JP3109978B2 - Voice section detection device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice section detecting apparatus used in a voice coding apparatus for coding and transmitting only a voice section.

[0002]

2. Description of the Related Art Conventionally, there has been known a voice section detection device which divides an input voice into frames of a predetermined length and detects whether the frame is a voice section. In this specification, input speech is classified as follows according to its properties. The voice section has some information of the input voice to be transmitted as voice, and indicates a portion that needs to be coded by the voice coding device. The non-speech section is a part obtained by removing the speech section from the input speech, and is a part that does not need to be encoded. A sound segment is
It shows a portion of the input voice where the voice signal is above a predetermined threshold. It does not matter whether the audio signal in the sound section contains information to be transmitted. That is,
It may be just noise. The silent section indicates a portion of the input voice excluding the sound section.

[0003] The most basic method for detecting a voice section is
In this method, the average voice power for each frame is compared with a predetermined threshold value, a frame whose average voice power is larger than the threshold value is determined as a voiced section, and the voiced section is regarded as a voice section as it is. Under the condition that there is no background noise or the level is very low, the voice section almost coincides with the sound section, so that accurate voice section detection is possible. On the other hand, under conditions where the background noise level is high, the number of sections determined to be sound sections increases, and voice sections cannot be detected correctly.
Alternatively, when the threshold for determining a sound section is increased and the number of sections determined to be a sound section is reduced, there is a problem that a sound section is determined to be a silent section, and a sound break occurs. Further, a method using a feature amount of the input speech such as a prediction error based on the stability of the spectral parameters of the input speech or a linear prediction analysis, the number of zero crossings of the input speech as a threshold, or an average speech in a sound section and a silent section By a method of determining a sound section based on power and varying a threshold value, for background noise such as white noise, for example, background noise exists at a high level of about 20 dB in SN ratio. However, a device that can correctly detect a voice section has been developed.

[0004]

However, in the above-mentioned conventional apparatus, general environmental noises, such as factory noises and street noises, both have a large variation in average voice power and voice feature amount, and the determination value is low. Since a problem such as frequent rise and fall around the threshold occurs, there is a problem that it is difficult to correctly detect a voice section.

[0005] The present invention solves such a conventional problem. Even under the condition that a highly fluctuating environmental noise exists at a high SN ratio, for example, about 20 dB, the voice section in the input voice is recognized. It is an object of the present invention to provide a voice section detection device capable of correctly detecting.

[0006]

In order to achieve the above object, the present invention provides a periodicity detecting means for analyzing an input voice to detect a periodicity, and a sound section based on power information of the input voice. Voiced section detection means for detecting, and voice section determination means for detecting a voice section in accordance with rules for determining a predetermined voice section and a non-voice section from current and past detection results of these two detection means , The sound section detection means
Is the average for calculating the average audio power for each frame of the input audio.
Average voice power calculation means, and average voice power of the previous frame and current
Short-time power ratio to calculate the ratio with the average audio power of the frame
The calculation means and the average audio power for each frame
Average voice power and current frame averaged over time
A long-term power ratio calculator that calculates the ratio to the average audio power of the system
And a voice section determining means for determining a state of the input voice.
Has a status counter indicating the status according to a predetermined rule.
Status counter updating means for updating the status counter;
The judgment map containing the rules and the value of the status counter
The voice section is determined by comparing with a predetermined threshold.
And comparison determining means .

In the present invention, the status counter updating means may include:
Periodicity judgment value from periodicity judgment means and short-time power ratio calculation
Short-term power ratio from the output means and long-term power ratio calculation means
Based on the long-term power ratio and the past judgment result.
A state counter that holds a value that estimates the state of the force voice
The rule to determine the increase / decrease value of the state counter based on the value of
Update the value of the status counter with reference to the judgment map
It is characterized by that.

[0008]

According to the present invention, a voice section can be correctly detected by the above configuration even under a condition in which the input noise has a highly fluctuating environmental noise, for example, at a high SN ratio of about 20 dB.

[0009]

An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing a configuration of a voice section detection device according to one embodiment of the present invention. In FIG. 1, 10
1 is a periodicity detecting means for detecting the periodicity of the input voice, 10
2 is an average voice power calculating means for calculating an average voice power of each frame of the input voice, 103 is a short-time power ratio calculating means for calculating a ratio between the average voice power of the previous frame and the average voice power of the current frame, 104 is A long-time power ratio calculating means for calculating a long-time average voice power by averaging the average voice power for each frame over m frames and calculating a ratio between the long-time average voice power and the average voice power of the current frame; 105
Is a state counter updating means having a state counter indicating the state of the input voice and updating the state counter in accordance with a predetermined rule, 106 is a judgment map containing predetermined rules, 107 is a state counter value and a predetermined value This is a comparison determination unit that determines a voice section by comparing the threshold value with a threshold value. 108 is an input voice, 109 is a periodicity judgment value,
110 average speech power of the current frame, 111 short time power ratio, 112 long time power ratio, 113 state counter value 114 is a speech segment determination result of the determination. And
Average voice power calculating means 102 and short-time power ratio calculating means 1
03 and the long-time power ratio calculating means 104 constitute a sound section detecting means 115, and the state counter updating means 105, the determination map 106 and the comparing means 107 constitute the voice section determining means 1.
Constituting No. 16.

Next, the operation of the above embodiment will be described. In FIG. 1, an input voice 108 is input to a periodicity detecting unit 101 and an average voice power calculating unit 102. The periodicity detecting unit 101 analyzes the input voice and outputs a periodicity determination value 109. The periodicity determination value 109 may be binary information indicating the presence or absence of periodicity, or may be continuous value information indicating the degree of periodicity. Average voice power calculation means 10
2 calculates and outputs the average audio power 110 of the current frame. The short-time power ratio calculation means 103 calculates the average audio power 1
A ratio between 10 and the held average audio power of the previous frame is calculated and output as a short-time power ratio 111. Thereafter, the stored average audio power of the previous frame is updated by the average audio power of the current frame. Similarly, the long-time power ratio calculating means 104 calculates the ratio between the average voice power 110 and the long-term average power obtained by further averaging the held average voice power of the past m frames, and calculates the ratio as the long-time power ratio 112. Output. Thereafter, the stored average audio power of the past m frames is updated by the average audio power 110 of the current frame. The state counter updating means 105 calculates the periodicity judgment value 1
09, the short-time power ratio 111, the long-time power ratio 112, and the increase / decrease value of the state counter based on the value of the state counter holding the value for estimating the current state of the input voice based on the past determination result. Map 106 containing rules for determining
And update the status counter. Comparison judgment means 10
7 compares the updated state counter value 113 with a predetermined threshold value to determine whether the current frame is a voice section or a non-voice section.

Next, the periodicity judgment value 109, which is the characteristic amount of the voice used in the above embodiment, and the average voice power 11
The principle of voice section detection using 0, short power ratio 111, and long power ratio 112 will be described below.

FIG. 2 shows an SN ratio of 20 dB in the above embodiment.
Shows the change of each voice feature amount when a voice to which street noise is added is input, and 201 denotes the average voice power 1 for each frame.
Reference numerals 10 and 202 denote a short-time power ratio 111, 203 denotes a long-time power ratio 112, and 204 denotes a change in the periodicity determination value 109, respectively. 201, 202, and 203 set the value in decibels [d
B], in which reference numeral 204 denotes a section determined to have periodicity (stationary section) by a peak, and a section determined to have no periodicity (unsteady section) by a valley. The frame length of the input voice was 20 ms, and the number m of frames for calculating the long-term average voice power was 5. Both the short-time power ratio 111 and the long-time power ratio 112 greatly change at a portion where the average audio power 110 changes significantly, that is, at the rising and falling portions of the audio. Therefore, it can be said that the short-time power ratio 111 and the long-time power ratio 112 are characteristic amounts suitable for detecting the rise and fall of the voice.
The difference between the short-time power ratio 111 and the long-time power ratio 112 is that the short-time power ratio 111 tends to change too rapidly instead of following the average voice power even if the average voice power repeatedly changes rapidly in a short time. On the other hand, the long-time power ratio 112 has a tendency that the change is stable but cannot follow a rapid change. By using both characteristics in combination, not only the condition without background noise,
For example, even under a condition in which a highly fluctuating noise having an SN ratio of about 20 dB is added, a more accurate rising and falling portion of a voice can be detected. Further, the periodicity judgment value 109 is a feature amount effective for detecting a stationary part in the voice.

The voice section detection apparatus according to the present embodiment detects the rising and falling portions of the voice based on a combination of the short-time power ratio and the long-time power ratio, and detects the stationary voice section by detecting the periodicity. The principle of operation is to detect a voice section by comprehensive judgment of the above, and an example of voice section detection using this principle of operation will be described below. Here, the state counter updating means 1 in FIG.
The range of values taken by the state counter held by 05 is 0 to 1.
When the value of the state counter is in the range of 0 to 5, it is determined to be a voice section.

FIG. 3 shows an example of the judgment map 106. A plane having the short-time power ratio on the vertical axis and the long-time power ratio on the horizontal axis is represented by nine map areas 301 from area 1 to area 9. The increase / decrease value 302 of the state counter is assigned to each area. When the held state counter value before update indicates a non-voice section, the state counter updating unit 105 determines to which region on the determination map 106 the received short-time power ratio 111 and long-time power ratio 112 belong. , And updates the status counter with the increase / decrease value of the corresponding status counter. If the updated state counter value is in the range of 0 to 5, the current frame is determined to be a voice section. That is, the rising of the voice section is detected. Further, when the state counter value before the update indicates a voice section, a steady section is detected based on the periodicity determination value 109. If it is determined that the section is a steady section, the state counter is cleared to 0, and the state is determined to be an unsteady section. If it is, the state counter is updated and the state counter is updated in the same manner as when the state counter value indicates the non-voice section. If the updated state counter value is in the range of 6 to 18, the current frame is determined to be a non-voice section. That is, the falling of the voice section is detected.

FIG. 4 is a diagram showing a result of detection of a voice section by the voice section detection apparatus of the present embodiment, where 401 indicates a change in average voice power of voice to which street noise having an SN ratio of 20 dB is added, and 402 indicates A result of detecting a voice section from a voice to which street noise with an S / N ratio of 20 dB is added, 403 indicates a change in average voice power of voice without background noise, and 404 indicates a voice section from a voice without background noise. The detection result is shown. In 402 and 404, a peak indicates a voice section and a valley indicates a non-voice section.
As shown in the figure, although there are some differences in the detection intervals, it can be seen that the voice intervals are detected stably regardless of the presence or absence of background noise.

Note that the speech frame in the above embodiment need not always match the speech frame of the speech encoding device to be combined, and if the delay is allowed, the speech frame of the speech segment detection device is replaced by the speech frame. By displacing the frame ahead of the frame of the encoding device and pre-reading the input voice, it is possible to detect the power change of the input voice more quickly and detect the voice section more accurately.

Also, in the above embodiment, since the average voice power is not directly used as the feature value used for detecting the voice section, the voice section can be detected almost without being affected by the input level of the input voice. There are features. Therefore, the voice section can be detected even under the condition that the input level is very low. However, depending on the application, a section in which the average voice power is equal to or less than a predetermined threshold value needs to be determined as a non-voice section. In order to use such a purpose, a decision rule that directly depends on the average sound power is added to the decision map,
If the average voice power is equal to or less than a predetermined threshold, it may be determined to be a non-voice section.

It should be noted that the voice section detection device of the present invention is
Speech coding apparatus having means for detecting speech periodicity generally called pitch extraction, such as P (Code Excited Linear Prediction coding) and MBE (Multi Band Excitation). When used in combination with the above, the periodicity determination value can use the pitch information obtained in the speech encoding process, eliminating the need for independent periodicity detection means, and significantly reducing the amount of computation required for speech section detection. I do. Therefore, the speech section detection device of the present invention is very suitable for being combined with a speech encoding device having pitch extraction means.

[0019]

As described above, according to the present invention, in a voice section detection apparatus for dividing an input voice into frames of a predetermined length and detecting whether or not the frame is a voice section, Periodicity detecting means for analyzing the input voice to detect periodicity; voiced section detecting means for detecting a voiced section based on the power information of the input voice; current and past detection results of these two detecting means Therefore, since the voice section is provided with voice section determination means for detecting a voice section in accordance with a rule for determining a predetermined voice section and a non-voice section, the voice section can be recognized even when the background noise level is high and the fluctuation is severe. It can be detected accurately.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of a voice section detection device according to an embodiment of the present invention.

FIG. 2 is a characteristic diagram showing a feature amount of a voice according to an embodiment of the present invention.

FIG. 3 is a schematic diagram showing a determination map according to one embodiment of the present invention.

FIG. 4 is a characteristic diagram showing a detection result of a voice section in one embodiment of the present invention.

[Explanation of symbols]

Reference Signs List 101 periodicity detecting means 102 average voice power calculating means 103 short-time power ratio calculating means 104 long-time power ratio calculating means 105 state counter updating means 106 determination map 107 comparison determining means 108 input voice 109 periodicity determining value 110 average of current frame Voice power 111 Short-time power ratio 112 Long-time power ratio 113 State counter value 114 Decision result 115 Voiced section detection means 116 Voice section determination means 201 Average voice power 202 Short-time power ratio 203 Long-time power ratio 204 Periodicity determination value 301 Map area on decision map 302 Increase / decrease value of state counter 401 Average voice power of voice with background noise added 402 Detection result of voice section from voice with background noise added 403 Average voice power of voice without background noise 404 Background Voice without noise The voice section detection

────────────────────────────────────────────────── ─── Continuation of front page (56) References JP-A-2-238493 (JP, A) JP-A-60-200300 (JP, A) JP-A 1-159697 (JP, A) JP-A-60-200 57396 (JP, A) JP-A-60-499 (JP, A) JP-A-63-235999 (JP, A) JP-A-63-163495 (JP, A) JP-A-1-255897 (JP, A) Patent 2648779 (JP, B2) Japanese Patent Publication No. 1-2519 (JP, B2) Japanese Patent Publication No. 4-64074 (JP, B2) Proceedings of the Acoustical Society of Japan Spring Meeting, 1989, I-7, 3-7- 15, Yumi Takizawa, "Development of Noise-Resistant Speech Recognition System (1)-Section Detection Method-", p. 117-118 (Issued March 14, 1989) (58) Field surveyed (Int. Cl. ⁷ , DB name) G10L 11/00-21/06

Claims (2)

(57) [Claims] 1. A speech section detection device for dividing an input speech into frames of a predetermined length and detecting whether the frame is a speech section, detects the periodicity by analyzing the input speech. Periodicity detecting means, and a sound section detecting means for detecting a sound section based on power information of the input voice,
A voice section determining means for detecting a voice section in accordance with a rule for determining a predetermined voice section and a non-voice section from current and past detection results of these two detecting means ;
The sound interval detecting means calculates an average of the input voice for each frame.
An average voice power calculating means for calculating voice power;
The ratio of the average audio power of the current frame to the average audio power of the current frame.
Short-time power ratio calculation means to calculate and average sound for each frame
Long-time flatness that averages voice power over m frames
Calculate the ratio between the average audio power of the current frame and the average audio power
Long time power ratio calculating means,
The stage has a state counter indicating the state of the input sound, and
A status counter that updates the status counter according to a set rule
Counter updating means, a judgment map storing the rules,
The value of the state counter is compared with a predetermined threshold value.
A comparison determining means for determining a voice section.
Characteristic voice section detection device. 2. A state counter updating means, comprising: a periodicity judgment value from a periodicity judging means, a short-time power ratio from a short-time power ratio calculating means, and a long-term power ratio from a long-time power ratio calculating means. Based on the value of the state counter that holds the value for estimating the current state of the input voice based on the past determination results, refer to the determination map containing rules for determining the increase or decrease of the state counter. updating the value of the counter speech segment detection device of claim 1, wherein.