Vapi vs Vercel

Vapi's recent releases concentrate on production-grade voice infrastructure rather than new capability surfaces. The transcriber lineup is expanding — Soniox is now GA, Deepgram Flux gained multilingual support, and an autofallback plan lets the platform pick a backup transcriber mid-call if the primary fails. Real-time signals for UI consumers are also maturing: assistant.speechStarted went GA with per-word timing on ElevenLabs and cursor-based word progress on Minimax, opening clean integrations for live captions and karaoke-style UI. Squads agent handoffs picked up a previousAssistantMessages context type, and Monitoring graduated to GA in mid-April with trigger-based rules and dashboard alerts.

The shipping cadence is weekly and the through-line is consolidation from feature-shipping to production-ready surfaces — GA flags, fallback plans, monitoring. The transcriber expansion is the most directional piece: Soniox plus Deepgram Flux plus autofallback selection is both a hedge against single-provider dependence and a clear play for multilingual workloads. The crawler is picking up duplicate stub entries per week alongside the content-bearing ones, which inflates the apparent volume but does not reflect duplicate releases.

Expect a TTS-side mirror of the transcriber autofallback work next, given the symmetry of the voice stack, plus deeper Monitoring integrations — likely structured alert webhooks and custom rule templates. The previousAssistantMessages handoff type suggests more granular context-shaping primitives for Squads are queued.