A classical singing performance occurring in different rooms is likely to vary for different reasons. This study investigates to which extent this variation is due to different acoustic conditions. To analyse the performance of four singers rendering four musical pieces in eight different rooms, room acoustical parameters were used to predict musical performance features extracted from recordings based on linear mixed-effects models. Considering the common behaviour of all singers, only a small proportion of the variance in performance can be explained. Instead, rather individual patterns indicate that each singer developed a specific strategy of adaptation to the varying acoustic environment.
1. Introduction
A singer performing a given musical piece in different venues does not repeat exactly the same performance. Changes in the interpretation may occur due to multiple factors related to the singer and the environment, such as the mood and physical state of the singer or the visual and auditory appearance of the room (Parncutt and McPherson, 2002).
Previous research has investigated the influence of room acoustics on the musical performance mostly for instrumentalist musicians. For a solo instrumentalist who performed a largely identical musical program on a concert tour through different concert venues, for example, more than 50% of the variance of a set of performance features could be explained by room acoustical parameters (Schärer Kalkandjiev and Weinzierl, 2013). In another study investigating the reaction of instrumentalist musicians and a singer to various virtual acoustical environments rendered through loudspeakers in an anechoic chamber, it was found that the singer reduced the loudness of his voice with increasing sound strength of the room response (G), decreased the intensity of vibrato with increasing reverberation time (RT), and lowered the harmonic strength with increasing RT and G (Kato et al., 2015).
Other research (Ternström, 1989) focused on choir singing by analysing the performance of three choirs in three different rooms. Results suggested that room acoustics could influence both the long-time average spectrum and the sound power of the recorded voices. In particular, it was observed that the singers were louder in the less reverberant rooms.
The present paper investigates the extent to which variations between singing performances can be explained by room acoustical conditions, analysing how both common and individual patterns are used within this adaptation process.
2. Experiment
2.1 Measurements
The experiment took place in six rooms, two of which were equipped with banners to modify the RT, resulting in a wide range of eight acoustical environments from a dry recording studio up to a church. In each room, impulse responses were measured on stage at the position of the singer using, on the one hand, a dodecahedral source (Norsonic Nor276 and amplifier Nor280, Norway) and an omnidirectional microphone (NTI Audio M2230, Liechtenstein), and on the other hand, a directional broadband loudspeaker (Fostex 6301B, Japan) placed just before the mouth of an artificial head (Neumann KU100, Germany) facing the audience.
The singers were recorded using a professional audio interface (RME Fireface UFX, Germany) and a close field microphone (DPA 4060, Denmark) placed 3 cm from the edge of their mouth. Each singer sung a cappella excerpts of the three musical pieces (Table 1). They were asked to choose pieces they were familiar with in order to reduce variability of the performance which might be due to a weak knowledge of the piece. These musical excerpts had various lengths, ranging from 15 s to 1.5 min. As the singers had different vocal ranges, the pieces were different, except for the last one, where they were asked to sing a version of “Happy birthday” in lyrical style and in their own vocal register to obtain data common to all singers. To prevent the singers from over-adapting to the different rooms, they were not informed of the final goal of the study. They were simply asked to perform their musical program as if they were in front of an audience, in the empty room. They were not asked to pay special attention to the room response.
Singers and musical pieces involved in the experiment.
Voice type . | Age . | Musical pieces . | Pace . |
---|---|---|---|
W. A. Mozart - Ach ich liebte (Aria from Entführung aus dem Serail) | Fast | ||
Soprano | 23 | G. Puccini - Aria from Manon Lescaut | Medium |
G. Verdi - Aria from Rigoletto | Slow | ||
W. A. Mozart - Aria from Le nozze di Figaro | Fast | ||
Mezzo-soprano | 27 | J. S. Bach - Esurientes implevit bonis (Aria from Magnificat) | Medium |
F. Schubert - Nur wer die Sehnsucht kennt | Slow | ||
G. F. Händel - Aria from Tamerlano | Fast | ||
Tenor | 21 | F. Schubert - Ganymed | Medium |
A. Cesti - Intorno all idol mio | Slow | ||
G. F. Händel--Honor and arms (Aria from Samson) | Fast | ||
Baritone | 26 | E. Humperdinck - Aria from Hänsel und Gretel | Medium |
G. Fauré - Au cimetière | Slow | ||
All singers | Happy Birthday (in a lyrical style) | Medium |
Voice type . | Age . | Musical pieces . | Pace . |
---|---|---|---|
W. A. Mozart - Ach ich liebte (Aria from Entführung aus dem Serail) | Fast | ||
Soprano | 23 | G. Puccini - Aria from Manon Lescaut | Medium |
G. Verdi - Aria from Rigoletto | Slow | ||
W. A. Mozart - Aria from Le nozze di Figaro | Fast | ||
Mezzo-soprano | 27 | J. S. Bach - Esurientes implevit bonis (Aria from Magnificat) | Medium |
F. Schubert - Nur wer die Sehnsucht kennt | Slow | ||
G. F. Händel - Aria from Tamerlano | Fast | ||
Tenor | 21 | F. Schubert - Ganymed | Medium |
A. Cesti - Intorno all idol mio | Slow | ||
G. F. Händel--Honor and arms (Aria from Samson) | Fast | ||
Baritone | 26 | E. Humperdinck - Aria from Hänsel und Gretel | Medium |
G. Fauré - Au cimetière | Slow | ||
All singers | Happy Birthday (in a lyrical style) | Medium |
2.2 Room acoustical parameters and musical features
Room acoustical parameters were calculated from the measured impulse responses according to the dedicated standard (ISO 3382-1, 2009) by means of the ITA-toolbox (Berzborn et al., 2017). Based on clusters of acoustical parameters correlated with certain perceptive dimensions described in previous research (Lokki et al., 2012), four acoustical parameters were selected for the present study, corresponding to four different dimensions, as shown in Table 2. In addition, the Speech transmission index (STI), considered as a measure of speech intelligibility, was estimated from the impulse responses (IEC 60268-16, 2011) measured by means of the directional speaker and the artificial head facing the audience, within an average background noise level on the order of 35 dB. It is noted that these room acoustical parameters, except for , were developed to describe characteristics of the sound field received in the audience. Although the proportion of direct sound is much stronger when measured on stage, with the receiver being only 1 m away from the source, considerable variations across rooms could still be observed, reflecting the diversity of acoustic environments the singers were exposed to.
Six musical performance venues representing eight different room acoustical conditions. Five room acoustical parameters, all measured on stage, were selected for the current analysis: Early decay time (EDT), STI, BR, Late interaural cross-correlation (IACClate), and Early stage support (STearly). All parameters were averaged over the 500 Hz and the 1 kHz octave bands, except for IACClate, which was averaged over the 125 Hz to 4 kHz octave bands, and over the 250 Hz to 2 kHz octave bands according to ISO 3382-1 (2009). Abbreviations w/ & w/o stand for “with” & “without.”
Room name . | Volume (m3) . | EDT (s) . | STI . | BR . | . | (dB) . |
---|---|---|---|---|---|---|
Recording studio (w/o banners) | 420 | 0.7 | 0.78 | 1.18 | 0.41 | −5.5 |
Recording studio (w/ banners) | 420 | 0.5 | 0.86 | 1.55 | 0.42 | −7.1 |
Kammersaal | 590 | 1.0 | 0.77 | 1.17 | 0.48 | −4.9 |
Cabaret theater Distel | 1700 | 0.5 | 0.90 | 1.23 | 0.42 | −9.3 |
Joseph Joachim Saal (w/o banners) | 3660 | 1.4 | 0.85 | 0.94 | 0.55 | −11.1 |
Joseph Joachim Saal (w/ banners) | 3660 | 0.8 | 0.89 | 1.21 | 0.62 | −11.4 |
St Eduard Church | 9360 | 4.0 | 0.77 | 0.97 | 0.43 | −13.9 |
Philharmonic Berlin | 22 000 | 1.2 | 0.96 | 0.88 | 0.40 | −17.6 |
Perceptual dimension | Reverberance | Voice clarity | Bassiness | Envelopment | Loudness |
Room name . | Volume (m3) . | EDT (s) . | STI . | BR . | . | (dB) . |
---|---|---|---|---|---|---|
Recording studio (w/o banners) | 420 | 0.7 | 0.78 | 1.18 | 0.41 | −5.5 |
Recording studio (w/ banners) | 420 | 0.5 | 0.86 | 1.55 | 0.42 | −7.1 |
Kammersaal | 590 | 1.0 | 0.77 | 1.17 | 0.48 | −4.9 |
Cabaret theater Distel | 1700 | 0.5 | 0.90 | 1.23 | 0.42 | −9.3 |
Joseph Joachim Saal (w/o banners) | 3660 | 1.4 | 0.85 | 0.94 | 0.55 | −11.1 |
Joseph Joachim Saal (w/ banners) | 3660 | 0.8 | 0.89 | 1.21 | 0.62 | −11.4 |
St Eduard Church | 9360 | 4.0 | 0.77 | 0.97 | 0.43 | −13.9 |
Philharmonic Berlin | 22 000 | 1.2 | 0.96 | 0.88 | 0.40 | −17.6 |
Perceptual dimension | Reverberance | Voice clarity | Bassiness | Envelopment | Loudness |
The cross-correlation within the set of room acoustical parameters over the rooms was never higher than 0.67 (Table 3). Thus, collinearity in the statistical analysis was not considered problematic.
Cross-correlation of the room acoustical parameters over the eight rooms.
. | . | EDT . | . | BR . | STI . |
---|---|---|---|---|---|
1.00 | |||||
EDT | −0.07 | 1.00 | |||
−0.02 | −0.47 | 1.00 | |||
BR | −0.06 | −0.53 | 0.67 | 1.00 | |
STI | 0.05 | −0.38 | −0.62 | −0.13 | 1.00 |
. | . | EDT . | . | BR . | STI . |
---|---|---|---|---|---|
1.00 | |||||
EDT | −0.07 | 1.00 | |||
−0.02 | −0.47 | 1.00 | |||
BR | −0.06 | −0.53 | 0.67 | 1.00 | |
STI | 0.05 | −0.38 | −0.62 | −0.13 | 1.00 |
The singing recordings were first analysed by means of onset detection of each note, by comparing a MIDI score with the corresponding recording using a dynamic time warping algorithm (Lerch, 2008). Manual correction was required due to the long attack time of singing voice, to ensure all onsets were correctly detected. A collection of low-level audio features [inter-onset interval, sound level, spectral centroid, MFCCs (Mel-Frequency Cepstral Coefficients), etc.] was estimated between two consecutive onsets. Linear combinations of these features yielded higher-level features that convey a larger extent of musical meaning. This methodology has been exposed in more detail in a previous paper (Luizard et al., 2019). The musical features that characterise the performances constituted four groups related to temporal (tempo, agogic, rhythmisation), dynamical (loudness, dynamics), timbral (hardness, brightness, fullness, bandwidth) performance properties, and a combination thereof (articulation, expression).
2.3 Statistical analysis
With four singers singing four pieces in each of the eight rooms, the data are structured in different levels of variance. Linear mixed-effects models (LMMs) are particularly suited to analyse such a hierarchical data structure (Hox et al., 2017). The models were computed for each musical feature, where both fixed effects (the room acoustical parameters) and random effects (the musical pieces) could be simultaneously taken into account. All features were z-transformed prior to the analysis so that their relative predictive contributions could be derived from the regression estimates.
3. Results
As a first step, intraclass correlation coefficients (ICCs) based on random intercept only models were computed to estimate the relative contribution of the respective random effects in the explanation of the overall variance in the data. In contrast to the variance attributable to performers () and musical pieces (), the overall variance due to the rooms () was small. This confirms that the variance of performance between different singers and different musical pieces is larger than the variance due to a given singer and a given piece performed in different rooms, as expected. However, the variance caused by the different room acoustical conditions as the core subject of the present study was investigated in more detail in the next steps, accounting for the side effects of the factors Singer and Piece. Then linear mixed-effect models were computed for each musical feature, including the selected room parameters as independent variables (fixed effects) and the singer ID and piece ID as random intercepts. Results revealed no significant relationship between the musical features and the room acoustical parameters. The , representing the amount of variance explained by the fixed effects (Nakagawa and Schielzeth, 2013), indicated that, considering the common behaviour of all singers, the variance explained by the room acoustical parameters was low, with a mean across musical features of 1.5% and a maximum value of 5%.
In order to consider not the common but the individual behaviour of all singers, models were calculated using only the piece ID as random intercept. The averaged across singers, as depicted in Fig. 1, showed much higher values than the models generalising across singers, with a mean across musical features of 11% and a maximum of 20%. In addition, high variations across singers were observed regarding the values for different performance features. For Singer 2, for example, 67% of the variance in agogic could be explained by room acoustical parameters, whereas the values were below 4% for all other singers. The significant relations obtained between musical features and room acoustical parameters (see Fig. 2) showed different patterns among the four singers. Considering the effect of different room acoustical parameters, Bass ratio (BR) was involved in most significant relations, followed by . On the dependent variable side, most of the musical features which could be significantly predicted were related to tone colour and sound level, with Hardness and Dynamics most often involved. The direction of the effect, however, was by no means uniform for the different singers: While Singers 1 and 4, for example, showed lower levels of Hardness in their singing with increasing BR of reverberation, Singer 3 has acted contrary to this, which again underlines the individual adaptation patterns of the different musicians.
(Color online) Variance in the musical features explained by room acoustical parameters () in several LMMs computed separately for each singer.
(Color online) Variance in the musical features explained by room acoustical parameters () in several LMMs computed separately for each singer.
Number of significant relations between the room acoustical parameters and the musical features for each singer (top left), room parameter (top right), and musical feature (bottom).
Number of significant relations between the room acoustical parameters and the musical features for each singer (top left), room parameter (top right), and musical feature (bottom).
A more detailed view on these results is presented in Fig. 3, showing the sign of the regression estimates of the significant predictors for each model. It can be seen that Singers 1 and 2 used mainly temporal and dynamic features, whereas Singer 3 used timbral performance features to adapt to the different acoustic conditions, and Singer 4 used a mix of all aspects.
(Color online) Regression slopes of the significant relations between the room acoustical parameters and the musical features based on the LMM analysis for each singer.
(Color online) Regression slopes of the significant relations between the room acoustical parameters and the musical features based on the LMM analysis for each singer.
4. Discussion and conclusion
The current field study has analysed the recorded performances of four musical pieces sung by four singers in eight rooms, and related performance features to the acoustical parameters of these rooms by means of linear mixed models.
No common pattern of adaptation to room acoustics was observed when considering all singers together. Accounting for the singers individually, however, revealed several significant relations, suggesting that each singer developed a specific strategy to adapt to changing room acoustical conditions. Since we did not test several singers for each vocal register, the collected data does not allow us to decide whether the adaption strategies found are specific for the individual singers or for the different vocal registers, although we consider the latter less plausible.
These patterns mostly included an adaptation of loudness and timbre (measured by Dynamics and Hardness), as already observed in a previous study (Ternström, 1989), while only little variance in tempo could be explained by the room acoustical parameters. This is in accordance with the previous comparison of a singer to instrumentalists (Kato et al., 2015), the latter having generally adapted to room acoustics in terms of tempo, in agreement with other research (Schärer Kalkandjiev and Weinzierl, 2013, 2015b).
Also on the side of room acoustics as the influencing factor, most parameters involved in significant relations were related to sound level and tonal colour, as measured by stage support () and BR.
Looking at the slope of the regression coefficients shows not only different, but sometimes even opposite relations. While Singers 1 and 4 reacted with reduced Hardness to an increase in BR, i.e., they adapted the timbre of their voice to the timbre of the room, Singer 3 showed an opposite behaviour, i.e., he tried to compensate the effect of the room with his vocal intonation. A similar grouping can be seen when considering the relation between acoustic envelopment, as measured by and the vocal timbre. While Singers 1 and 4 reacted with increased Hardness to an increase in Envelopment, Singer 3 again showed the opposite behaviour.
In total, 11% of the variance in performance could be explained by the acoustics of the room, on average across singers. The extent of the adaptation and the number of significant relations between acoustic and performance features varied between the singers, ranging from 7 significant relations and a mean of 5% over all musical features for Singer 1 to 21 significant relations and a mean of 15% over all musical features for Singer 3. These values are comparable to a previous study involving 12 instrumentalists (Schärer Kalkandjiev and Weinzierl, 2015a), where the mean explained variance was slightly lower with R2 = 8%.
The results are not only meaningful with regard to the extent and manner in which classical singers are influenced by the acoustics of their performance spaces; they also show which acoustic properties are most relevant for this influence and deserve special attention in the design of stage acoustics in these venues.
Acknowledgments
This work was supported by a grant of the Alexander von Humboldt Foundation. The authors would like to thank the singers for their participation, the room managers for their collaboration, and Erik Brauer for his support in the recording sessions.