---
title: "The Three Futures of Voice AI (Substack)"
url: https://stacklist.com/card/f19e38c4-896f-4094-ae19-08d4041ee6d1
source_url: "https://fdaudens.substack.com/p/the-three-futures-of-voice-ai"
stack: https://stacklist.com/c/technology/stack/8521690e-e828-43bf-a05b-91287c60612b
summary: "Voice AI insiders disagree on fundamental questions about the future of voice technology, including whether AI should sound human or distinctly artificial, and whether cascade or end-to-end architectures are superior. The consensus centers on latency as the critical technical barrier to overcome for natural human-AI interaction."
tags: "voice-ai, speech-recognition, latency, agentic-future, real-time-processing, ai-interaction"
key_entities: "Florent Daudens (person), Russ d'Sa (person), Brandon Yang (person), Justin Uberti (person), Scott Stephenson (person), Dylan Fox (person), OpenAI (organization), Deepgram (organization), AssemblyAI (organization), Cartesia (organization), LiveKit (organization), Sierra (organization), MiniMax (organization), Wispr Flow (organization), GPT-Realtime-2 (technology), GPT-5 (technology), latency (concept), uncanny valley (concept), auditory Turing test (concept), Cerebral Valley Voice Summit (event), Cerebral Valley (location)"
classification: "analysis"
content_hash: "sha256:7858589253923b68b97d1259beba5f82a5bbe6d6f0bb8cf84dd280d7495828b0"
acp_version: "0.2"
token_counts_approximate: 2379
visibility: public
agent_accessible: true
status: "final"
---

# The Three Futures of Voice AI (Substack)

What 13 voice AI insiders can’t agree on And why that disagreement is the most useful signal in the room Florent Daudens May 14, 2026 4 1 Share Francis Barraud’s 1898 painting of Nipper looking into an Edison Bell cylinder phonograph - Wikipedia Voice could change how we interact with machines forever. It’s the obvious UX layer for the agentic future: how we’ll dictate to agents, brief AI assistants, run our work without keyboards. Or it stays what it is today: a feature we use occasionally, mostly for dictation, mostly when typing is annoying. Which of those happens depends on a question the people building voice AI can’t agree on. Newcomer just published all 13 videos from last week Cerebral Valley Voice Summit . A full day of stage time from the people actually building voice AI: Sierra, OpenAI, Deepgram, AssemblyAI, MiniMax, Wispr Flow, LiveKit, Cartesia. the same week, OpenAI launched GPT-Realtime-2, a new voice model that handles speech in and speech out directly, with GPT-5-class reasoning baked in (among other models). I went in expecting a unified “voice is the next platform” narrative. I came out with the opposite. The builders barely agree on anything, and that's probably a more useful signal than any of the individual talks. What they agree on (and why it’s a small list) The one real consensus: the last technical wall is latency. Russ d’Sa of LiveKit explained it most cleanly: “In real life, you’re used to [someone] acknowledging you in 250 milliseconds on average. That’s your prior. To cross the uncanny valley, you have to think about latency as this critical component.” It’s a brain-prior we evolved with. Brandon Yang built Cartesia around the lowest-latency text-to-speech models on the market. Justin Uberti’s bet at OpenAI on real-time speech-to-speech is the same wager from the model side: make the round trip fast enough that the model can do something with the signal it’s hearing. Scott Stephenson (Deepgram) puts a date on the broader version of the same problem: “The auditory Turing test: totally not solved yet, but I’d say by the end of this year it will be.” That’s where the agreement ends. After this, the room splinters in ways that have nothing to do with rivalry and everything to do with people genuinely believing different things about the future they’re building. Disagreement #1: Should AI sound human, or sound like AI? Stephenson believes naturalness wins. “You want the natural response back. I think most people are going to really take to that in a very short period of time.” The end-state is a voice you’d want to talk to, indistinguishable from a person. Dylan Fox (AssemblyAI) thinks the opposite, and his version is the more counterintuitive bet. Once people trust AI agents, he argues, machines pretending to be human start to feel uncomfortable. “If you don’t know that it’s an AI, a machine, you continue on. Then if you find out later, you’re like, ‘Whoa, this is weird.” - Dylan Fox In Fox’s future, we want AI to announce itself. He’d rather his son have an engaging conversation with a device, but know it’s a device. The first invests in warmth, ambient empathy, voice as companion. The second invests in clear non-human signals, transparency, voice as tool. Disagreement #2: Cascade or end-to-end? Justin Uberti, who runs OpenAI’s voice work, made a technically interesting argument for real-time speech-to-speech: “You have a [trillion-parameter] model being governed by a million-parameter model. It’s almost like having a child supervising an adult.” What he means: in today’s voice stacks, a tiny model is in charge of deciding when the AI gets to speak and what bits of audio reach the big model. That tiny model isn’t smart enough to make those calls well. The cascade approach (speech-to-text > LLM > text-to-speech) bottles the smart part behind a dumber gatekeeper. Going end-to-end lets the big model actually use the signal it’s hearing. If you really want to feel the AGI, if you really want to hear the AGI, we just need something that is a different type of approach that can just feel much more human in terms of the way you’re able to interact with it. - Justin Uberti His bet now has production receipts. Zillow, one of GPT-Realtime-2’s launch customers last week , reports its call-success rate on the hardest adversarial test set jumping from 69% to 95%. That’s a 26-point lift in real customer interactions on the new end-to-end model. Stephenson, from the cascade side, is making an industrial point. “Well over 95% of business use cases use speech-to-text plus an LLM plus text-to-speech right now, not a speech-to-speech model.” The reason, in his telling, is two-fold: First, controllability : a cascade gives engineers a handle on each stage: what was transcribed, what the LLM did with it, how it gets voiced. Which matters a lot when you’re a bank or a hospital and need to log and audit every step. End-to-end models are a black box by comparison. Second, inertia : “Once somebody has built something like that and it works and serves a purpose, they’re not re-evaluating.” Architectures that win on a benchmark don’t always win in production. The cascade isn’t elegant; it’s installed. This is maybe the most consequential of the three disagreements, because it decides where the money lands. If Stephenson is right, voice AI looks like cloud computing: modular and layered, with specialists winning at each layer (one company for transcription, another for the LLM, another for synthesis) and customers free to mix and match. Lots of viable winners. If Uberti is right, it looks like operating systems. When the whole stack lives inside one end-to-end model, the company that owns the model captures most of the value, and everyone else is a wrapper integrating to its API. One or two real platforms; everything else is downstream. Disagreement #3: Is voice even the primary interface? Bret Taylor (Sierra) thinks we’re on the brink of something big. The implication is that voice unlocks an entire generation of applications that weren’t possible before. We just haven’t seen them yet. “It feels like we’re in the stage of the internet before broadband... you couldn’t have YouTube before broadband.” - Bret Taylor Eugenia Kuyda — who built Replika, and has thought about human-AI voice connection for a long time — pushes back hard. The problem with voice as a primary interface isn’t the model; it’s the use case. “Discovery is almost impossible. Unless I really know what I need to say, [it’s hard] to discover what is what.” Most of what we actually do on our phones is browse. We glance at a feed, decide quickly, move on. You can’t do that with voice. Olivia Moore (a16z), who tracks consumer voice closely, offers a third reading. The dictation products that are actually working — Wispr Flow, Granola, Aqua — aren’t winning because they’re voice products: “What’s cool about Wispr Flow and Granola is their interface. What’s cool about it isn’t the fact that it’s voice. Voice is like a component, and they’ve reimagined the interface.” Granola is a notepad that joins your Zoom call. Wispr Flow is a keyboard that types faster. Voice is the engine, but it’s hidden behind the surface you actually use. Three very different optics. Taylor is right about what’s possible . Kuyda is right about what’s behavioral . Moore is right about what’s already shipping . What’s the next generation voice AI? Linda Sheng, CEO of MiniMax (a frontier multimodal lab), had the most contrarian take: voice AI has barely moved in nine months. “It looks a little bit similar compared to six months ago or even nine months ago.” “The innovation is coming from one dimension above. It’s the visual plus the video and the omni models.” - Linda Sheng In her view, the most interesting breakthroughs in expressiveness and emotion in voice AI right now aren’t coming from voice specialists, but from video labs. That should make every voice-pure startup nervous. It’s the most contrarian claim in the room, and it came from someone whose company is well-positioned to make it true. The futures are unevenly distributed By the end of the day, three distinct futures are sitting on the table: The broadband moment (Taylor’s bet). Voice AI is the slow-build platform shift. The killer apps haven’t been built because the infrastructure isn’t ready yet. When latency closes and quality crosses a threshold, a generation of products gets built on top. The infrastructure layer (Stephenson’s and d’Sa’s bet). There is no winner-take-all. Cascade stacks persist because they work and they’re already installed. Voice becomes plumbing under everything else; never a destination, always a feature. Lots of winners. No platforms. The multimodal absorption (Sheng’s bet). Voice doesn’t get its own moment. It gets folded into bigger omni-models that handle voice, video, and context together. Pure-voice startups are at a local maximum. The companies that win the next round are the ones already operating at a higher dimension. The really interesting thing is that all three are being built simultaneously, by serious operators, with real money behind them, and by people who all have the same information. It’s a good time to be paying attention. Especially to the disagreements. Original Newcomer post: Newcomer 13 Videos From the Cerebral Valley Voice Summit: Sierra's Bret Taylor, Wispr Flow's Tanay Kothari, MiniMax's Linda Sheng & More Read more 2 months ago · 47 likes · Madeline Renbarger 4 1 Share Previous Next