Large language models (LLMs) are impressive, but they can be slow. Generating text token by token, like typing one letter at a time, creates a bottleneck. Researchers are constantly looking for ways to speed up this process, and a technique called speculative decoding is showing real promise. Imagine the LLM having a helpful assistant that drafts multiple words at once. The LLM then checks the draft and either accepts it or makes corrections. This ‘draft-and-verify’ approach can significantly accelerate text generation, but the effectiveness hinges on the quality of the drafts. If the drafts are poor, the LLM spends more time correcting than it saves. A new research paper proposes a clever way to improve these drafts using something called Connectionist Temporal Classification, or CTC. Traditionally used in speech recognition, CTC helps the draft model understand the relationships *between* words in a sequence, generating more coherent and accurate drafts. This means the LLM accepts more of the draft, leading to faster overall performance. Experiments show this CTC-based drafting method produces a notable speed boost, especially with smaller LLMs. While there’s still room for improvement in balancing drafting complexity and speed, this research offers a compelling path toward faster and more efficient LLMs.