The purpose of this article is to quantitatively determine the effects of emotional distortions (in particular, those of anger) on the long term spectra (LTS) analysis used for speaker identification (SPID). (For authoritative SPID review articles, see e.g. Hollien 2013, 2016.) This is performed through a careful and replicable methodology.
The objective of the SPID process is to identify an unknown speaker through voice analysis, usually under conditions which are not ideal. The process often involves comparing one unknown voice against the voice of a known speaker (and thus, emotional state differences between recordings could be problematic). One of the fundamental challenges that SPID faces is the determination of whether or not intraspeaker variability is smaller than interspeaker variability, and of how this relationship holds for different conditions (Hollien, 2002). These conditions include distortions from varied sources, such as technological distortions due to the equipment used for the recordings, and environmental distortions caused by noise or harsh sounds in the background. In particular, speakers may often be the source of distortion themselves, as a variety of feelings including fear, anger, and anxiety may be present (a situation likely to occur, for example, in forensic SPID when the speaker might be committing a crime). These emotions trigger a change in speech production that reveals itself as a shift in the values of the measured signal parameters (such as frequencies and speech speed) (Williams and Stevens, 1972; Banse and Scherer, 1996; Johnstone, 2001). As voice production consists of air pulses caused by the vibration of the vocal folds (which are then modified by the supralaryngeal vocal tract), the dominant factors of vocalization are the respiration patterns and the varying tension of the muscles involved in the process, and these will most likely modify both the fundamental frequency produced and the position of the formants. Since respiration and muscle tone are highly correlated to emotions, it is therefore very likely that these changes in tension are detectable in the acoustic sound wave (Scherer, 1986).
This topic, however, has not been extensively studied largely because there are ethical and methodological constraints that prevent the induction of controlled high magnitude emotions. Current consensus (Johnstone, 2001) is that standard controlled laboratory conditions are feasible only for low-intensity emotional responses, for which it is difficult to observe appreciable changes in SPID effectiveness. Another difficulty is how to induce the correct kind of emotion. Martin (1990)
gives an overview on some of the possible techniques used to induce emotions, including emotional music, pictures, and self-generated techniques such as imagination and memories. These methods are classified according to how the emotions are produced. For the purpose of this article, the autobiographical recall approach was selected. In this technique, subjects are asked to remember mood-evoking events in order to generate the desired emotion. While this is not the only method applicable to the problem, it was selected for its simplicity and because it allowed the subjects to have privacy while recording. Furthermore, experiment blindness was assured by having the participants themselves assessing their level of anger, as described below.
The remainder of this section discusses the workings of the LTS SPID process. Although there are many different markers (also called “vectors”) that help to distinguish two recordings from different speakers, ever since the pioneering work ofHollien and Majewski (1977), one of the most common methods used in SPID is LTS analysis, also known as long term average
. The LTS analysis evinces the time-averaged quality (timbre) of the voice, an acoustic feature that allows a person to distinguish between, for example, a clarinet and a violin playing the same note (frequency) with the same intensity. The distribution is obtained computing the Fourier transform of the signal over a long period of time, say, 30 seconds; the idea is that the speaker has enough time to move through all the sound phase space.
This vector has been widely studied in terms of how efficiently identification is performed and has been found to be one of the most reliable ones, mainly because it continues to function in the presence of noise and limited passband (Hollien, 2002).
One of the challenges of using this vector is how to define the correlation between two spectra. There are several approaches to this problem, including assigning a specific number to each LTS set (according to some algorithm), or even visual inspection of the plots. For the purposes of this article, a more sophisticated approach was required, as the sought differences turned out to be subtle
. Two correlation coefficients were considered for this purpose: the Standard Deviation of the Differences Distribution (SDDD)(Harmegnies, 1988) and the Bravais-Pearson cross-correlation coefficient, (Stanton, 2001). Exploratory experiments performed by our group without the anger component (Ortega-Rodríguez et al., 2015) showed that the Bravais-Pearson correlation coefficient gives the best results for this line of research and was therefore chosen.
In the Bravais-Pearson approach, a spectrum in the LTS analysis is considered to be a -dimensional vector, with frequency channels. Thus, the spectrum may be defined as
where is the level of the -th frequency component (Harmegnies, 1988). In this context, the coefficient measures how related two LTS samples are. is defined as:
where stands for the mean of each spectrum, and are the respective standard deviations. The Bravais-Pearson coefficient provides several advantages. Not only does it have a high discriminating ability, but also it is independent of the overall differences in the intensities of the two spectra (Harmegnies, 1988). This allows for comparison of recordings that were conducted under different microphone placement or environmental conditions.
Most of the studies on the relationship between speech and emotions have focused on the ability to distinguish between different emotional states of a speaker (e.g. anger, sorrow, fear) through signal analysis (Williams and Stevens, 1972; Fuller Jr, 1972; Scherer, 1986; Johnstone, 2001; Harnsberger et al., 2009). In particular, LTS analysis has been used to attempt to identify emotional states and depression (Pittam, 1987), sometimes making use however of human filtering (perception tests) in the research process (Banse and Scherer, 1996).
Less common is research on how emotions or disguise affect SPID. Rodman and Powell (2000) recommend and plan out research to study the effects of disguise on SPID, although they do not carry it out. More related to our article, Hollien and Majewski (1977) studied the effects of stress (induced via electroshocks on the subjects) on the LTS of a signal, but they did not find a significant deviation from adequate SPID. Other than this, to the best of the authors’ knowledge, there is no study on the effect of emotions on LTS SPID.
2 Materials and methods
2.1 Recorded subjects
The SPID problem has many variables. Consider, for example, the gender, geographical origin, and age of the subjects (Hertrich and Ziegelmayer, 1987; Linville, 2002; Hollien and Majewski, 1977; Pittam, 1987; Yüksel and Gündüz, 2017). For this reason, and in order to work in a smaller phase space in this exploratory endeavor, subjects with similar characteristics were chosen, in order to highlight the effects of emotion and reduce the overall complexity of the problem. The subjects met the following restrictions: they are males in the age range 18–25, and their geographical origin is Costa Rica’s Central Valley Metropolitan Area (this region consists of the four most populous cities of the country, located in the country’s densely populated mid-region). Subjects also attended primary school in this region. Variations in speech would presumably be less marked under these conditions.
2.2 Recording conditions
Recording conditions were standardized to ensure homogeneous results. Each of the speech samples was collected in a private location (which was not, however, the same for all participants) where the subjects would feel comfortable. Additionally, the speech was required to be fluid and not scripted, and top of the range smartphone microphones were used, such as the iPhone and Samsung Galaxy series.
According to the emotional state, there were two types of recordings: Normal recordings, in which subjects were asked to speak about their day-to-day life, in order to have as little emotional response to the topic as possible; angry recordings, in which subjects were asked to invoke an angry state by means of describing a situation that made them angry. The interviewer was not present in order to avoid inhibiting the subjects.
The sample length was 45 and 60 seconds for the normal and angry cases, respectively (extra time for the angry cases allowed for some transition time for anger to presumably build up).
32 subjects who met the selection criteria described in Section 2.2 were interviewed. This number was chosen so as to have a significant sample (National Institute of Standards and Technology, 2012) of the Costa Rican Central Valley Metropolitan Area population. Each interview consisted of three recordings. For 16 of the 32 interviews, the following recording order was used: normal-normal-angry. For the other 16 interviews, we used instead a normal-angry-normal order. The motivation behind having two types of order was to reduce the effects of possible systematic errors due to sampling order.
The subjects were taken to a private place where a script was read to them. The interviewers explained that this was a research project for the Universidad de Costa Rica and that the contents of the recordings would not be used or heard. It was emphasized that only the acoustic parameters of the samples were to be used and that the level of anger would be only determined by the subjects’ self-rating at the end of the recording process (a characteristic that also provided blindness to the methodology). Furthermore, the nature of the project was described only in general terms to avoid affecting the natural outcome of the recordings.
In the case of normal recordings, the speakers were instructed to talk for 45 seconds about their life (for example, their day, their pet, a recent event, or their interests). In the case of anger recordings, the subjects were asked to invoke a state of anger by speaking for one minute about something that made them angry, for example, a person they hated or some event that especially bothered them. The subjects were instructed not to fake the emotion (for example, by forcing it with shouts). They were left alone with the recorder and were told to speak whenever they felt ready.
Once the anger recording was completed, the participants were asked to rank the anger they had during the recording in a scale from 1 to 5, where 1 meant “not able to get angry,” 3 meant “moderately angry,” and 5 meant “furious.”
Note that when the first order was used, subjects did not know about the emotional-response part until they had to do the anger recording, whereas, in the case of the normal-angry-normal order, the speaker knew about this aspect of the research during his last normal recording. As mentioned, both orders were used and averaged so as to reduce order bias.
3 Data processing
The processing in which the present research is based was developed by our group (Ortega-Rodríguez et al., 2015) for SPID (without the anger component) and has been extensively tested and optimized for best identification results.
The data processing corresponding to the present paper can be summarized as follows. Thirty-second segments were obtained from the original recordings. In the case of the anger samples, this was particularly important to filter out the possible transition from normal to angry state. The Audio processing software Audacity 2.1.0 (Audacity Team, 2015) was used to obtain the Fast Fourier Transform (FFT) for each recording. A Hanning window was used with 4096 frequencies sampled. This parameter was empirically determined to render the best results in our previous exploratory studies. Each FFT was stored as a text file for further processing.
Next, a C++ code computed the Bravais-Pearson correlation coefficient for the samples. Finally, the average, the sample standard deviation (SD) and the standard error of the mean (SE) for the correlation coefficients were obtained for the cases of normal-normal-angry and normal-angry-normal recordings.
Table 1 displays the results of the process described in the previous section.
|Standard error of the mean||0.005||0.005|
32 subjects were interviewed and the Bravais-Pearson correlation coefficient was calculated to obtain the normal-to-normal correlation and the normal-to-anger correlation. There is a noticeable difference between the two obtained averages. To understand how significant this difference actually is, it is useful to compare it to the results of the aforementioned research performed by our group (Ortega-Rodríguez et al., 2015), which is a simplified version of the process described in the present paper (as the anger element was absent), although the experimental conditions were the same. There, the average correlation coefficient between two different speakers was found to be 0.890 with an SE of 0.010, whereas same-speaker correlation had a mean coefficient of 0.955 with an SE of 0.005.
Comparing these three correlations: 0.950 (normal-normal, same speaker), 0.934 (normal-angry, same speaker), and 0.890 (different speakers, both normal), we see that the effects of anger deviate the signal (from what would be otherwise expected) a significant 33% of the difference in correlation between the same speaker under normal circumstances and two different speakers under normal circumstances. This is remarkable since the average self-reported anger was 2.9 in the 1-5 scale described in Section 2.3, meaning that on average the subjects were only moderately angry. The distribution of the level of anger for our subjects is presented in Table 2
|Number of subjects||Self-reported anger level|
|Standard error of the mean||0.12|
Higher anger ratings were expected to generate larger effects on the LTS SPID. To test this hypothesis, we selected only the five angriest subjects (those who self-rated 4 on the 1–5 scale). Table 3 shows the results obtained for this case. For strong anger, the deviation from normality is close to 50% of the shift to a different speaker. This shows that the effects of anger on the LTS SPID do indeed grow as anger increases.
|Standard error of the mean||0.013||0.015|
Even though some authors praise LTS SPID for being robust to speaker stress (Hollien and Majewski, 1977), we have found that there is, in fact, a significant distortion on the human voice due to anger for LTS SPID considerations. Even as the emotional response obtained from the participants was only moderate, we found an appreciable difference in the correlation coefficients between the cases of normal-normal recordings and normal-anger recordings. Moderate anger offsets SPID results by a significant 33% in the direction of a different speaker. Furthermore, these results were obtained with a method which can be fully automatized, providing an objective approach independent of human perception errors. The method also avoids assessing the sincerity of the participants, and is therefore in agreement with the code of practice of the International Association for Forensic Phonetics and Acoustics (The International Association for Forensic Phonetics and Acoustics, 2004).
Our results are relevant for forensic research as LTS analysis has been traditionally considered a robust vector in SPID, especially since it is not sensitive to changes in the intensity of the speech sample, works well for short recordings, and continues to function in the presence of noise and limited passband. The results of the present paper, however, indicate that one should be cautious when using LTS SPID to calculate likelihood ratios in a scenario of anger, even if this anger is not strong.
As other emotions might also affect significantly the effectiveness of LTS SPID, they are also worthy of future automated study. The study could also be performed on women, or on speakers of a different language.
This work was supported by grant 805-B2-175 of the Universidad de Costa Rica’s Vicerrectoría de Investigación and the CIGEFI. The authors would also like to acknowledge the useful comments of two Speech Communication anonymous reviewers.
- Audacity Team (2015) Audacity Team (2015). Audacity (version 2.1.0), audio editor and recorder.
- Banse and Scherer (1996) Banse, R. and Scherer, K. R. (1996). Acoustic profiles in vocal emotion expression. Journal of personality and social psychology, 70(3):614.
- Fuller Jr (1972) Fuller Jr, F. H. (1972). Detection of emotional stress by voice analysis. Technical report, Decision Control Inc Bethesda MD.
- Harmegnies (1988) Harmegnies, B. (1988). SDDD: A new dissimilarity index for the comparison of speech spectra. Pattern Recognition Letters, 8(3):153–158.
- Harnsberger et al. (2009) Harnsberger, J. D., Hollien, H., Martin, C. A., and Hollien, K. A. (2009). Stress and deception in speech: evaluating layered voice analysis. Journal of forensic sciences, 54(3):642–650.
- Hertrich and Ziegelmayer (1987) Hertrich, I. and Ziegelmayer, G. (1987). Sexual dimorphism in the long term speech spectrum. Human Evolution, 2(3):255–262.
- Hollien (2013) Hollien, H. (2013). Barriers to progress in speaker identification with comments on the trayvon martin case. Linguistic Evidence in Security, Law and Intelligence, 1(1):76–98.
- Hollien (2016) Hollien, H. (2016). An approach to speaker identification. Journal of forensic sciences, 61(2):334–344.
- Hollien and Majewski (1977) Hollien, H. and Majewski, W. (1977). Speaker identification by long-term spectra under normal and distorted speech conditions. The Journal of the Acoustical Society of America, 62(4):975–980.
- Hollien (2002) Hollien, H. F. (2002). Forensic voice identification. Academic Press.
- Johnstone (2001) Johnstone, T. (2001). The effect of emotion on voice production and speech acoustics. Retrieved from thesiscommons.org/qd6hz; accessed 15 August 2017.
- Kinnunen et al. (2006) Kinnunen, T., Hautamäki, V., and Fränti, P. (2006). On the use of long-term average spectrum in automatic speaker recognition. In 5th Internat. Symposium on Chinese Spoken Language Processing (ISCSLP’06), Singapore, pages 559–567.
- Linville (2002) Linville, S. E. (2002). Source characteristics of aged voice assessed from long-term average spectra. Journal of Voice, 16(4):472–479.
- Martin (1990) Martin, M. (1990). On the induction of mood. Clinical Psychology Review, 10(6):669–697.
- National Institute of Standards and Technology (2012) National Institute of Standards and Technology (2012). Sample sizes required, Engineering Statistics Handbook. http://www.itl.nist.gov/div898/handbook/prc/section2/prc222.htm. Online; accessed 16 November 2017.
- Ortega-Rodríguez et al. (2015) Ortega-Rodríguez, M., Solís-Sánchez, H., Valverde-Méndez, D., Venegas-Li, A., Agüero Botero, C., and González, A. (2015). Forensic Voice ID. Paper presented at Mini-Congreso, November 2015, Centro de Investigaciones Geofísicas, Universidad de Costa Rica.
- Pittam (1987) Pittam, J. (1987). The long-term spectral measurement of voice quality as a social and personality marker: a review. Language and Speech, 30(1):1–12.
- Rodman and Powell (2000) Rodman, R. and Powell, M. (2000). Computer recognition of speakers who disguise their voice. In The international conference on signal processing applications and technology ICSPAT2000.
- Scherer (1986) Scherer, K. R. (1986). Voice, stress, and emotion. Dynamics of stress: Physiological, psychological and social perspectives, pages 157–179.
Stanton, J. M. (2001).
Galton, pearson, and the peas: A brief history of linear regression for statistics instructors.Journal of Statistics Education, 9(3):1–16.
- The International Association for Forensic Phonetics and Acoustics (2004) The International Association for Forensic Phonetics and Acoustics (2004). Code of Practice. https://www.iafpa.net/the-association/code-of-practice/. Online; accessed 6 January 2018.
- Williams and Stevens (1972) Williams, C. E. and Stevens, K. N. (1972). Emotions and speech: Some acoustical correlates. The Journal of the Acoustical Society of America, 52(4B):1238–1250.
- Yüksel and Gündüz (2017) Yüksel, M. and Gündüz, B. (2017). Long term average speech spectra of turkish. Logopedics Phoniatrics Vocology, pages 1–5.