been thinking a lot lately about cascaded voice systems vs native audio voice models that take audio in as well as output it and which will be more prevalent in the future…
cascaded models make sense when you look at them through the lens of where we started - they’re relatively easy to stitch together via different components like STT, LLM, and TTS and allow us to pick and choose the best component as each component improves. They’re also really beneficial for auditing text at every point in the chained pipeline which is great for compliance reasons.
However, my main concern with cascaded systems is the fact that they throw away at least half of what makes speech so important - all the paralinguistic context such as tone, prosody, emotion, etc… for me, personally, how something is said can sometimes be more important than what was actually said and I think we’ll want to build, and interact with, systems that capture both of those things in the future. If what we care about, as a society, is naturalness and AI systems that can steer their expressiveness to match your current state, then I can’t see us using voice models that lose the ‘how’ in the process.
Having said this, as with everything, it will also depend on the use case and context - task-oriented use cases will probably care more about what was said as opposed to the how and whether or not the task was completed whereas other use cases will most definitely want to understand how something was expressed in order to respond with the right tone and prosody.