Abstract: Non-autoregressive (NAR) text-to-speech (TTS) models have gained significant attention for their parallel inference and global style consistency. However, their reliance on accurate character-level or global duration predictions remains a critical bottleneck, limiting both synthesis fidelity and naturalness. In this paper, we propose ElasticDLM (Elastic-length Diffusion Language Model), a novel variable-length NAR TTS model that eliminates the strong dependency on accurate duration predictions. Specifically, we introduce two functional tokens, along with a Differentiated Length-Scaling Training Scheme (DLTS) and a Hierarchical Confidence-Guided Inference (HCGI) strategy tailored for speech characteristics, enabling the model to learn length-scaling features and automatically adjust sequence length during inference. Experiments show that ElasticDLM produces high-fidelity, natural speech from arbitrary-length inputs, demonstrating a more flexible and simplified approach to high-quality speech synthesis.

Model Architecture

Main Results

We conducted our evaluations on the seed-test-zh and seed-test-en benchmarks from Seed-TTS, using MaskGCT as our primary baseline. We designed three distinct settings for the MaskGCT baseline. The first, termed Normal, represents the standard MaskGCT which relies on its built-in duration model. The second, Fix, simulates a poor duration estimate by initializing the sequence to a fixed length (4s for Chinese, 4.5s for English). Finally, to serve as a performance upper bound, the GT setting provides MaskGCT with the ground-truth audio length.

Normal Len

文本 (Transcript)	Ground Truth	MaskGCT (GT Len)	MaskGCT (Normal Len)	Ours (Normal Len)
中文 (Chinese)
这是专门为山地地区，作战研发的全新坦克。
而且全球百分之二十的，新发癌症病人在中国。
他们宁愿躺在家里睡上半年，也不肯麻烦一点到南方去。
标题为中国如此害怕印度的五大理由。
上下班耗费的时间，影响着人们对工作的满意度。
英文 (English)
He lives in a tent, smokes a pipe and plays the harmonica.
It was the first time either had touched the other.
There is no universal definition of intelligence, but everyone agrees that the ability of learning belongs to it.
With his promotion to colonel, he was named director of personnel.
He also regularly attends events with the Royal Family.

Fix Len

文本 (Transcript)	Ground Truth	MaskGCT (Fix Len)	Ours (Fix Len)	Ours (w/o Len Loss)	Ours (w/o Inference)
中文 (Chinese)
大象卷住你妈，使劲一甩，你妈妈扑通一声被甩进了河里。
榜单前五位分别为阿里巴巴，海尔集团，百度，华为和联想。
他不得不骑在马上颠簸赶路，而不能够躺在床上安详熟睡。
兔妈妈胀麻的身子一下子被抽空，久违的轻松回来了。
其中又有大部分百分之四十一点九认为，由该品牌影楼拍摄。
英文 (English)
Will you be happy when I am in the earth?
Smith was a fellow of the British Academy.
He was educated at Wellington College, Berkshire.
Do the roses compete in athletic events?
The four group winners had the opportunity to promote directly.

Different Initial Length

To validate the powerful length adaptation capability of our model, this section tests its speech generation performance under conditions of varying initial sequence lengths. This experiment demonstrates that the model does not rely on precise prior length information and can robustly generate high-quality target speech from varied initial states, highlighting its flexibility and robustness.

Prompt Audio	鞋子和鞋垫可以使用抑菌喷剂，或散剂进行消毒。
Target Audio (GT)	并采取实际行动，去培训快递员这方面的法律意识。

Initial Length	MaskGCT	Ours
1s
4s
7s
10s
13s
16s

Analysis of Application Flexibility

To assess application flexibility, we fine-tuned our model and the MaskGCT baseline on a highly expressive single-speaker dataset. An AB preference test was then conducted, where participants evaluated speech synthesized from held-out texts in four emotional styles: Fearful, Sad, Coquettish, and Excited.

Prompt Audio

每个人的生活节奏都不一样，你也不必太在意别人的眼光啊。

Emotion	Transcript	MaskGCT	Ours
Excited (喜悦)	哇塞！生活处处是惊喜呀！
Excited (喜悦)	嘿嘿，和你聊天超欢乐！
Sad (悲伤)	我的心就像被掏空了一样，孤独和悲伤一直围绕着我。
Sad (悲伤)	日子太难熬，满心疲惫，真不知何时能轻松些。
Coquettish (撒娇)	人家好难过哦，你都不陪陪我，真的太过分了。
Coquettish (撒娇)	哼，你要是再不陪我玩，我就要闹小脾气喽。
Fearful (恐惧)	我不想待在这儿了，到处都透着诡异，我快受不了啦！
Fearful (恐惧)	我真的好后悔来到这个地方，早知道会这样，我打死也不会来。