DE10244699A1 - Voice activity determining method for detecting phrases operates in a portion of an audio signal through phrase detection based on thresholds - Google Patents

Abstract

Energy values (EV) for a time segment in an audio signal are recorded in a histogram. Using EV-detected distribution determines voice (6) and pause (5) thresholds, which are compared with a current EV to produce a decision on the limits of a phrase between speech and pause so as to determine and store a 'start' label (8) for the beginning of each phrase and a 'stop' label (7) for its end.

Description

The invention relates to a method to determine speech activity in a signal section of an audio signal by a threshold-based Phrase detection.

The development of robust phrase limit detectors is important for automatic recognition of fluent speech. Such detectors are used, for example, in signal processing in the mobile radio sector used to improve detection rates and data reduction to reach the relevant information. Other areas of application are in the areas of command word recognition, echo cancellation and noise reduction.

The phrase detection must be economical Use certain requirements. These include alongside the robustness of the detection the fast adaptation to changing Environmental conditions and a relatively low resource requirement both data storage as well as the necessary computing effort.

Very simple and time efficient Phrase detection algorithms are related to the security of the detection generally not very efficient. The phrase boundaries of the phrases to be detected often become inaccurate found so that there is data loss on the one hand and on the other false detections can occur. One understands data loss the lack of consideration relevant information, such as spoken sounds, that belong to the utterance, that are partially meaningful and by the phrase detector marked as not belonging to the utterance were. False detections in the other case are marked as phrases Signal sections that do not represent a linguistic expression.

At the current state of development, three performance classes of phrase limit detectors can be identified for phrase detection. The first class includes simple, energy threshold based and time domain detectors like in the DE 100 26 872 A1 shown. These evaluate time signals after a threshold value determination of the determined energy of a specific signal section (window) and are therefore generally fast and can be implemented with little modeling effort. The detection rate determined in this way is strongly dependent on the signal and its background noise.

The second class includes more powerful, detectors operating in the frequency domain as described in the ETSI EN 301 708 V7.1.1 (1999-12), from December 1999. This values transformed into the frequency range and divided into frequency channels Signals from are therefore common complex and need a big Computational effort. A higher one Detection reliability can be achieved as many parameters (pitch, signal-to-noise ratio, Peak-to-average ratio, etc.) used for decision making become.

The third class summarizes the elaborate and extensive statistical procedures. Through the evaluation the probability density function (PDF) or the creation of models, with the help of an HMM (Hidden Markov Model), can be computationally expensive high detection reliability can be achieved. A more detailed description in Son, Jongseo: "A Statistical Model-Based Voice Activity Detection ", IEEE Signal Processing Letters, vol. 6., No. 1, January 1999.

For the implementation of phrase limit detectors in systems with low Resources therefore only come from detectors of the first performance class in question. So far, these simply implemented detectors but with insufficient detection reliability and adaptation to changing Environmental conditions.

The invention is therefore the object based on a method for determining language activity in a Specify signal section of an audio signal in which the opposite reduced between reliable detection and low computing effort and that a robust demarcation of the language from time-varying Background noise realized.

According to the invention the object solved by that in a first step energy values of a period of time Audio signal recorded in a histogram that in a second step, based on the determined distribution of the energy values, a language threshold and a pause threshold are set and that by comparing the thresholds with the current energy value made a phrase boundary decision between language and pause becomes.

At the beginning, a signal section (time window) is examined in chronological order and its energy is determined. This short-term energy value is arranged in a histogram that estimates the long-term distribution of the signal energy. The parameters mean X and variance s are determined for this estimated distribution. The ThrVoice language threshold and the ThrPause pause threshold are determined in the threshold adaptation based on these two parameters. By using two thresholds, the robustness of the phrase limit decision against smaller energy fluctuations is increased, as in 4 shown.

In one embodiment of the invention it is envisaged that the determination of the language threshold and the pause threshold, keeping pace with the signal curve, before or after a phase limit decision he follows.

Robust and fast adaptation to changing environmental conditions is achieved by determining the thresholds, which keeps pace with the signal curve. The calculation of the signal energy is short-term based for the length of a time window. The interval between two successive time windows (the continuation rate) controls the temporal resolution of the phrase limit decision. With a low continuation rate, good resolution in the time domain is achieved.

Realizing a phase boundary decision before updating the thresholds, based on the existing thresholds of a previous process the classification of the energy values occurring only during the breaks. Since the window continuation rate alone decides how much time between the currently considered period and the period in which the thresholds have been adapted, has passed and this is usually relative is small, the decision-making error is kept small.

In a further embodiment of the Invention is provided that as a result of the phase boundary decision a label "Start" for the beginning a phrase and a label "stop" for the end a phrase per phrase is determined and that the respective label and the associated Times are saved.

The phrase limit decision compares the current energy value with the determined thresholds and determined the state of the signal. Two states are used in phrase limit detection distinguished. The first state characterizes the break or the background noise and marks the beginning of the area that does not contain a phrase with the label "Stop". This state is carried out after the break below the break threshold for the first time the signal energy is ingested and continues until it goes from a second state superseded becomes. The second state is assumed when a phrase is present, that is if the signal energy exceeds the speech threshold for the first time. The beginning of this Area is identified by a "start" label. The duration of this area is only when the level falls below again the break threshold, through which the signal energy ends.

In one embodiment of the invention it is intended that a minimum and a maximum phrase length as well a minimal pause length be determined and a plausibility check is carried out in such a way that label their associated Time intervals do not correspond to the phrase lengths or the pause length the label track can be eliminated.

Robust phrase limit detection is not only guaranteed by the adaptation of the thresholds. Wrong decisions in phrase limit detection are made by avoided correcting the decision. The correction will then be made carried out, when there is an entire phrase. It consists of a review of the minimum break length, the minimum and maximum expected phrase duration. Correcting the minimum break length causes pauses detected within the audio signal, for example by shorten gaps within of words are not marked as breaks. Control of minimum phrase length removed short sections marked as phrases and the check of the maximum phrase duration removed long and unexpected segments.

mit einemIn one embodiment of the invention it is provided that the energy values according to the equation

with a

RMS value X of a signal section the width N can be calculated.

This energy value is in the histogram or classified in the distribution contained therein such that the number of values in the histogram in the steady state State remains constant. The steady state is reached if there are sufficient values in the histogram, because only after a certain number of values in the histogram becomes the actual Distribution enough accurately estimated by the histogram. So that not all signal energy fluctuations affect the distribution of energy negatively affect only energy values in the histogram added that not too far from the maximum of the current distribution lie away. This decision can be made by linking Distribution variance and distribution mean are taken.

erfolgt.In a special embodiment of the invention it is provided that after the first step the energy values recorded in the histogram are smoothed according to the formula

he follows.

This results in the smoothed histogram entry X '(N) of the Nth histogram interval from the sum of the weighted two left and the two right neighboring histogram entries X (N - 2) , X (N - 1) , X (N + 1) and X (N + 2) as well as the weighted entry X (N) itself.

In a further embodiment of the invention it is provided that the break threshold, with a Adaptation factor α to control the speed of adjustment and a parameter β to determine the distance of the break threshold from the mean X , according to the equation ThrPause '= (1 - α) ThrPause + α ( X + βs) is determined.

The pause threshold ThrPause, which is important in the phrase limit decision for the detection of the phrase end, is determined from the above equation. In the case of threshold adaptation, the distribution parameters mean the mean X and variance s determines the thresholds. The adaptation factor a controls the sensitivity to adaptation. If this value is close to zero, the adaptation is very slow, but very close to one. The parameter β determines how far the break threshold is from the mean X is placed away.

In a special configuration the invention provides that the speech threshold, with an adaptation factor α for control the rate of adaptation and a parameter γ to determine the distance of the speech threshold from the pause threshold, according to the equation ThrVoice '= (1 - α) ThrVoice + α (ThrPause + γs) determined becomes.

The adaptation of the ThrVoice language threshold is based on the calculated pause threshold ThrPause. In the above equation is α again the adaptation factor with which the speed of the adaptation is controlled, and γ determines how big the Distance of the ThrVoice language threshold to the ThrPause pause threshold is. The link between the speech and pause threshold shown in the equation with the variance s of the distribution has the advantage that the distance of the two thresholds depends on the distribution of short-term energy. This positions itself differently for changing Background noise represents.

In quiet environments (static Case) the distribution of short-term energy is narrow, which is indicated by a small one Expressed variance becomes. This means, the short-term energy fluctuates relatively little around its long-term average. Dynamic changes in background noise usually occur in noisy environments, which result in a wide distribution of short-term energy. In this case, the variance is large because the short-term energy is strong fluctuates around their long-term mean. By a small distance the thresholds in the static case and by a large distance the thresholds in the dynamic case become incorrect detection of the phrase boundaries limited.

The invention is based on the following two embodiments are explained in more detail. In the associated Shows drawings

1 a first variant of the procedure,

2 a second variant of the procedure,

3a a histogram with energy values,

3b a smoothed distribution and derived parameters,

4 a sample phrase with thresholds,

5 an example of the threshold adaptation to a changing background noise and

6 a possible energy distribution for voice and interference signal.

The process according to the invention can be carried out in different ways Areas are used. In speech signal processing can detection of useful signals and reliable start and end point detection for one Command word recognizer can be realized. The procedure enables the Realization of noise suppression, at which is a break detection for adaptation processes it is necessary to determine a speaker activity for the area echo cancellation or a determination of the channel load in the Telephony area.

A first variant of the process flow is in the 1 shown. In a first step, a signal section of an audio signal (time window) and its energy values are examined in chronological order 1 certainly. These energy values 1 are divided into a distribution in the form of a histogram 2 according to 3a classified and thus created or updated the distribution. The energy values 1 are in the histogram 2 or arranged in the distribution contained therein such that the number of in the histogram 2 values in the steady state remains constant. The steady state is reached when there are sufficient values in the histogram 2 are included, because only after a certain number of values in the histogram 2 can be spoken of a distribution. So that not all signal energy fluctuations negatively influence the distribution of energy, only energy values 1 into the histogram 2 recorded that are not too far from the current distribution. This decision can be made by linking distribution variance 3 and average distribution 4 to be hit.

After smoothing the distribution, the method evaluates the histogram 2 and determines an average X 4 and the variance s 3 , as in 3b shown. Using these distribution parameters 3 and 4 the pause threshold becomes ThrPause 5 determined according to the specified calculation rule. ThrPause '= (1 - α) ThrPause + α ( X + βs)

The adaptation factor α controls the sensitivity to adaptation. If α is close to zero, the adaptation is carried out very slowly; if α is close to one, the adaptation takes place very quickly. The parameter β influences the Distance of the break threshold 5 from the determined mean X 4 ,

Basis for determining the ThrVoice language threshold 6 are the previously determined pause threshold ThrPause 5 and the variance s 3.

The determination is made using the equation: ThrVoice '= (1 - α) ThrVoice + α (ThrPause + γs)

With α again the speed the adaptation set. The distance of the language threshold from the The pause threshold is influenced by γ.

The relationship of the ThrVoice language threshold shown in the equation 6 with the variance s 3 The distribution has the advantage that the distance between the two thresholds 5 and 6 depends on the distribution of short-term energy. This is different for changing background noise (see 5 ). In quiet environments (static case) the distribution of short-term energy is narrow, which is due to a small variance 3 is expressed. This means that short-term energy fluctuates relatively little around its long-term average. In noisy environments, dynamic changes in background noise usually occur, which result in a wide distribution of short-term energy. In this case, the variance 3 large because the short-term energy fluctuates strongly around its long-term average. By a small distance between the sleepers 5 and 6 in the static case and a large distance between the thresholds 5 and 6 in the dynamic case, incorrect detections of the phrase boundaries are restricted.

By a subsequent comparison of the determined thresholds 5 and 6 with the current energy value 1 a phrase boundary decision is made between language and pause. A distinction is made between two states. The first state characterizes the pause or the background noise. The beginning of this area, which does not contain any phrase, is marked with the label “Stop” 7. This state is assumed by the signal energy after the pause threshold has been undershot for the first time and continues until it is replaced by a second state. The second state becomes then taken when a phrase is available, that is when the signal energy exceeds the speech threshold for the first time. The beginning of this area is marked by a "start" label 8th characterized. The duration of this area is only ended when the pause threshold is exceeded again due to the signal energy. The result of this phrase detection is in 4 presented result. In the illustration is a sample phrase 9 with the currently determined thresholds 5 and 6 and the set start 8th and stop label 7 shown.

Robust phrase boundaries are not only achieved by adapting the thresholds 5 and 6 guaranteed. Wrong decisions in phrase limit detection are avoided in a subsequent step by a plausibility check. The check is carried out when an entire phrase is available. It consists of a check of the minimum break length as well as the minimum and maximum expected phrase duration. Checking the minimum pause length causes detected pauses within a phrase 9 are not marked as breaks. Checking the minimum phrase length eliminates short sections marked as phrases. The control of the maximum phrase duration filters out long and unexpected segments. An almost error-free sequence of labels with associated time intervals can thus be transferred to a subordinate stage, for example a command word recognizer.

A second variant of the process flow is in the 2 shown. The difference to the first variant is that the phrase limit decision is not made after determining the thresholds 5 and 6 but is carried out before their determination.

After the energy has been determined over a considered time window, the thresholds present in the previous sequence are used 5 and 6 made a phrase boundary decision. Since the window continuation rate alone decides how large the time difference between the currently considered period and the period in which the thresholds were adapted, and this is normally kept relatively small, the decision-making error can be kept small.

The advantage of the phrase limit decision before the histogram adaptation is that with the help of this decision the energy values that occurred during the breaks are classified 1 is possible. Thus the generated distribution is only a distribution of the background noise. Through the exclusive consideration of energy values 1 from breaks, the distribution in the histogram adjusts 2 quickly to the ambient conditions and is narrow in many cases, i.e. there is a small variance 3 in front. If one looks at the distribution of all energies that occur, in many cases one of the results 6 shown image. Two maxima of the distribution are clearly visible. In the left part of the 6 it is the distribution of the background noise and in the right part of the figure the distribution of the spoken utterances.

1

energy values

2

histogram

3

distribution variance s

4

Distribution mean X

5

pause threshold ThrPause

6

speech threshold ThrVoice

7

"Stop" label

8th

Start label

9

example sentences

Claims (8)

Method for determining the speech activity in a signal section of an audio signal by a threshold-based phrase detection, characterized in that in a first step energy values ( 1 ) a period of the audio signal in a histogram ( 2 ) that in a second step, based on the determined distribution of the energy values ( 1 ), a language threshold ( 6 ) and a break threshold ( 5 ) and that by comparing the thresholds ( 5 and 6 ) a phrase limit decision between speech and pause is made with the current energy value. A method according to claim 1, characterized in that the determination of the language threshold ( 6 ) and the break threshold ( 5 ), keeping pace with the signal curve, before or after a phase limit decision. A method according to claim 1, characterized in that a label "start" ( 8th ) for the beginning of a phrase and a label "Stop" ( 7 ) for the end of a phrase is determined for each phrase and that the respective label ( 7 and 8th ) and the associated times are saved. Method according to Claims 1 and 3, characterized in that a minimum and a maximum phrase length and a minimum pause length are defined and a plausibility check is carried out in such a way that label ( 7 and 8th ) whose associated time intervals do not correspond to the phrase lengths or the pause length from which the label track is eliminated. A method according to claim 1, characterized in that the energy values ( 1 ) according to the equation

with an effective value X of a signal section of width N can be calculated. A method according to claim 1, characterized in that after the first step a smoothing of the in the histogram ( 2 ) recorded energy values ( 1 ) according to the formula

he follows. A method according to claim 1, characterized in that the pause threshold ( 5 ), with an adaptation factor α to control the speed of adaptation and a parameter β to determine the distance of the break threshold ( 5 ) from the mean X ( 4 ), according to the equation ThrPause '= (1 - α) ThrPause + α ( X + βs) is determined. A method according to claim 1, characterized in that the speech threshold ( 6 ), with an adaptation factor α to control the speed of adaptation and a parameter γ to determine the distance of the speech threshold ( 6 ) from the break threshold ( 5 ), according to the equation ThrVoice '= (1 - α) ThrVoice + α (ThrPause + γs) is determined.