ECCD decoding method cuts TTS speech hallucinations by 55.6%

Language model-based text-to-speech (LM-based TTS) systems can hallucinate—producing speech that deviates from the target text, sometimes reading words that aren't there. Existing fixes rely on architectural changes or extra training, but a new paper from researchers Chenlin Liu, Minghui Fang, Zhonghao Bi, Zekai Su, Rong Wang, and Jiqing Han proposes a decoding-time solution that requires no retraining. Their method, Experience-Calibrated Contrastive Decoding (ECCD), distinguishes between text-derived alignment information and "experience information" supplied by acoustic context and learned speech regularities. It then uses predictions from the same speech LM with and without text conditions to strengthen alignment support, ensuring the model follows the script while preserving natural prosody.

Across four TTS models, ECCD reduced word error rate (WER) and character error rate (CER) by up to 55.6% on all SeedTTS-Eval settings and in 24 of 25 multilingual CV3-Eval settings. A listening test showed a CMOS gain of +0.644—meaning human listeners rated it significantly more natural—while speaker similarity stayed strong. The analysis also revealed that alignment influence varies within linguistic units, with lower decision-level gains at first-error boundaries than at matched correct boundaries. This identifies conditional information control as a promising new direction for mitigating speech hallucination at inference time, without modifying model weights or training pipelines.