VapiCon 2026: The phone is the main event
Phone calls still dominate voice AI. At VapiCon, the hard questions were how to make agents reliable and when speech-to-speech is ready for production.
VapiCon is the largest Voice AI event. I didn’t attend VapiCon last year, but I followed it. The conference this year was 2 days with twice as many sessions. In many ways, the conference itself mirrors the larger industry - double the intensity with a fractionally larger audience.
I came away from the event with three main observations:
- Voice AI is about phone calls (not web or mobile apps)
- Everyone is doing Sims and Evals to improve reliability
- The industry is skeptical about Speech-to-Speech
Voice AI improves 1870s technology
For most of the last 150 years, we’ve made calls over the public switched telephone network (PSTN). Today we have other options - like WhatsApp calling or WebRTC for calling in a browser or native app - but most of the talks and discussions I had at VapiCon were still about ordinary phone calls. Vapi’s own promotional photos showed employees holding phones. The fortune cookies had a phone number you would dial to get your fortune. They even were giving away plushy rotary phones.
A couple of examples used WebRTC or WebSockets. One was the talk by OpenAI’s Justin Uberti on GPT-Voice, which is primarily used on the web or in an app. Even there, Justin mentioned its native SIP interface.
I am a long-time fan of WebRTC (see webrtcHacks), so I have an inherent bias toward it. It costs less, offers better quality, and gives developers more flexibility. Still, better technology does not always mean a better market. As Michael Carpenter from Twilio stated, the PSTN is the universal API that can reach anyone. The big money is in more traditional call center use cases and those involve phone calls. Phone calls are a huge market.
Sims and evals are the new test and monitoring
I also noticed a lot of focus on reliability in the talks. Admittedly I am also biased here since this is an area I am working on with Cyara. Following the theme above, enterprise customers expect enterprise reliability. Traditional call centers would provide assurance via testing and monitoring. In Voice AI, you do this with:
- simulations to emulate realistic customer interactions before you load up actual live customers and
- evaluations to check that your agent did what it was supposed to do
The Production vs. Demo panel covered practical issues that trip up voice agents, including interruptions, background noise, accents, and callers changing their requests mid-sentence. The Monitor, Simulate, Evaluate panel picked up from there and looked at what teams do when those calls go wrong.
Voice AI is so much better than a year ago, but clearly it will be a long while before anyone trusts these systems without intense oversight.
Speech-to-Speech Skepticism and Scrutiny
The discussion about interruptions also came up in the architecture talks, where a recurring question was how speech-to-speech (S2S) compares with the usual STT → LLM → TTS cascade and its separate turn-taking logic. Justin Uberti of OpenAI described a two-model approach for S2S. GPT-Live keeps listening while it speaks and can delegate reasoning and tool calls to an LLM of your choice. That model passes its result back to GPT-Live. This is a simpler architecture, but GPT-Live becomes a black box that can only be steered. In addition, you are subject to the speech and voice options of GPT-Live.
Meanwhile, the cascaded approach continues to be optimized. The most significant (re)announcement in this area was from Deepgram with their Real-Time Context Standard (RCS) announcement (not to be confused with Realtime Communication Service for mobile messaging). RCS is an open specification for passing information between the speech-to-text, LLM, and text-to-speech components. Today it's a bit of a Wild West, with vendors using their own formats and tags. With RCS, a tone classifier could tell the LLM that the caller sounds frustrated, and the LLM could tell the speech synthesizer to slow down when reading a confirmation number In Pipecat’s demo of RCS, first announced at their own show 2 weeks ago, the Tone Reader infers emotion and sarcasm from the caller’s transcript and the agent’s previous reply, and the Tone Tagger adds delivery instructions for the speech synthesizer.
The spec is still at 0.1, and what actually works depends on what each component supports. It also remains to be seen if DeepGram can convince anyone else to adopt their format, but making it part of a popular framework like Pipecat provides an incentive.
Most of the crowd is using a cascaded approach. Most said they would continue using the cascaded approach. S2S is too new and unproven.
S2S may not be a good fit for most Voice AI deployments right now. However, given the rapid pace of AI improvements, it is certainly not wise to ignore it.
What is the best way to bring the Voice AI industry together?
As a former technology event organizer, I was also thinking about who gets to convene this industry. I’m a little disappointed that the best Voice AI can do is a vendor-led conference. I appreciate that VapiCon brought in the broader ecosystem, but it would have been nice to hear from some of Vapi’s direct competitors, too. Then again, starting an industry event takes time and money. Industry events usually aren’t very profitable at first. It is hard to compete against venture funding. Sadly, I missed the smaller and more focused Voice Agent Forum earlier this month, but it is encouraging to see other events.
Thanks to Vapi for putting together a great event! I do hope to see you all at the next Voice AI event.