Speed-accuracy tradeoffs in human speech production

Abstract

Speech motor actions are performed quickly, while simultaneously maintaining a high degree of accuracy. Are speed and accuracy in conflict during speech production? Speed-accuracy tradeoffs have been shown in many domains of human motor action, but have not been directly examined in the domain of speech production. The present work seeks evidence for Fitts' law, a rigorous formulation of this fundamental tradeoff, in speech articulation kinematics by analyzing USC-TIMIT, a real-time magnetic resonance imaging data set of speech production. A theoretical framework for considering Fitts' law with respect to models of speech motor control is elucidated. Methodological challenges in seeking relationships consistent with Fitts' law are addressed, including the operational definitions and measurement of key variables in real-time MRI data. Results suggest the presence of speed-accuracy tradeoffs for certain types of speech production actions, with wide variability across syllable position, and substantial variability also across subjects. Coda consonant targets immediately following the syllabic nucleus show the strongest evidence of this tradeoff, with correlations as high as 0.72 between speed and accuracy. A discussion is provided concerning the potentially limited applicability of Fitts' law in the context of speech production, as well as the theoretical context for interpreting the results.

Fig 1. Schematic representation of the Task Dynamics framework.

The variable X is the displacement of the controlled variable in task space and X₀ is the target. The Forward Dynamics component implements a second-order dynamical system, conforming to Eq 3, that transforms (via inverse kinematics) the error signal, ΔX, into the second derivative of the articulator-space variable u. The integrals $\dot{u}$ and u function as motor commands to the Plant, or speech production apparatus.

Fig 2. Schematic representation of the VITE neural model [13].

Note the many similarities of this structure to that of Task Dynamics in Fig 1. TPC is a representation of the target position, which produces a target position X₀. The DV population compares the target to the system’s current position, and computes the task-space dynamics of the network. The PPC population integrates the DV population activation into position information. The network dynamics have the form described in Eqs 6 and 7.

Fig 3. A key concept behind the methodology developed in the present work is that motor tasks in speech articulation can be viewed as a sequence of movements toward and away from target points in articulatory space.

Those targets are assumed to be approached and approximated, but not necessarily reached, at the temporal center of each phone interval. The initial position for a given task is assumed to be the target immediately preceding the current one.

Fig 4. Illustration of the key relationships in calculating ID from articulatory data, with most variable names taken from the text.

Target vectors are defined in the high-dimensional articulatory space, represented in the illustration by features x₁, x₂, x₃. In the analysis, this articulatory space is actually composed of L total features. The articulatory target vector F_g is the target of the previous movement, and represents the starting point of the current movement. The target of the current movement is F_h. The distance to the target is the Euclidean distance between these two vectors. The width around the target is calculated with respect to a hypersphere around the current target, which is used to estimate the density of other target vectors that are not the current one.

Fig 5. Illustration of the different syllable position-specific task categories used in the present analysis, shown on a traditional, generic syllable structure tree.

Categories are numbered outward from the nucleus, and include tasks leading into and out of the nucleus (1 & 2), tasks between consonants in the onset and coda (3 & 4) and tasks leading from one syllable to the next (5).

Fig 6. Images illustrating stages in the data pre-processing pipeline for a single vocal tract posture.

Shown are (a) an image of a single posture, in its original form, (b) the same image with low-variance pixels masked out (c) the image again, reconstructed as an image, but using only the L PCA-generated features.

Fig 7. Example high- and low-ID tasks for subject M2.

The top row, (a)-(b), represents one of the highest ID tasks, while the bottom row (c)-(d) represents one of the lowest. Images were reconstructed from the L articulatory features in Z (see text).

Fig 8. Movement time (MT) vs. index of difficulty (ID) for subject M2.

All context-target tasks are shown, divided by syllable position-based category (see text for details concerning categories).