Synchronization Among Speakers Reduces Macroscopic Temporal Variability

A recent method for restricting inessential variation in speech is presented. Synchronous Speech is obtained by having two subjects read a prepared text in synchrony. Past results demonstrate that this is easy for subjects to do, and that some prosodic variability is greatly reduced when reading synchronously. Particular advantage has been found in the analysis of pauses and fundamental frequency variation, where synchronous speech has been demonstrated to exhibit markedly less inessential variability, thus furthering analysis and modeling. Here, duration ratios within a phrase are compared across synchronous and solo conditions. Variables associated with global timing and with the relationships between phrases are shown to be more consistent in the synchronous condition, while smaller units are not noticeably affected by the speaking condition. No systematic artifacts are found to be introduced by asking subjects to read in synchrony. A Method for Restricting Variability Synchronous Speech is obtained with the simple expedient of having two subjects read a prepared text together, with the minimal instruction to attempt to maintain synchrony (Cummins, 2002). The reason for constraining subjects in this manner is perhaps best appreciated by analogy with the difficult task of attempting to reconstruct a musical score, based only on a recording of a specific musician (Heijink et al., 2000). This task is interestingly similar to the work of the theoretically minded phonetician who attempts to uncover control and timing information, along with combinatorial units, from the continuous stream of speech. If one were faced with this task, it is worth considering which musician would give one more tractable data: the soloist, or the 14th violin player in the string section. Neither will reproduce the durations (or pitches) specified in their score exactly, of course, due to the inherent underspecification of the score. Both players will overlay some inherent biophysical noise, along with conventional timing variability, such as the predictable decellerando at the end of a phrase. The soloist will add additional complexity, however, in keeping with her role as the expressive focus in performance, making the inverse mapping from the recording to the score considerably more difficult. Now return to the position of the laboratory phonologist (or theoretical phonetician). An overarching goal is to deduce the units of control which relate to the linguistic message being uttered, and to uncover their mutual relations. This is not so different in kind from the above musical analog, though additional levels of complexity undoubtedly arise. Signal variability which is related to the linguistic content is relevant, while (for many purposes) one might like to find a way to reduce or exclude variability of paraor non-linguistic origin. The approach which I and colleagues have recently been following is to constrain the speaker to speak in time with another co-speaker. For this purpose, speakers read through a given text silently to familiarize themselves with it, and then commence reading together on a signal from the investigator. For many purposes, recording using near field head-mounted microphones onto the left and right channels of a single stereo file is sufficient to separate the two speakers while preserving the relative temporal alignment of speech events. We call speech collected in this manner Synchronous Speech, and both the task and the product have provided us with much food for thought (Cummins, 2001; Cummins, 2002; Cummins and Roy, 2001; Cummins, 2003). In this paper, I will summarize those findings which have best revealed the advantages of this novel method, then provide some new results which examine the variability of intervals below the whole phrase, and finally provide pointers to areas I believe might benefit from adoption of the method. Properties of Synchronous Speech Synchronizing with a co-speaker, without extensive practice, turns out to be simple for subjects to do (Cummins, 2002; Cummins, 2003). After reading through a simple text once, and being given a start signal, subjects typically manage to keep inter-speaker lags to average values of around 60 ms at phrase onsets, and 40 ms or less after the first syllable or so. Rather surprisingly, extensive practice at the task does not improve the degree of synchrony significantly (Cummins, 2003), although with repeated readings of the same text, and with the same co-speaker, a slight improvement may be detected. Visual contact with the co-speaker does seem to have a small beneficial effect on synchrony, even though subjects are typically attending to a read text in front of them (Cummins, 2003). In experiments done to date, speakers have not been carefully matched for familiarity, intrinsic speaking rate or volume. Among the heterogeneous pairs of speakers we have studied to date, most appear to be collaborating, producing speech at a relatively slow rate (but faster than some of the slowest speakers’ natural reading rate). We have not yet (in over 60 pairs of speakers) found a speaking pair in which one speaker consistently lagged behind the other. Rather, they seem to genuinely speak together, with a high degree of synchrony.