KR101616112B1 - Speaker separation system and method using voice feature vectors - Google Patents

Abstract

The present invention relates to a system and method capable of speaker separation using a feature vector of a speech signal on a single channel including at least one speaker, and a system configuration thereof includes at least one speaker A feature vector extracting unit for extracting a feature vector from a speech signal, a feature vector storing unit for storing the feature vector extracted through the feature vector extracting unit as a speech segment or a non-speech segment, A speaker separator for separating the speaker using the pattern recognition technique based on the accumulated data, and a speaker separator for outputting the speaker separated data output through the speaker separator and the cumulative data of the non- To detect the start and end times of the speaker's per-speaker interval It includes chapters.

Description

[0001] SPEAKER SEPARATION SYSTEM AND METHOD USING VOICE FEATURE VECTOR [0002]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to speech separation, and more particularly to a speaker separation system and method capable of speaker separation using a feature vector of a speech signal on a single channel including at least one speaker .

Generally, in a call center, when an incoming call is made between an agent and a customer, the call is defined as one consultation, and the consultation contents are recorded and stored. The recorded recording file can be used as evidence for customer complaints or as a sample for the evaluation of the counselor.

At this time, it is difficult to record and store counseling line and customer line separately if a professional recording environment is not provided in the call center. Therefore, the voice of the agent and the customer can be recorded and stored together in a single channel.

For example, when the agent and the client line are recorded separately, it is divided into stereo, and it is possible to listen to the agent voice and the customer voice separately in the left and right. When recording on a single channel without dividing the line, both the voice of the agent and the voice of the customer can be heard together.

However, as the service using the recording file of the call center has diversified recently, it is a reality that there is a limit to the utilization of the recording file of the single channel which does not distinguish the speaker such as the agent and the customer.

Related prior arts include Korean Patent Registration No. 10-0642455 entitled " Method of Automatically Evaluating Customer Satisfaction for Corporate Telephone Service, Date of Notification: November 02, 2006 ".

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned problems, and it is an object of the present invention to extract a speech feature vector from a speech signal on a single channel including at least one speaker, The present invention also provides a system and a method for separating the ignition section.

To this end, a speaker separation system using speech feature vectors according to an embodiment of the present invention divides a speech signal including at least one speaker into a predetermined frame unit and extracts a feature vector from the speech signal of the frame interval A feature vector extractor; A feature vector storage unit for classifying and storing feature vectors extracted through the feature vector extracting unit into a speech segment or a non-speech segment; A speaker separator for separating the speaker using the pattern recognition technique based on the accumulated data when the feature vectors of the voice segments stored in the feature vector storage unit are accumulated; And a speech segment detection unit for detecting the start and end times of the speech segment for each speaker by synchronizing the speaker separation data output through the speaker separation unit and the cumulative data of the non-speech segment previously stored in the feature vector storage unit on the time line do.

Meanwhile, a speaker separation method using a voice feature vector according to an embodiment of the present invention is a method in a speaker separation system for separating a speaker from a voice signal including at least one speaker, Extracting a feature vector from the signal; Classifying and storing the feature vector extracted by the speaker separation system into a speech section or a non-speech section; If the feature vectors of the stored speech segments are accumulated, separating the speakers using the pattern recognition technique based on the accumulated data; And a step of the speaker separation system synchronizing the speaker separation data outputted through the speaker separation and the data classified and stored in the non-speech interval on the time line to detect the start and end time of the speaker-specific speaking interval .

According to the present invention, a speaker and a speaker can be separated from each other by using a speech feature vector and pattern recognition from a speech signal including at least two or more speakers.

In addition, due to the above-described speaker separation technique, it is possible to move not only in time or frame unit but also in a speech unit of a specific character when moving between voice data in an audio player and a moving picture player. And the application range is varied.

Furthermore, it can be applied to the dictation function utilizing the STT (Sound To Text) technology, which converts voice commands and voice to text using speech recognition technology, in order to provide a better analysis by applying a range of words or phrases There will be.

1 is a configuration diagram of a speaker separation system according to an embodiment of the present invention;
2 is a detailed configuration diagram of a feature vector extracting unit in a speaker separation system according to an embodiment of the present invention;
3 is a detailed configuration diagram of an error correction filter unit in a speaker separation system according to an embodiment of the present invention;
4 is a flowchart illustrating a speaker separation method according to an embodiment of the present invention.

It is noted that the technical terms used herein are used only to describe specific embodiments and are not intended to limit the invention. It is also to be understood that the technical terms used herein are to be interpreted in a sense generally understood by a person skilled in the art to which the present invention belongs, Should not be construed to mean, or be interpreted in an excessively reduced sense.

Also, the singular forms "as used herein include plural referents unless the context clearly dictates otherwise. In the present application, the term "comprising" or "comprising" or the like should not be construed as necessarily including the various elements or steps described in the specification, Or may be further comprised of additional components or steps.

Further, the suffix "module" and "part" for components used in the present specification are given or mixed in consideration of ease of specification, and do not have their own meaning or role.

Hereinafter, the present embodiment will be described in detail with reference to the accompanying drawings.

1 and 2, FIG. 1 is a configuration diagram of a speaker separation system according to an embodiment of the present invention, and FIG. 2 is a detailed configuration diagram of a feature vector extracting unit according to an embodiment of the present invention.

1, a speaker separation system according to an embodiment of the present invention includes a feature vector extraction unit 10, a feature vector storage unit 20, a speaker separation unit 30, and a speech segment detection unit 40 do. In addition, the apparatus may further include an error correction filter unit 50 for correcting speaker separated data to clearly distinguish the band of each speaker.

The operation of each of these components can be controlled by the system.

The feature vector extraction unit 10 extracts at least one feature vector from a speech signal on a single channel including at least one speaker using a signal feature. The feature vector is at least one feature vector extracted using Pitch, Energy, and Mel Frequency Cepstral Coefficients (MFCC). For the analysis of the speech signal, a segmentation using a sampling window Frame.

2, the feature vector extracting unit 10 includes a speech signal dividing unit 11, a sampling windowing unit 13, a feature vector generating unit 15, a non-speech region extracting unit 17, . ≪ / RTI >

The voice signal dividing unit 11 divides a voice signal generated on a single channel into a predetermined frame unit. This is to improve the accuracy of the extraction of the feature vector by cutting out the speech signal into smaller units and at the same time to accurately recognize the change of the feature with time.

The sampling windowing unit 13 can multiply the audio signal divided by frame units by a window of a predetermined overlap range and output the result. The sampling windowing unit 13 performs a process of sampling the divided signals so as to be used for frequency analysis, thereby reducing unnecessary signals and enhancing the reliability of analysis by emphasizing necessary signals. For example, in this embodiment, by applying a Hamming window as a window function, it is possible to reduce the influence of undesired high-frequency components on the feature extraction.

In addition, the sampling windowing unit 13 may be variously selected as needed as a window function, such as a hanning window, a triangle window, and the like.

The feature vector generating unit 15 generates a feature vector by using at least one extraction technique of pitch, energy, and mel-cam stream (MFCC) from the voice signal sampled through the sampling windowing unit 13 Can be extracted. Pitch refers to the high and low of the speech signal frequency, which can be the most basic and important feature element for analyzing the speech signal. Energy is an indicator of the size of the signal occupied by the voice, and is one of the characteristics that react sensitively to the recording environment or speaker characteristics. The Mel-Cam Strobe (MFCC) can extract a number of mathematical coefficients for modeling speech signals. The Mel-Cam Strobe (MFCC) is a sinusoidal wave form of the frequency spectrum of a voice signal expressed in Mel-Scale. Expressed easily, the frequency distribution is expressed differently so that it can be similar to the characteristic of listening to the human ear.

The feature vector generator 15 may further include a feature vector for at least one of a pitch, an energy, and a mammal's muscle (MFCC) of the speech signal, Can be calculated. Here, the mean and standard deviation values of the pitch, energy, and mel-cam signal (MFCC) of the voice signal can be used as feature vectors for analyzing the characteristics of the voice signal sender.

The feature vector generating unit 15 may further include a delta value extracting function for extracting a temporal change amount of at least one of a pitch, energy and melancholy muscle (MFCC) . In this case, the temporal variation of the pitch, energy, and mel-cam stream (MFCC) of the voice signal can be used as a feature vector for analyzing the characteristics of the voice signal sender.

The feature vector generator 15 may further calculate an average and a standard deviation value of a delta value with respect to at least one of a pitch, an energy, and a mammal's muscle (MFCC) of a speech signal . In this case, the mean and standard deviation values of the delta value for at least one of the pitch, energy and MFCC of the speech signal are also used as characteristic vectors for the characteristics analysis of the speech signal sender Can be utilized.

The non-speech section extracting section 17 extracts a non-speech section from the speech signal. This is to improve the performance of the system by limiting the data so that only the feature data of the voice can be correctly analyzed in the pattern recognition process described below.

The non-speech section extracting section 17 may be located before, after, or between the speech signal dividing section 11 and the sampling windowing section 13.

The non-speech frames extracted through the non-speech interval extracting unit 17 may be stored separately in the silent interval by extracting only the time information. In addition, the feature vectors including the time information can be accumulated and stored for the frames discriminated by speech through the non-speech interval extracting unit 17. [ The storage is performed in a feature vector storage unit (20 in Fig. 1).

Referring again to FIG. 1, the feature vector storage unit 20 stores the data (frame) of the non-speech region and the data (frame) of the speech region extracted by the feature vector extraction unit 10 and stores the data. The reason for distinguishing the voice section from the non-voice section is to provide only the voice section signal to the speaker separation section 30 to reduce the error due to the non-voice section signal.

The input data for the speaker separation in the section determined as the voice is accumulated as a feature vector including the time information. The data determined as the non-speech section is used for restoring the speaker separation data including only the voice section into the existing time line Only time information is recorded separately.

The speaker separation unit 30 extracts the feature vector of the speech interval and recognizes the speaker including at least one speaker using the pattern recognition technology based on the accumulated data.

Learning methods for pattern recognition are divided into two categories: supervised learning and unsupervised learning. When the number and type of targets to be classified are determined, the map learning method is a method of training the classification standard so that the machine can distinguish the result known by using the data for training. The autonomous learning method refers to a method in which a machine analyzes the data set itself to derive an arbitrary pattern. In this embodiment, since a speaker-independent system recognizes an arbitrary speaker and separates the voices of the respective speakers, a method of detecting an arbitrary pattern in a given data and classifying cumulative data is used by using an autonomous learning technique.

In this embodiment, an example of clustering data using a representative K-means algorithm among autonomous learning techniques will be described.

The K-means algorithm aims to divide the input data into K clusters, and finds a set that minimizes the variance of the distance difference between the center of each cluster and the member data. In order to classify the K clusters defined in the system in the input data set composed of the feature data of the speech interval, the K-means algorithm can define the representative value (center) of the K initial clusters. At this time, an arbitrary value may be used for defining the cluster representative value (center), or the data distribution range may be divided into K by using the difference between the maximum and minimum values of the data.

Equation (1) is a formula expressing the K-means algorithm as an equation. It finds a value V that minimizes the distance from the representative value mu representing k clusters for all x including the feature vector set S of the speech interval. The purpose.

If the initial cluster representation is defined by the K-means algorithm, ① measure the relationship between representative values of each cluster and all the data included in the input data set. The relationship diagram measurement in this embodiment uses the Euclidean distance.

Equation (2) below expresses a Euclidian formula for finding the distance between two points in an n-dimensional space.

(2) When the distance for each cluster is obtained for all the input data, each input data is allocated to the closest cluster. At the end of the assignment process, the center of the new cluster is defined using only the data of each cluster.

The K-means algorithm searches the center of the cluster, finds the optimal classification by repeating the above steps ① and ② to measure the relationship (distance) between the center of the cluster and each data (distance) Perform until the center no longer changes. The fact that the center of the cluster no longer changes means that it has found an optimal boundary to divide the data set into K clusters. The optimal boundary can be regarded as a reference point separating at least one speaker.

When the speaker separator 30 has converged on the result of minimizing the distance between all the data and the center of each cluster in the speaker separator 30, the cumulative data of the non- And synchronizes the data with the time line represented by the voice signal of the mobile terminal.

 Next, according to the filter setting defined by the system, the data less than a certain standard may pass through the error correction filter unit 50 for forcibly correcting the error in consideration of the surrounding situation. The data output through the error correction filter unit 50 can be defined as the start and end times of each speech segment including all speakers and non-speech intervals (silent intervals).

3 is a detailed configuration diagram of an error correction filter unit in a speaker separation system according to an embodiment of the present invention.

The error correction filter unit 50 according to the present embodiment largely proceeds to two stages, where one stage is applied to an extraction unit and the second stage is applied to an analysis unit.

First, the first correction unit 51 forcibly corrects irregular data appearing instantaneously between consecutive flows in the speaker separation result existing in the extraction unit.

For example, when one other data set as a threshold value is mixed among 30 pieces of data indicating the speaker A, the corresponding data can be forcibly corrected to A. For a voice for a hundredth of a second between data of another speaker, it is regarded as an error due to the ambiguity of the frequency, or even a voice of another speaker is meaningless. The error correction filter unit 50 provides a first threshold for determining and correcting such data as an error, and the first threshold can be added and modified according to the application site or voice signal.

Next, for the second correction unit 53, it is necessary to restore the timeline of the original voice signal by integrating data accumulated in the silent interval previously. The speaker separation data and the silence interval data are integrated to calculate the classification result for all the time included in the original input data, and the process is then converted back to the analysis unit time 1 second. For example, according to the set values provided in the present embodiment, 50 extraction unit data can be included for 1 second as the analysis unit time. The results are aggregated into 50 units, and the number and specific gravity data of each speaker and silence interval are calculated and stored for each analysis unit.

Here, if we look at the weighted data of the derived speakers, the upper data centered on a specific weight can be determined as a section representing the speaker, and the lower data can be determined as a two-stage correction target. At this time, an operation of forcibly correcting the correction target distributed around the representative section is performed. The corresponding section can be corrected based on the specific gravity value from the shortest interval to the longest interval among the correction target sections. The correction step is a part that needs to be adjusted according to the characteristics of the site or voice data, and it requires a careful setting of the user as it has a small influence on the reliability.

The result calculated through the first correcting unit 51 and the second correcting unit 53 can be divided into K + 1 classifications including each speaker and a silent section. Therefore, the data that occupies the highest weight based on the specific weight data of each speaker and the silent section existing in the analysis unit by the ignition interval detection unit 40 can be defined as the corresponding classification.

4 is a flowchart illustrating a speaker separation method according to an embodiment of the present invention.

First, in a first step S10, when a speech signal is input to the speaker separation system, the feature vector extraction unit 10 in the system divides one speech signal into predetermined frame units through the speech signal division unit 11. [

Next, in step S11, the feature vector extracting unit 10 multiplies the speech signal divided in frame units by a window having a predetermined range in the sampling windowing unit 13, and performs sampling. At this time, the window function can be applied to the Hamming window to reduce the influence of unwanted high frequency components on the feature vector extraction. In addition, Hanning window, Triangluar window and various other options are available.

Next, in step S12, the feature vector extracting unit 10 samples at least one voice signal sampled through the sampling windowing unit 13 using pitch, energy, and mel-cam stream (MFCC) The above feature vectors are extracted.

In extracting the feature vector, the feature vector extracting unit 10 is a feature vector for analyzing the feature of the speech signal sender, and includes at least one of pitch, energy, and MFCC of the speech signal The average and standard deviation values for the above can be additionally calculated. If necessary, a delta value indicating a temporal change amount of the pitch, energy, and MFCC of the audio signal, and an average and standard deviation value of the delta value can be calculated.

In step S13, the feature vector extracting unit 10 extracts a non-speech region from the speech signal through the non-speech region extracting unit 17. As described above, in order to improve the performance of the system by limiting the data so that only the feature data of the voice can be correctly analyzed in the pattern recognition process.

After extracting, the feature vector extracting unit 10 extracts only the time information, which is determined as non-speech, and stores the accumulated time information in the feature vector storing unit 20 as a silent period. In addition, the frame discriminated by speech accumulates the feature vectors including the time information and stores the feature vectors in the feature vector storage unit 20.

If the feature vectors of all the speech sections are accumulated, the speaker separating section 30 recognizes and separates the speaker including at least one or more based on the accumulated feature vectors in step S14. At this time, the speaker separation unit 30 can apply the pattern recognition technology using the K-means algorithm.

For example, if a speaker is separated from the recorded data of a conversation between two speakers, the system can define a set value of 2 in the initial cluster representative K of the K-means algorithm since the system needs to recognize and distinguish two speakers . The K-means algorithm is based on the step of defining the center of the cluster (cluster) to find the configuration that minimizes the dispersion of the distance difference between the center of each cluster (cluster) and the member data, And assigning it to the cluster is repeatedly performed.

Thereafter, in step S15, the speaker separation unit 30 confirms whether the center of the cluster (cluster) moves through the K-means algorithm. This step is performed only when the K-means algorithm is applied. If the pattern recognition technology of another technique is applied, the above process can be deleted or modified according to the control of the algorithm.

When the center of the cluster (cluster) moves, the relationship (distance) between the center of the cluster and each data is measured and assigned to the nearby cluster, and the optimum classification is searched for. Perform it until it no longer changes.

As a result, if the center of the cluster (cluster) stops moving, it finds an optimal boundary that divides the data set into K clusters.

When the speaker separation is performed, the error correction filter unit 50 can forcibly correct the error in consideration of the surrounding conditions and the like, with respect to data less than a certain standard in the next step S16. For example, when one other data set as a threshold value is mixed among 30 pieces of data indicating the speaker A, the corresponding data can be forcibly corrected to A.

Next, in step S17, the speech segment detection unit 40 combines the speaker separation result output through the error correction filter unit 50 or the speech segmentation unit 30 with the cumulative data of the previously stored non-speech segment, Synchronize the data with the timeline represented by the voice signal. Therefore, the start and end time of the speech section is detected based on all the speakers and the non-speech section (silence section).

In general, there is a limitation in separating a speaker based on a feature vector extracted from a speech signal. This is because of the characteristics of the frequency components of the feature vector. In the case of the male and female voices, the voice is usually located in a different frequency band. However, in cases where the voice of a gender such as a man, a man, a woman and a woman is to be separated, an error may occur because the similarity is large between the two speaker ranges. In this case, it is necessary to apply a correction algorithm capable of forcing ambiguity in the speaker separation result.

For the data correction, we introduce the concept of weight by expanding the data analysis time unit. In order to achieve this, the system manages two unit times: one is the extraction unit time for dividing the frame of the speech and extracting the features, the next is the analysis unit time for calculating the specific weight by grouping the speaker separation results classified by the extraction unit to be. For example, in the embodiment of the present invention, 50 feature vectors are extracted per second using an extraction unit time of 20 ms based on the 8 KHz sampling rate, and the speakers are separated.

The result of the speaker separation and correction work can be divided into K + 1 classifications including each speaker and a silent section. The utterance interval detection unit 40 defines the data that occupies the highest weight based on the weight data of each speaker and silence interval existing in the analysis unit.

According to this process, the speaker and the speaker can be distinguished from the speech signal generated on the single channel. Therefore, in the reproducing apparatus such as the music player and the moving picture player, the speech signal is not moved in time or frame unit, In the unit of the ignition interval. Also, in the case of a language learning machine, it can be used for movement of a sentence unit. Furthermore, it is possible to apply a range of words or sentences to the dictation function utilizing the STT (Sound To Text) technology for converting voice commands and voice to characters using the voice recognition technology to provide a better analysis will be.

The foregoing description is merely illustrative of the present invention, and various modifications may be made by those skilled in the art without departing from the spirit of the present invention. Accordingly, the embodiments disclosed in the specification of the present invention are not intended to limit the present invention. The scope of the present invention should be construed according to the following claims, and all the techniques within the scope of equivalents should be construed as being included in the scope of the present invention.

10: feature vector extraction unit 20: feature vector storage unit
30: speaker separation unit 40: ignition interval detection unit
50: error correction filter unit 11: audio signal division unit
13: sampling windowing part 15: feature vector generating part
17: Non-speech section extraction section 51: First speech recognition section
53: Second correction unit

Claims (10)

A feature vector extracting unit for dividing a speech signal including at least one speaker into a predetermined frame unit and extracting a feature vector from the speech signal of the frame period;
A feature vector storage unit for classifying and storing feature vectors extracted through the feature vector extracting unit into a speech segment or a non-speech segment;
A speaker separator for separating the speaker using the pattern recognition technique based on the accumulated data when the feature vectors of the voice segments stored in the feature vector storage unit are accumulated;
Accumulating cumulative data of a speaker-specific voice section separated through the speaker separating section and cumulative data of a non-voice section pre-stored in the feature vector storing section on a time line and extracting, based on the non- A loudspeaker section detecting section for detecting the start and end times of the loudspeaker section per speaker;
And an error correction filter unit,
Wherein the feature vector storage unit stores,
And storing and accumulating feature vectors including time information for a frame section classified into the voice section, extracting and storing only time information for a frame section classified into the non-voice section,
Wherein the error correction filter unit comprises:
A first correcting unit correcting irregular data instantaneously appearing between consecutive flows with respect to cumulative data of the voice section existing in an extraction unit before the speaker is separated;
Calculating a specific weight value of each speaker and a non-speech section from the original voice signal obtained by synchronizing the cumulative data of the speaker's voice section separated by the speaker separating section and the cumulative data of the non-voice section on the time line, A second correcting unit
Wherein the speech characteristic vector is a speech signal. The method according to claim 1,
The feature vector extracting unit
A speech signal dividing unit for dividing a speech signal including at least one speaker into a predetermined frame unit;
A sampling windowing unit for sampling and multiplying a window function having a predetermined size by each divided frame unit through the audio signal dividing unit;
A feature vector generating unit for calculating feature vectors using at least one of pitch, energy, and MFCC for a speech signal output through the sampling windowing unit; And
A non-speech section extracting section for extracting a non-speech section (silence section) from the feature vector calculated by the feature vector generating section and classifying the speech section into a speech section and a non-
Wherein the speech feature vector comprises a speech feature vector. delete The method according to claim 1,
The speaker separator
A representative value of each cluster is defined to classify cumulative data of the voice interval into clusters by a speaker using the K-means algorithm according to the following equation as the pattern recognition technique, And a step of determining a distance between the accumulated data and allocating the distance to the nearby cluster is repeatedly performed.

(Where x means to calculate a value V that minimizes the distance between all the input data including the feature vector set s of the speech interval and the representative value mu representing k clusters for all x) delete A method in a speaker separation system for separating a speaker from a speech signal comprising at least one speaker,
Extracting a feature vector from a speech signal input by the speaker separation system;
Classifying and storing the feature vector extracted by the speaker separation system into a speech section or a non-speech section;
If the feature vectors of the stored speech segments are accumulated, separating the speakers using the pattern recognition technique based on the accumulated data;
Wherein the speaker separation system synchronizes the accumulative data of the separated voice section of the speaker with the accumulated data of the non-voice section pre-stored through the step of classifying and storing the divided voice section on the timeline, Detecting a start and end time of a speaker-based speech section based on the non-speech section from a star speech signal;
And correcting the error before and after the step of separating the speaker,
The step of classifying and storing the feature vector into a speech section or a non-
Wherein the speaker separation system stores and accumulates feature vectors including time information for a frame section classified into the voice section, extracts and stores only time information for a frame section classified into the non-voice section,
The step of correcting the error includes:
A first correcting step of correcting irregular data instantaneously appearing between consecutive flows with respect to cumulative data of the voice section existing in an extraction unit before the speaker is separated;
Calculating a specific weight value of each speaker and a non-speech section from the original voice signal obtained by synchronizing the cumulative data of the speaker-specific voice section and the accumulated data of the non-voice section on the time line through the step of separating the speaker, A second correction step of correcting the data to occupy a high specific gravity
And extracting a speech feature vector from the speech feature vector. The method according to claim 6,
Wherein the extracting of the feature vector comprises:
Dividing a speech signal including the at least one speaker into a predetermined frame unit;
Multiplying each of the divided frame periods by a window function of a predetermined size and sampling;
Extracting feature vectors using at least one of a pitch, an energy, and a mammal's muscle (MFCC) for the sampled speech signal;
Extracting non-speech sections (silence sections) from the extracted feature vectors and classifying them into speech sections and non-speech sections
And extracting a speech feature vector from the speech feature vector. delete The method according to claim 6,
The step of separating the speaker
Using the following equation,
A representative value of each cluster is defined to divide the cumulative data of the voice interval into clusters by speaker and a representative value of the cluster and a distance between the cumulative data are obtained and assigned to nearby clusters repeatedly A speaker separation method using a speech feature vector characterized by separating a speaker.

(Where x means to calculate a value V that minimizes the distance between all the input data including the feature vector set s of the speech interval and the representative value mu representing k clusters for all x) delete

Images