TiCo: Time-Controllable Training for Spoken Dialogue Models

Forum topic · 小凯 · 2026-03-25

Summary

TiCo is a simple post-training method that enables spoken dialogue models (SDMs) to follow time-constrained instructions and generate responses with controllable duration. The authors observe that existing open-source and commercial SDMs, despite generating natural spoken responses, lack time awareness and frequently fail to satisfy duration-related instructions such as producing a reply lasting about 15 seconds. TiCo addresses this by training models to estimate elapsed speaking time via speech time markers (STMs), e.g. <10.6s>, interleaved during generation. These markers help the model maintain time awareness and adjust remaining content to reach a target duration. The method is efficient: it requires only a small amount of data, no additional QA pairs, and relies on self-generated data combined with reinforcement learning. This capability is valuable for real-world spoken language systems such as voice assistants and interactive agents, where controlling response length improves interaction quality. Paper: arXiv 2603.22267.

Overview

Field: NLP Authors: Kai-Wei Chang, Wei-Chih Chen, En-Pei Hu, Hung-yi Lee, James Glass Published: 2026-03-23 arXiv: 2603.22267

Abstract

We propose TiCo, a simple post-training method for enabling spoken dialogue models (SDMs) to follow time-constrained instructions and generate responses with controllable duration. This capability is valuable for real-world spoken language systems such as voice assistants and interactive agents, where controlling response duration can improve interaction quality. However, despite their strong ability to generate natural spoken responses, existing models lack time awareness and struggle to follow duration-related instructions (e.g., "Please generate a response lasting about 15 seconds"). Through an empirical evaluation of both open-source and commercial SDMs, we show that they frequently fail to satisfy such time-control requirements.

TiCo addresses this limitation by enabling models to estimate elapsed speaking time during generation via speech time markers (STMs), such as <10.6s>. These markers help the model maintain time awareness and adjust the remaining content to hit the target duration. TiCo is simple and efficient: it only requires a small amount of data, needs no extra QA pairs, and relies on self-generated data and reinforcement learning.

--- *Auto-collected on 2026-03-25.*

TiCo: Time-Controllable Training for Spoken Dialogue Models

Summary

Overview

Abstract

Tags