Large language models (LLMs) have shown incredible promise in generating realistic speech from text, but they're not perfect. One persistent problem is that LLMs can sometimes 'hallucinate' in speech synthesis, resulting in repeated words, missing words, or misaligned audio. Think of it like a slight stutter in the AI's voice. This issue is especially noticeable when the text contains the same word multiple times. Researchers at NVIDIA dug deep into this problem, focusing on a type of LLM called an encoder-decoder transformer model. They found that certain parts of these models, called cross-attention heads, try to learn the alignment between the text and the generated speech, but they don't always get it right. This imperfect alignment is the root cause of the stuttering. To fix this, the team developed a new training technique that encourages monotonic alignment—basically, making sure the AI reads the text in the correct order. They used something called a Connectionist Temporal Classification (CTC) loss, which helps guide the LLM to correctly match the text and speech. They also introduced 'attention priors' that act like a roadmap for the LLM, showing it a near-diagonal alignment pattern initially. Importantly, these techniques don't require any changes to the LLM architecture itself, making them an efficient fix. The results are impressive. Their improved LLM showed a significant reduction in errors, particularly with challenging texts that include repeated words. Not only was the synthesized speech more intelligible, but the quality and naturalness also improved significantly. This research paves the way for more robust, reliable, and natural-sounding LLM-based speech synthesis, making it a game-changer for applications like virtual assistants, audiobooks, and even generating realistic voices for games and virtual worlds.