from creating a voice pattern.

On the other hand, the speech pattern created by the speech pattern creation section 10 under noise within an acceptable range is pattern-matched by the identification section 14 with a large number of standard patterns stored in advance in the standard pattern memory 13. Among the standard patterns, the most similar standard pattern (that is, the standard pattern with the smallest error margin) is stored as a recognition result in the identification unit 14 along with the degree of similarity (which has a reciprocal relationship with the error margin).

The identification unit 14 ultimately determines whether or not the recognition result stored in the identification unit 14 is valid based on the recognition threshold of the recognition threshold setting unit 15.

Here, a method of setting a recognition threshold in the recognition threshold setting section 15 will be explained. When expressing the degree of similarity by error, the recognized darkness value Dt
It is determined by following the noise average level ave(N) of s2 to Te2), and is determined, for example, as in the following example.

Dt = pXave (N) + q Here, p and q are constants.

That is, the recognition threshold Dt is set large when the surrounding noise is large.

Therefore, the identification unit 14 validates the recognition result if the degree of similarity of the recognition result is greater than the recognition threshold Dt which varies according to the level of ambient noise (if the recognition result is similar). Even if the input pattern is slightly deformed depending on the magnitude of the noise level, this can be absorbed and a recognition result can be derived.

The voice recognition device described above can be used, for example, as a means for operating a car stereo in a car, and in this case, the car stereo itself becomes the object of ambient noise. In addition to inputting directly to the noise level input terminal 5 from the output line of the audio equipment, ambient noise can also be collected from the microphone by using a microphone and an analog/digital converter.

(G) Effects of the Invention According to the audio extraction method of the present invention, the threshold for detecting the time domain of audio at that level from an acoustic signal containing audio can be dynamically set according to the ambient noise level. Effective audio region detection is possible even in a fluctuating audio reproduction environment. Furthermore, according to the speech recognition device that employs the speech extraction method of the present invention, it is possible to detect speech regions more appropriately, and it is expected that the accuracy of speech recognition processing will be improved.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing the configuration of the speech recognition device of the present invention, and FIG. 2 is a signal diagram showing the speech cutting method of the present invention adopted in the device of FIG. DESCRIPTION OF SYMBOLS 1...Microphone, 2...Speech analysis part, 3...First extraction threshold control part, 4...First sound buffer, 5...
- Noise level input terminal, 6... First extraction threshold setting section, 7... Second extraction threshold control section, 8... Second audio buffer, 9... Second extraction threshold setting section, 10.
...Speech pattern creation section, 11...Noise level buffer, 12...Level comparison section, 13...Standard pattern memory, 14...Identification section, 15...Recognition threshold setting section.

Claims (8)

[Claims] (1) In a sound extraction method that detects the presence of sound in a time region in which the level of the sound signal in which the sound exists reaches a specific threshold or higher and cuts out the sound region, detection is performed using a sound input means different from the above-mentioned sound signal. The above threshold is set according to the ambient noise level, and the acoustic signal region is cut out using the threshold,
A sound extraction method for extracting the acoustic signal region as a sound region. (2) When the ambient noise level is smaller than a predetermined value, the threshold value is set based on the acoustic signal itself.
The audio extraction method described. (3) In a sound extraction method that detects the presence of sound in a time region in which the level of the sound signal in which sound is present reaches a specific threshold or higher and cuts out the sound region, detection is performed using a sound input means different from the above-mentioned sound signal. Due to the ambient noise level
A threshold value is set, a first acoustic signal region is cut out using the threshold value, and then, for the acoustic signal in which this first acoustic signal region exists in the center, furthermore, based on the ambient noise level, A sound extraction method that sets a second threshold at a low level, uses the threshold to cut out a second audio signal region that includes the first audio signal region, and extracts the second audio signal region as an audio region. (4) If the level of ambient noise in the extracted audio region is higher than a value set according to the level of the acoustic signal in the audio region, the extraction of the audio region at this time is invalidated according to claims 1 and 2. , or the audio extraction method described in 3. (5) The audio extraction method according to claim 1, 2, 3, or 4, wherein the ambient noise is an audio signal directly input from an output terminal of an audio device that reproduces sounds such as music. (6) A microphone that inputs voice, a voice analysis unit that analyzes the acoustic signal obtained from the microphone and extracts a time series of voice characteristic parameters, and a voice pattern is extracted based on the time series of characteristic parameters obtained from the voice analysis unit. A voice pattern creation unit to create, a standard voice pattern memory that stores a plurality of standard voice voice patterns as standard voice patterns in advance, and a pattern matching between each voice pattern in the memory and the voice pattern to generate the voice pattern. an identification unit for identifying, an audio input terminal for inputting ambient noise, and a first extraction for setting a first audio extraction threshold based on a noise sound level from an audio device connected to the input terminal and which is a source of ambient noise. a threshold value setting section; a first extraction control section that detects a first acoustic signal region from the acoustic signal obtained from the microphone using a first extraction threshold set by the setting section; a first acoustic signal detected by the control section; a second cutout threshold setting unit that sets a second threshold value lower than the first threshold value based on the ambient noise level for the acoustic signal in which the region is centered; a second cutout set by the setting unit; A second cutout control section detects a second acoustic signal region including the first acoustic signal region using a threshold value, and the second acoustic signal region detected by the second cutout control section is regarded as an audio region. The speech recognition device generates a speech pattern in the speech pattern creation section based on the feature parameter time series present in the speech region among the feature parameter time series obtained from the speech analysis section. (7) The identification unit outputs a recognition result signal that is associated with a standard voice pattern that can perform pattern matching with the input voice pattern at this time with the smallest error, and when the minimum error is smaller than a predetermined recognition error. 7. The speech recognition device according to claim 6, wherein the recognition error is variably set according to ambient noise in the speech region. (8) the audio input terminal is coupled to an output terminal of a vehicle-mounted sound reproduction device that becomes ambient noise for audio input to the microphone, and the identification result signal of the identification section is output to a control circuit of the vehicle-mounted sound reproduction device; The voice recognition device according to claim 6 or 7, wherein an in-vehicle audio reproduction device is operated by voice recognition.