A Model for Generating On-Off Speech Patterns in Two-Way Conversation

TLDR

The paper proposes a stochastic model that generates on‑off speech patterns observed in two‑way telephone conversations and explores its applications to studying conversational motivations and predicting voice‑device behavior. The model represents a conversant as occupying one of six states (three speaking, three silent) with Poisson‑driven transitions parameterized by six rates, and its validity is assessed by simulating 16 conversations and comparing cumulative distribution functions of ten speech‑pattern events to real data. The model fits most speech‑pattern events well, except for the timing of speech before interruption where it predicts later interruptions, and analysis shows all events arise from concatenated exponential state durations, making exponential fits adequate for talkspurts but not for pauses.

Abstract

This paper describes a model that generates on-off speech patterns representative of those in experimental two-way telephone conversations. The model assumes a conversant to occupy one of three speaking or one of three silent states. Transitions among the states arc determined by Poisson processes governed by six parameters (one for each state). The validity of the model is tested by comparing the model computer simulation of 16 conversations with 16 real conversations. Cumulative distribution functions are compared for ten events (such as talkspurts, pauses, mutual silences, and so on) defined on the speech patterns. The model yields good fits to all events except “speech before interruption;” when an interruption occurs, a model speaker tends to interrupt the other's talkspurt later than a real speaker does. Theoretical behavior of the model is also studied. All events consist of concatenations of exponentially distributed “state durations,” even though most events are not themselves exponential. For some purposes, the exponential distribution is a satisfactory empirical fit to talkspurts, but not to pauses. Possible applications of the model include studying people's motivations to talk and fall silent on different circuits, and predicting statistical behavior of voice operated devices on the circuits.