OpenAI Could Replace Speech APIs

The real threat is not that OpenAI becomes another speech vendor, it is that it turns speech into one bundled model call and makes separate transcription APIs easier to skip. In the modular stack, builders pick STT, then pass text into an LLM, then send the answer to TTS. OpenAI now offers a realtime voice model that can take audio in, reason, use tools, and answer in audio inside one session, which removes the handoff points where vendors like AssemblyAI win usage.1 2 3 4

• Model agnostic voice platforms like Retell, Vapi, and Bland were built around swapping components. Retell charges a platform fee on top of pass through LLM, STT, and TTS costs from providers like OpenAI, Deepgram, Cartesia, and ElevenLabs. That architecture weakens lock in for any one speech API, but it still depends on modular pieces existing.5 6
• Integrated speech vendors are already moving in the same direction. Deepgram has expanded from STT into TTS and a Voice Agent API, while Cartesia now sells STT, TTS, and full agent deployment. The market is shifting from best single component to who can own more of the live call workflow.7 8
• AssemblyAI still has room where teams want a dedicated speech layer, especially when they need to plug transcription into many products beyond live agents. But as more new voice apps start with a speech to speech session instead of a chained pipeline, the buying decision moves up from STT quality alone to the whole agent stack.1 2 4

The next step is a split market. High control teams will keep mixing vendors for latency, cost, and workflow tuning, while a growing share of builders will accept a single realtime model if it is good enough and faster to ship. That pushes specialized speech companies to move upward into agents, orchestration, and workflow software where the budget is larger and the customer relationship is stickier.2 5 7 8