Responsiveness—the speed at which a text-to-speech (TTS) system produces audible output—is critical for real-time voice assistants yet has received far less attention than perceptual quality metrics. Existing evaluations often touch on latency but do not establish reproducible, open-source standards that capture responsiveness as a first-class dimension. This work introduces a baseline benchmark designed to fill that gap. Our framework unifies latency distribution, tail latency, and intelligibility within a transparent and dataset-diverse pipeline, enabling a fair and replicable comparison across 13 widely used open-source TTS models. By grounding evaluation in structured input sets ranging from single words to sentence-length utterances and adopting a methodology inspired by standardized inference benchmarks, we capture both typical and worst-case user experiences. Unlike prior studies that emphasize closed or proprietary systems, our focus is on establishing open, reproducible baselines rather than ranking against commercial references. The results reveal substantial variability across architectures, with some models delivering near-instant responses while others fail to meet interactive thresholds. By centering evaluation on responsiveness and reproducibility, this study provides an infrastructural foundation for benchmarking TTS systems and lays the groundwork for more comprehensive assessments that integrate both fidelity and speed.

Effective turn-taking is fundamental to conversational interactions, shaping the fluidity of communication across human dialogues and interactions with spoken dialogue systems (SDS). Despite its apparent simplicity, conversational turn-taking involves complex timing mechanisms influenced by various linguistic, prosodic, and multimodal cues. This review synthesises recent theoretical insights and practical advancements in understanding and modelling conversational timing dynamics, emphasising critical phenomena such as voice activity (VA), turn floor offsets (TFO), and predictive turn-taking. We first discuss foundational concepts, such as voice activity detection (VAD) and inter-pausal units (IPUs), and highlight their significance for systematically representing dialogue states. Central to the challenge of interactive systems is distinguishing moments when conversational roles shift versus when they remain with the current speaker, encapsulated by the concepts of “hold” and “shift”. The timing of these transitions, measured through Turn Floor Offsets (TFOs), aligns closely with minimal human reaction times, suggesting biological underpinnings while exhibiting cross-linguistic variability. This review further explores computational turn-taking heuristics and models, noting that simplistic strategies may reduce interruptions yet risk introducing unnatural delays. Integrating multimodal signals, prosodic, verbal, visual, and predictive mechanisms is emphasised as essential for future developments in achieving human-like conversational responsiveness.

This study addresses a significant gap in voice assistant research by evaluating the responsiveness - the speed at which a TTS system generates speech in reaction to input, crucial for maintaining natural, real-time interactions - of open-source text-to-speech (TTS) models—an often overlooked yet critical component for real-time applications. While extensive benchmarking has been performed on speech-to-text and large language models, little work has focused on how efficiently TTS systems respond in live settings—largely because TTS research has historically prioritized subjective quality metrics like naturalness and intelligibility, which are easier to assess through human listening tests than real-time performance; additionally, the lack of standardized, reproducible tools for measuring latency and responsiveness has further limited progress in this area. This work presents the first comprehensive benchmark focused on responsiveness—assessing TTS latency, tail latency, and real-time processing performance across 13 prominent open-source, readily available models, in contrast to commercial systems like Amazon Polly or Google Cloud TTS, which are closed-source and paywalled. Using a standardized single-stream evaluation inspired by MLPerf Inference, the study measures model responsiveness under controlled conditions and also investigates trade-offs between speed and audio quality. Results reveal substantial variability across models, with some achieving sub-second latency suitable for interactive systems, while others fall short of real-time standards. The benchmark highlights performance bottlenecks in autoregressive architectures and identifies parallel and flow-based models as more efficient for low-latency scenarios. Importantly, the proposed framework provides a reproducible foundation for comparing TTS models in latency-sensitive environments and sets a baseline for future research. By focusing on responsiveness, this work contributes to the development of more effective and natural voice interfaces.