A16Z AI Voice Update 2025

I don't get it -- textual support chatbots have been around for decades. Even if we accept the premise that people would rather speak to them by voice, how do voice agents represent some kind of sea change in availability?

(And I personally find customer support chatbots deeply frustrating to use for reasons that have nothing to do with the modality or the quality of the AI model. I only ever need to use one when the question I have is not answered in the documentation, which is often the extent of the chatbot's business-specific training data. Inevitably I end up being led in circles, screaming for a human.)

> It is the most frequent (and information dense)

Second, this is false. Voice is effective when the sensory context is available to both people, e.g. in the dinner table where "pass the salt" makes immediate sense. Otherwise it is an erratic form of communication, prone to misunderstanding, often repetitive and redundant.

It is not more information dense, but it is the most immediate. The latency of AI applications makes its immediacy less useful.

Voice is simply natural to humans. Downloading an app to learn about the departure of the next bus is not.

I used voice bots to let my 5-year-old play role-playing games (e.g., checking into a hotel) or let my parents (60+) call a fake car dealership.

It's amazing to observe. They behave as if they're talking to a human, especially when doing it via a phone. That is exactly the UX a computer system should have—simply a phone number and voice.

As soon as people have to learn something new (a new webpage, a new app, etc.), something is wrong.

That’s… quite the claim. I guess we’re picking the worst people, the best voice-based AI, the easiest of scenarios, and a total desire for humanity to remove other human from interaction.

Pretty dark and sinister if you ask me.

My beef with AI voice is it's so fucking slow. As someone use to podcasts at 3-4x speed, I can't wait to ditch human interaction if as voice agents adopt variable speech rate.

I will note that the model has been successively nerfed, massively, from launch, you can watch some demo pre-launch videos, or just try out some basic engagement, for instance, try asking it to talk to you in various accents and see which ones Open AI deems “inappropriate” to ask for and which are fine. This kind of enshittification I think is pretty likely when you are the only one in town with a product.

That said, even moderately enshittified, there’s something magic about an end to end trained multimodal model — it can change tone of voice on request. In fact, my standard prompt asks it to mirror my tone of voice and cadence. This is really unique. It’s not achievable through a whisper -> LLM -> Synthesizer/TTS approach. It can give you a Boston accent, speculate that a Marseille accent is the equivalent in French, and then (at least try) to give you a Marseille accent. This is pretty strong medicine, and I love it.

There’s been so much LLM commoditization this year, and of course the chains keep moving forward on intelligence. But, I hope Ms. Moore is correct that we’ll see better and more voice models soon, and that someone can crack the architecture.

I'll take a professional actor over TTS any day - incomparably better quality even with the best TTS.

Anthropomorphism is to AI what skeuomorphism is to UIs. I can’t wait for us to move into the “flat design” era of AI, where instead of being patronized with phrases like “Hi! I’m Bobby! Your intelligent AI assistant, how can I help you?” we just get something cold and straight to the point like “Ready for Instructions”, in some crunchy byte encoding. Sorry for the rambling, I’m a little drunk.

But the whole point of this medium is that you want the humanity and personality. Otherwise just use text.