APPLE FINDS A WAY TO MAKE AI SPEECH FASTER WITHOUT SACRIFICING NATURAL SOUND

Apple researchers have introduced a new method that significantly speeds up artificial speech generation while preserving clarity, tone, and speaker identity.

- 

The technique, called Principled Coarse-Grained Acceptance, rethinks how speech models decide whether a predicted sound unit is “close enough” to be used, replacing rigid token matching with acoustically informed group verification.

The work, published by Apple's machine learning research team, addresses one of the largest bottlenecks in text-to-speech and voice AI systems: the time required to generate natural-sounding audio, especially in large language models trained for speech.

Rather than forcing models to agree on an exact sound token, the new approach allows models to accept acoustically similar sounds as equivalent, dramatically increasing the speed at which spoken audio can be produced.

The technique is designed for speech-based large language models that generate acoustic tokens, which are discrete units representing small segments of sound rather than words or letters.

WHY CURRENT SPEECH MODELS ARE SLOWER THAN THEY SHOULD BE

Modern speech systems often rely on autoregressive decoding, meaning each sound token must be generated one at a time, with each step depending on the previous one. While this approach can produce high-quality results, it is computationally expensive and slow, especially for long sentences or real-time interactions.

To improve speed, many models use speculative decoding, a process where a smaller “draft” model quickly proposes several tokens at once, and a larger “target” model checks whether they are correct. If they match, the system can skip steps and move forward faster.

The problem is that speech tokens rarely need to match perfectly.