Method and apparatus for speech analysis

i

METHOD AND APPARATUS FOR SPEECH ANALYSIS 1 531-O1AP.DOC D-13 6.8.2001

METHOD AND APPARATUS FOR SPEECH ANALYSIS FIELD OF THE INVENTION 5 The present invention relates to the field of voice and speech analysis and in particular to the analysis of acoustic and prosodic features of speech. BACKGROUND OF THE INVENTION It is long known that certain voice characteristics carry information regarding 10 the emotional state of the speaker. Lynch described already in 1934 differences in timing and pitch characteristics between factual and emotional speech. (Lynch, G. E. (1934). A Phonophotographic Study of Trained and Untrained Voices Reading Factual and Dramatic Material, Arch. Speech, 1, 9-25.)

Since then, many studies have demonstrated different correlates between 15 various non-verbal speech characteristics and various emotional states. Research efforts were directed to several different aspects of the emotional speech phenomena: One line of research focuses on identifying the carriers of emotion within the speech signal and studies show complex correlation patterns of pitch, amplitude, timing, duration, pace, envelope contours and other speech variables to the emotional 20 content. A second research area tries to explore the expression of different emotional dimensions in speech - the studies suggest correlates to emotional dimensions of strength, valence, activity and others. A further research effort focuses on revealing the distinctive speech correlates to various emotional states including primary emotions (e.g. anger), secondary emotions (e.g. boredom) and specific stressful 25 situations such as anxiety, workload and lying. Yet another area of research tries to point out the differences in emotional speech patterns of different individuals, different groups of individuals (e.g. sex, age, culture and personality) and different physiological states of the same individuals. 1

Three extensive literature reviews, summarizing the various findings regarding the vocal expression of emotion, were published by Murray, I. R. and

Amott J. L., (1993), Towards the Simulation of Emotion in Synthetic Speech: A

Review of the Literature on Human Vocal Emotion, Journal of the Acoustical 5 Society of America, vol. 93 (2), 1097-1108, by Frick, R. W. (1985), Communicating Emotion: The Role of Prosodic Features, Psychology Bulletin, 97, 412-429, and by Scherer, K. R. (1986), Vocal Affect Expression: A Review and a Model for Future Research, Psychology Bulletin, 99, 143-165. All the writers emphasize the fragmented nature of the research in the field and point out that the vocal emotion 10 research is forming a very small and isolated part of the general emotion literature and the general speech analysis literature. The reviews support the notion that human voice characteristics vary in relation to expression of emotion; yet, it highlights the complexity of the interplay between physiology, psychology and speech regarding emotions. They also stress the need for generalized models for a more coherent 15 understanding of the phenomena.

In recent years, a few studies approached the task of automatic classification of vocal expression of different emotional states by utilizing statistical pattern recognition models. The results indicate relative success (Dellaert, F., Polzin, T.S. and Waibel, A. (1996), Recognizing emotions in speech. In Proc. ICSLP, 20 Philadelphia PA, USA, 1996; Amir, N. and Ron, S. (1998), Towards an automatic classification of emotions in speech. In Proc. ICSLP, Sydney, 1998.)

The field of emotion in speech is gaining interest and a special workshop on this topic was held in Belfast in September 2001 (ISCA workshop on Speech and Emotion - presented papers: http://www.qub.ac.uk/en/isca/proceedings/index.html). 25 The papers, theoretical and empirical, reveal once more the complexity of the phenomena, the lacking data and the various aspects that are involved.

In respect with the detection of the emotion through speech analysis, the literature highlights several problems that still wait for a solution. We would like to emphasize two of the major problems: 2

The first is the lack of a unified model of emotional acoustic correlates, which will enable to address different emotional content in speech with one general scale. The current state of the research enables only to point out isolated acoustic correlates of specific emotional states. 5 The second is the difficulty to overcome the different speech expression patterns of different speakers, which tend to mask the emotional differences. The research tried to confront the latter problem by obtaining prior reference speech characteristics of the tested individual (or of specific groups of individuals). The reference was either prior baseline measurements (non-emotional) of a specific 10 subject, or the specific emotional speech profiles of relatively homogenous groups of subjects (e.g. all subjects suffering from depression).

Several patents regarding this field were registered during the years. These patents suffer mainly from the same limitations described above in regard of the academic research, namely, focusing on specific emotional states and depending on 15 prior reference measurements. The patents also vary significantly in their measurement procedures and parameters.

Fuller, in three U.S. Patents from 1974 (No. 3,855,416; No. 3,855,417 and No. 3,855,418), suggested a method for indicating stress in speech and for determining whether a subject is lying or telling the truth. The suggested method 20 measures vibratto content (rapid modulation of the phonation) and normalized peak amplitude of the speech signal, in particular of a subject under interrogation.

Bell et. al., in 1976 (U.S. patent No. 3,971,034), also suggested a method for detecting psychological stress through speech, based mainly on the measurement of infrasonic modulation changes in the voice. 25 Williamson, in two U.S. patents from 1978 and 1979 (No. 4,093,821 and No. 4,142,067) described a method of analyzing frequency perturbations in the speech pattern in order to determine the emotional state of a person. Analysis is based mainly on measurements of the first formant frequency of speech. The differentiation 3 between emotional states is not specified clearly - in the first patent the apparatus indicates mainly stress versus relaxation and in the latter the user of the device should apply “visual integration and interpretation of the displayed output” for “making certain decisions with regard to the emotional state”.

Jones, in 1984 (U.S. patent No. 4,490,840), suggests a method for determining patterns of voice-style (resonance, quality), speech-style (variable-monotone, choppy-smooth, etc.) and perceptual-style (sensory-internal, hate-love, etc.), based on different voice characteristics including six spectral peaks and pauses within the speech signal. Yet, the inventor states, “the presence of specific emotional content... is not of interest to the invention disclosed herein.”

Silverman, in two U.S. patents from 1987 and 1992 (No. 4,675,904 and No. 5,148,483) suggests a method for detecting suicidal predisposition in a person by identifying in his speech substantial decay on utterance conclusion and low amplitude modulation during the utterance.

Ron, in 1997 (U.S. patent No. 5,647,834), describes a speech based biofeedback regulation system that enables a subject to monitor and to alter his emotional state. An emotional indication signal is extracted from the subject’s speech (the method of measurement is not described in the patent) and compared to online physiological measurements of the subject that serve as a reference for his emotional condition. The subject can then try to alter the indication signal in order to gain control over his emotional state.

Bogdashevsky, et. al., in a U.S. patent from 1999, (No. 6,006,188) suggests a method for determining psychological or physiological characteristics of a subject based on the creation of specific prior knowledge bases for certain psychological and physiological states. The process involves creation of homogenous groups of subjects by their psychological assessment (e.g. personality diagnostic groups according to common psychological inventories), analyzing their unique speech patterns (based on cepstral coefficients) and forming specific knowledge bases for 4 these groups. By comparing the speech patterns of an individual (who is asked to speak a 30-phrase text similar to the text used by the reference group), to the knowledge bases characteristics, matching to certain psychological and physiological groups can be accomplished. The patent claims to enable verbal psychological 5 diagnosis of relatively steady conditions (e.g. mental status before and after therapy, personality profile).

Pertrushin, in 2000 (U.S. patent No. 6,151,571), describes a method for monitoring a conversation between a pair of speakers, detecting an emotion of at least one of the speakers, determining whether the emotion is one of three negative 10 emotions (anger, sadness or fear) and then reporting the negative emotion to a third party. Regarding the emotion recognition process, the patent details the stages required for obtaining such results: First, conducting an experiment with the target subjects is recommended, in order “to determine which portions of a voice are most reliable as indicators of emotion”. It is suggested to use a set of the most reliable 15 utterances of this experiment as “training and test data for pattern recognition algorithms run by a computer”. The second stage is the feature extraction for the emotional states based on the collected data. The patent suggests several possible feature extraction methods using a variety of speech features. The third stage is recognizing the emotions based on the extracted features. Two approaches are 20 offered - neural networks and ensembles of classifiers. The previously collected sets of data (representing the emotions) should be used to train the algorithms to determine the emotions correctly. Exemplary apparatuses as well as techniques to improve emotion detection are presented.

Slaney in a U.S. patent from 2001 (No. 6,173,260), describes an emotional 25 speech classifying system. The device is based on an empirical procedure that extracts the best combination of speech features (different measures of pitch and spectral envelope shape), which characterize a given set of speech utterances labeled in accordance with predefined classes of emotion. After the system has been “trained” on the given set of utterances, it can use the extracted features for further 5 classification of these emotional classes of other utterances. The procedure doesn’t present any emotional scale and only assumes that different emotional features can be empirically extracted for different emotional situations.

Two published PCT applications by Liberman also relate to emotion in 5 speech. Liberman, in 1999 (WO 99/31653), suggests a method for determining certain emotional states through speech, mainly emotional stress and lying related states - untruthfulness, confusion and uncertainty, psychological dissonance, sarcasm, exaggeration. The procedure is based on measuring speech intonation information, in particular plateaus and thorns in the speech signal envelope, and it 10 uses previous utterances of the speaker as a baseline reference.

Liberman, in 2000 (WO 00/62270), describes an apparatus for monitoring unconscious emotional states of an individual from speech specimens that are provided over the telephone to a voice analyzer. The emotional indicators comprise of sub-conscious cognitive activity level, sub-conscious emotional activity level, 15 anticipation level, attention level, “love report” and sexual arousal. The method is based on frequency spectrum analysis of the speech. The frequency spectrum is divided into four frequency regions and it is claimed that higher percentage of frequencies in one of the regions reflects dominance of one of the emotional states above. It is suggested that cognitive activity would be correlated with lowest 20 frequencies, anticipation level with highest frequencies, emotional activity with high frequencies and attention/concentration with main spectrum frequencies.

None of the above reviewed patents suggests one general scale of emotional arousal by speech, for describing the speech expression of the emotional response created by a wide range of different emotional states. Most of these patents (Fuller, 25 Bell, Jones, Silverman and Liberman) claim to identify only specific emotional states such as stress, lie or tendency to commit suicide, by correlating specific speech features to these emotional conditions. Two of the patents (Williamson, Ron) assume that the appropriate speech correlates of the emotional states are given as input and ignore at all the task of describing any scale of emotional speech features. Other 6

<img img-format="tif" img-content="drawing" file="IL144818AD00021.tif" id="idf0001" />

<img img-format="tif" img-content="drawing" file="IL144818AD00022.tif" id="idf0002" />

three patents (Bogdashevsky, Petrushin and Slaney), suggest procedures for extraction of specific speech correlates by “learning” given emotional classes of speech utterances and also ignore the description of a general scale of emotional speech features.

Furthermore, none of the above reviewed patents suggests an emotional speech scale that is valid beyond the differences between specific individuals or specific groups of individuals. In order to overcome the individual differences, some of these patents (Fuller, Williamson), require a skilled expert to analyze manually the results. Some of the patents (Ron, Liberman), require a comparison of the subject’s 10 speech measurements to prior baseline measurements of the same individual, as reference. Other patents (Bogdashevsky, Petrushin and Slaney), require a prior learning process of the speech characteristics of specific groups of individuals or specific psychological phenomena, to be used as reference. 15

The patents also vary significantly in the extracted speech features that they use in order to determine the emotional content. 7

SUMMARY OF THE INVENTION

The present invention provides one general scale of emotional arousal by speech analysis, for describing the speech expression of the emotional response created by a wide range of different emotional states. This emotional speech scale is 5 valid beyond the differences between specific individuals or specific groups of individuals.

There is provided according to the present invention a method for detecting emotional arousal through speech analysis including analyzing secondary (extracted or calculated) speech parameters which are relatively less affected by individual 10 differences than are the primary pitch and amplitude parameters.

According to a preferred embodiment of the invention, the method includes analyzing irregularity and agility of pace, rhythm and pitch indicators of speech specimens being analyzed.

According to a preferred embodiment, the method includes analyzing a 15 combination of at least two, and preferably more, of the following parameters: Pause length average; Pause frequency; Short silences length average; Short silences frequency; Very short silences length average; Very short silences frequency; Equal pitch segments length average; Equal pitch segments frequency; Rising pitch segments length average; Rising pitch segments frequency; Falling pitch segments 20 length average; Falling pitch segments frequency; Unclear pitch segments length average; Unclear pitch segments frequency.

According to one embodiment, the method further includes analyzing voice amplitude to indicate magnitude of emotional arousal.

According to one embodiment, the method includes inputting a digitized 25 voice file, pre-processing the voice file to produce consecutive cells of pitch and amplitude values, processing the pre-processed voice file to measure the emotional arousal values, and outputting the process results to an output device.

The step of inputting can include capturing of speech specimens and sampling and digitizing the voice in a voice sampling and digitizing unit. 8

<img img-format="tif" img-content="drawing" file="IL144818AD00023.tif" id="idf0003" />

<img img-format="tif" img-content="drawing" file="IL144818AD00024.tif" id="idf0004" />

The step of pre-processing preferably includes normalizing the samples’ values, data filtering, noise-reduction, segmentation of the sampled data into silence and speech segments, dividing the speech segments into cells, and auto-correlation processing to obtain pitch and amplitude parameters per cell.

The step of processing preferably includes marking a speech segment of a pre-defined number of seconds for processing; calculating silence periods within the speech segment; calculating the extracted speech parameters within the speech segment - number, lengths and averages of pauses and silences, number, lengths and averages of increasing, decreasing, stable and unclear pitch periods; classifying the 10 speech segment into one of several categories of typical parameter range; and determining the emotional arousal level of the speech segment based on comparison of the extracted voice parameters to pre-defined values that represent non-emotional speech.

Preferably, the step of processing also includes calculating a reliability grade 15 based on a combination of several factors: quality of the voice segment (noise level, size of sampled data, quality of sampled data), significance of the emotional arousal decision (number of participating parameters, degree of deviation of each parameter), and consistency of the specific segment results with the previous speech segments.

Optionally, the step of processing can also include pattern processing, to 20 detect certain emotional patterns that are revealed along the time axis or when compared to other results.

There is also provided, in accordance with the present invention, an apparatus for detecting emotional arousal through speech analysis including a voice input unit, a voice sampling and digitizing unit, a pre-processing unit, a main processing unit, 25 and a main indicators output unit. 9

BRIEF DESCRIPTION OF THE DRAWINGS

The present invention will be further understood and appreciated from the following detailed description taken in conjunction with the drawings in which:

Fig. 1 is a block diagram illustration of an apparatus constructed and operative 5 in accordance with one embodiment of the present invention;

Fig. 2 is a flow chart of a pre-processing unit constructed and operative in accordance with one embodiment of the present invention; and

Fig. 3 is a flow chart of a main processing unit constructed and operative in accordance with one embodiment of the present invention. 10 10

DETAILED DESCRIPTION OF THE INVENTION

The present invention relates to a method and apparatus for detecting emotional arousal through speech analysis. We use the term emotional speech in regard to a speech segment in which the speaker expresses in an emotional manner a current 5 feeling. Non-emotional speech refers to a speech segment in which the speaker does not express a current feeling of any kind in an emotional manner. Past descriptions of experiences and feelings or future expectations for desired or undesired events may be considered emotional speech only if the actual described or desired feeling is experienced and expressed currently in an emotional manner. The literature and 10 patents reviewed above support clearly the following phenomena: Different emotional states, when expressed vocally, are correlated with changes in the speech characteristics in comparison to the characteristics of non-emotional speech.

However, providing a general scale that can determine the emotional arousal level of a person through speech analysis is still a very difficult task, mainly because of 15 three factors: • Voice and speech characteristics vary significantly between individuals. • Different emotional states affect differently the speech characteristics. • Different magnitudes of emotions affect differently the speech characteristics.

In order to overcome these difficulties, most existing research and patents follow 20 two guidelines: They separate the measurement of different types of emotions and they use prior samples to obtain a comparable reference baseline.

The present invention suggests an automatic, real time, speech analysis method for determining the general emotional arousal level of a person at given time, beyond specific emotion states and beyond specific individual differences, without using a 25 reference speech baseline.

Eliminating the need for a reference baseline, as well as the generalization of emotional arousal voice characteristics beyond specific emotional states, are two innovations of the present invention. 11

The central assumption in the base of the present invention is that non-emotional speech reflects an equilibrium state and that emotional speech reflects a deviation from this balance. Emotional arousal is known to be a deviation from a physiological equilibrium in certain emotional states such as stress. It is expressed in changes in 5 autonomic system variables, such as heartbeat rate, muscle activity, galvanic skin resistance, blood pressure or temperature. In the same manner, the changes in the speech patterns during emotional states may reflect a deviation from the balanced, ordered non-emotional state. Therefore, the present invention is based on the principle that the speech characteristics during emotional arousal are less systematic 10 and less ordered in nature, in comparison to non-emotional speech. One clear example for that is the violation of the ordered speech rhythm when an individual is moving from a non-emotional to an excited speech.

Different emotional states may produce different speech characteristics, but it is suggested that the common factor is their irregularity in comparison to the more 15 systematic nature of non-emotional speech.

In the same manner, different individuals have different speech characteristics, but it is suggested that the common factor between them is their less ordered speech patterns when compared to a general non-emotional speech pattern.

The present invention focuses on the measurement of this common factor, in order 20 to highlight the individual’s general emotional arousal.

As stated above, voice and patterns of speech vary significantly from one person to another. Some of the differences are more general, for example, women have statistically higher pitch (basic frequency) than men. Other differences are more specific, for example, the personal typical pitch or any personal speech tendency 25 such as monotonous speech, paused speech etc.

In order to overcome these biasing effects the suggested speech processing proceeds on the following guidelines: • Use of secondary rather than primary voice pitch parameters: 12

The suggested approach uses mainly secondary voice and speech parameters and disregards primary parameters. For purposes of the application, the term secondary voice parameters indicates parameters extracted or calculated from the primary pitch and amplitude parameters. The Primary parameters are much more affected by individual differences. For example, the voice frequency itself (pitch) is totally ignored in all calculations, since it varies significantly between different people. On the other hand, pitch changes within the speech segments are emphasized, since they contribute relative, rather than absolute, values and are, therefore, less affected by the individual differences. • Selective use of secondary voice pitch parameters:

Secondary voice parameters are also sensitive up to a certain degree to individual differences. Some of the secondary parameters, which are more affected, are ignored by the speech processing as well. One example is the pitch change range parameter. This is a secondary parameter since it represents only the relative changes of the speaker’s pitch, not the pitch itself. Still, this parameter is rather correlated to the actual pitch value and therefore is often affected by individual differences. The speech processing ignores such parameters. • Disregarding voice amplitude parameters:

Voice amplitude parameters, primary and secondary, are much more affected than pitch parameters by individual speech differences, as well as by environmental effects, such as noise. The amplitude parameters do play a significant role in speech emotion interpretation, but their crucial contribution is to the determination of the magnitude of speech arousal. Therefore, the suggested speech processing limits the use of amplitude parameters mainly to this task, where relatively wide resolution is satisfactory. Determining the existence of emotional arousal is based mainly on the secondary parameters extracted from the pitch, pace, and rhythm characteristics. These parameters are less biased and carry more distinctive information. 13

<img img-format="tif" img-content="drawing" file="IL144818AD00025.tif" id="idf0005" />

10

As reported by the literature, the expression of different emotional states is “colored” by differences in certain speech characteristics. Some of the differences are correlated to the magnitude of the expressed emotion (see next paragraph) and some characterize the specific emotional state. However, two types of variables tend to characterize the emotional arousal itself, rather than the specific emotional state. The first is the agility level of the pitch changes within the speech. In general, emotional speech is related to more agile pitch changes. Many of the pitch changes are correlated either to individual speech tendencies or to magnitude differences, but after minimizing these effects it can be seen that emotional expression is characterized by higher fluctuation in pitch than non-emotional speech. The second type of “emotional” speech variable is the general less ordered and less systematic behavior of the speech features. Again, much of the parameters dispersion is caused by individual differences and magnitude differences, but when these are minimized, the irregularity relates also to emotional expression. The latter is true for various speech parameters related to pace and rhythm, as well as pitch behavior. 15 The magnitude of the emotional arousal has a significant effect on many of the speech characteristics. For example, when a person is shouting out of anger, usually his voice pitch, voice amplitude and speech speed will increase, causing an increase in many secondary speech parameters as well. This speech profile can be very different from the speech profile of a less excited form of anger, although both represent emotional arousal. 20 In order to minimize the effect of the arousal magnitude, two guidelines are followed:

The first is disregarding amplitude speech parameters for determination of emotional arousal. As mentioned earlier, the amplitude parameters are much more affected by arousal magnitude (as well as by individual differences) than are pitch 25 parameters. Disregarding amplitude parameters reduces the arousal magnitude effect.

Since secondary pitch parameters are also affected to some extent by the arousal magnitude, a second mean is used. The speech processing first classifies the speech segment according to one of several typical parameter range behavior. Then, the processing determines the existence of emotional arousal according to the 30 specific parameter range class. This enables to use different criteria in speech 14

<img img-format="tif" img-content="drawing" file="IL144818AD00026.tif" id="idf0006" />

10 15 20 25 segments that are significantly different in their parameter ranges, for reasons such as different magnitudes of arousal.

The speech characteristics that are associated directly with emotional arousal, after neglecting the parameters that are related to specific emotional states, individual differences and magnitude effects, are mainly the irregularity of pace and rhythm indicators as well as the degree of pitch agility indicators. Non-emotional speech will usually be characterized by average and consistent speech pace as well as low and consistent pitch fluctuations.

More specifically, the algorithm of the present invention uses a combination of at least two, and preferably more, of the following speech parameters: • Pause length average • Pause frequency • Short silences length average • Short silences frequency • Very short silences length average • Very short silences frequency • Equal pitch segments length average • Equal pitch segments frequency • Rising pitch segments length average • Rising pitch segments frequency • Falling pitch segments length average • Falling pitch segments frequency • Unclear pitch segments length average • Unclear pitch segments frequency

Pause segments are related here as relatively long silences (i.e., on the order of about 0.25-1.5 seconds) in speech - they usually appear between sentences. Short silences (on the order of about 0.15-0.25 seconds) and very short silences (on the order of about 0.03-0.15 seconds) are related here as silences that usually appear between words and between syllables. Equal speech segments are continuous 15 segments of speech that are characterized by relatively stable pitch (up to a tolerance). Rising and falling segments are characterized by a continuous and relatively clear (up to a tolerance) rising or falling trend of pitch. Unclear pitch segments are characterized by inconsistent or unclear pitch values. 5 Determination of emotional arousal requires that a combination of at least two, and preferably more, of the above parameters will simultaneously indicate a deviation from non-emotional values. The decision is also dependent on the degree of the deviation of each parameter. Ranges and values that characterize regularity for each parameter were determined by analysis of large samples of speech data 10 taken from the general population.

Referring now to Fig. 1, there is shown a block diagram illustration of an apparatus for detecting emotional arousal constructed and operative in accordance with one embodiment of the present invention. The apparatus includes a voice input unit 10, a voice sampling and digitizing unit 12, a pre-processing unit 14, a main 15 processing unit 16, and a main indicators output unit 18. Voice input unit 10 can be any device that carries human voice data in any form - microphone, wired telephone, wireless or cellular telephone, any audio-player device (such as tape-recorder, compact-disc), digitized voice files, Internet connection (voice over IP, cable, satellite or any other method). Voice sampling and digitizing unit 12 can be a 20 computer sound card, a specific DSP chip or any other sampling and digitizing device.

The emotional arousal determination procedure, according to the present invention, is as follows (with some variations between different embodiments of the apparatus) The flow chart in Fig. 2 details the pre-processing stage and the flow chart 25 in Fig. 3 details the main processing stage. 1. Pre-processing - The pre-processing function is to prepare the raw data for the processing itself. The processor is a CPU unit, such as a PC’s CPU, a specific DSP chip or any other processing device. Pre-processing includes the following 16 processing steps, in order to obtain pitch and amplitude parameters per cell of pre-defined length (Fig. 2): • Obtaining digitized voice samples (block 20) • Separation of the two voice channels’ samples (when required, for example, 5 when the input is sampled from a telephone conversation) (block 22) • Normalization of the samples’ values - performed for both channels simultaneously (block 24) • Data filtering - performed for both channels simultaneously (block 26) • Noise-reduction - performed for both channels simultaneously (block 28) 10 · Initiation of segmentation and basic parameters calculation for the first channel (block 30) • Silence and speech segmentation and dividing the speech segments into cells (block 32) are performed for. the first channel. • Auto-correlation of basic voice parameters (block 34) to calculate pitch and 15 amplitude is performed for the first channel. • The same steps (segmentation and auto-correlation) are now performed for the second voice channel, if one is present (blocks 36 and 38).

The results of the pre-processing are consecutive cells with pitch and amplitude values per cell. 20 2. The main processing procedure calculates the emotional arousal values. It can be performed on the same CPU processor where the pre-processing was performed or on a different CPU unit. It can be a PC’s CPU or a specific DSP chip or any other processing device. It includes the following processing steps, for each channel (Fig. 3): 25 · Marking a speech segment of a pre-defined number of seconds (the time frame) for processing (block 40). 17 • Calculating pauses related parameters (block 42). • Calculating silences related parameters (block 43). • Determining constant pitch segments (block 44). • Determining pitch up trends, pitch down trends and unclear pitch trends(block 46). • Calculating the pitch extracted voice parameters within the speech segment -number, lengths and averages of increasing, decreasing, stable and unclear pitch periods, (block 47) • Classifying the speech segment into one of several categories of typical parameter range (in order to differentiate segments with different speech magnitudes) (block 48). • Determining the emotional arousal level of the speech segment - based on comparison of the extracted voice parameters to pre-defined values that represent non-emotional speech and scoring the combination of irregularities (block 50). • If a second channel is present, i.e., taking specimens from two participants in a conversation (blocks 52 and 54), the same processing steps 40 to 50 are carried out on the sample from the second channel. • Preferably, the method also includes calculating a reliability grade - based on a combination of several factors: quality of the voice segment (noise level, size of sampled data, quality of sampled data), significance of the emotional arousal decision (number of participating parameters, degree of deviation of each parameter), consistency of the specific segment results with the previous speech segments (Emotional changes should follow reasonable patterns regarding the number of changes, their intensity, their length and switching between emotions in a given period of time). • Pattern processing - The processing may include another layer, which detects certain emotional patterns that are revealed along the time axis or when compared 18 to other results. For example, when analyzing a conversation, comparing the emotional states of the two speakers enables to detect patterns of the interpersonal communication - attachment, detachment, politeness, and conversation atmosphere. 3. Output - The emotion measurement results can be sent to various outputs in accordance with the specific apparatus and the specific application. Normally, the output will be sent to a user’s real time display (visual, vocal, or textual). It may be reported to a remote user through any kind of networking and it may be logged or stored to any sort of output or storage device or file.

Two basic apparatuses are presented for the patent implementation, although any other suitable apparatus can alternatively be employed: 1. A small, mobile, DSP chip based unit. This apparatus can serve as a small mobile unit for emotional arousal detection in real-time or offline analysis. It can be used as a stand-alone device in interpersonal face-to-face interactions. Alternatively, it can be connected to input or output devices such as computer, audio player, wired or wireless or cellular telephone, Internet or any other network, in order to obtain various local or remote voice inputs and to display or report to various local or remote outputs. It can also be integrated as hardware into other devices, such as wired, wireless or cellular telephones, computer games, toys, computers or any other electronic equipment. The apparatus includes a microphone (or any other input interface), digital sampler, processor and display (or any other output interface). 2. A software-based tool. This apparatus can serve as a computer-based tool for emotion arousal detection in real-time or offline analysis. It can be used as a stand-alone software tool for analysis of digitized voice files. Alternatively, it can be connected through the computer interfaces to any input/output device, in order to obtain any local or remote voice input, and display or report to various local or remote outputs.. Such devices may be a microphone, audio player, wired or 19 wireless or cellular telephone, Internet or any other network, other computers or any other electronic equipment. The software tool can also be integrated as a subsystem into another system. Such systems may be call/contact center software or hardware which monitors, records or analyzes conversations, various situation 5 and personal trainers or any monitoring, teaching or feedback system. The emotion software tool’s computer environment should include a microphone (or any other input interface), sampling and digitizing unit, processor, display (or any other output interface) and any other relevant external interface.

It will be appreciated that the present invention has a very wide range of 10 possible applications and implementations. A few of the possibilities are listed below, by way of example only. However, the use of the present invention is not limited to those applications described herein.

Emotion monitoring can be used to improve marketing, sales, service and relations with customers, especially in the call center environment. The emotion 15 monitoring, feedback and supervision of the service/sales interactions can be implemented in a real time environment, as well as in off line analysis. The monitoring can be implemented with both apparatuses described above: It can be integrated as a software tool into other call center products, such as recording tools, CRM (customer relation management) products, training tools or E-commerce 20 software. It can be installed as a stand-alone software tool in the call center, CRM or E-commerce environments and it can also be integrated into various hardware devices in these environments as a DSP chip unit. A small DSP chip unit can also be used as an independent small unit for monitoring face-to-face agent-customer interactions. 25 Emotion monitoring can be used to improve the training process of various professional personnel by improving awareness of emotional, as well as non-emotional, verbal patterns, as expressed in a speaker’s voice. In addition, the monitoring tool can be used for demonstration purposes (analyzing speech segments 20

<img img-format="tif" img-content="drawing" file="IL144818AD00027.tif" id="idf0007" />

of different emotions and different emotion expression patterns) and for training in controlling emotion expression (feedback of user’s emotions plus reward for altering emotion or expression pattern).

Emotional monitoring can be used as an assisting tool in various interpersonal 5 managerial tasks, such as interviewing, negotiating, meeting or just speaking on the telephone. Similarly, monitoring emotion is useful as an additional tool for psychological testing, and for diagnosis and treatment of specific illnesses, including psychiatric illnesses. This monitoring can be conducted during real time conversations, or in off line analysis (when the conversation is recorded), and it can 10 be operated in face to face or in remote interactions (i.e., telephone, Internet).

Advertising can also benefit from emotional monitoring, by adding significant value to the process of measuring and evaluating people’s attitudes in verbal questionnaires, focusing groups, and other methods.

Emotional monitoring can be used to aid in speech therapy and to increase 15 relaxation and to achieve more control over positive and negative emotional states. Altering the emotional state can be achieved either as a direct result of the increased awareness, or though a procedure similar to a biofeedback mechanism. One important application may be to assist the many programs aimed at reducing violent behavior among children and adults, where the monitoring can help to demonstrate 20 and to alter patterns of verbal anger.

The use of emotion monitoring can provide an added quality to computer and electronic games, both educational and recreational games. Emotion monitoring can also be part of toys and games that interact with a child and reflect to him his emotional state. 25 It can also be used to improve speech recognition in various applications, and to enhance the interaction between a computer or robot and its user, by permitting the electric device to respond to the emotional state of people around it. 21

Emotion monitoring can even be used as a tool for detecting some mental states which have distinctive voice characteristics, such as fatigue.

It will be appreciated that the invention is not limited to what has been described hereinabove merely by way of example. Rather, the invention is limited 5 solely by the claims which follow. 22