This talk introduces "perceptual debugging": fault-finding techniques that pinpoint which design decisions in building talking machines (speech synthesisers) that cause audible quality degradations. Just like a profiler or a debugger in software development, this information can help guide our work to improve speech-synthesis sound quality.

First, we dissect existing synthesisers in order to isolate the most important factors behind the recent success of DNN-based text-to-speech systems. Next, we step into the future, and show how we can identify critical design bottlenecks that limit the quality that both current and future parametric speech synthesisers may achieve. In a development that sounds like science fiction, we are able to simulate and listen to hypothetical, highly sophisticated synthesisers beyond the reach of current human engineering! This enables us to place upper bounds on achievable speech-synthesis quality that apply regardless of the underlying technology we use. Bounds identified in this way have recently been validated by the breakthrough success of Google's WaveNet model. We round off by describing how the proposed technique may be extended to answer further questions previously considered unanswerable.