Realtime APIs Threaten Modular TTS

The real risk is that voice becomes a bundled feature instead of a standalone purchase. In the older stack, a developer had to wire together speech recognition, an LLM, and a separate voice model, then manage streaming between each step. OpenAI’s Realtime API and similar native speech systems collapse more of that flow into one call, which cuts latency, removes failure points, and makes a modular TTS vendor easier to swap out or skip entirely.1 2 3 4

• Vapi’s own docs describe voice agents as three swappable modules, transcriber, model, and voice, and position Vapi as the layer that picks and manages those pieces. That architecture helps customers compare providers component by component, which turns TTS into a line item instead of the core product.2 5 6
• Vapi also supports its own house voice option, custom voice providers, bring your own credentials, and fallback plans if a voice provider fails. In practice that means an orchestration layer can reroute traffic, test cheaper voices, or replace Fish Audio without changing the rest of the app.5 6 7
• This pressure is already visible across adjacent vendors. Cartesia faces the same threat from bundled speech to speech systems, while Fish Audio still appears small at about $21M in latest estimated revenue. That makes distribution control and default placement inside partner stacks especially important.3 4

The market is likely to split in two. General purpose voice will get absorbed into full stack realtime platforms, while independent TTS vendors will need to win on something a bundle does not, such as better multilingual cloning, clearer brand voice control, or proprietary enterprise distribution that keeps them embedded in production workflows.1 3 4