OpenAI launches the Realtime API for low-latency voice apps

OpenAI released a public beta of the Realtime API, letting developers wire speech-to-speech conversation directly into their own applications rather than chaining together separate speech-to-text, language-model and text-to-speech calls.

The API streamed audio in both directions over a persistent connection, using the same six preset voices already available elsewhere in OpenAI’s products and handling conversational interruptions automatically. The pitch was capability already visible to consumers in ChatGPT’s Advanced Voice Mode, now exposed as infrastructure other companies could build on — customer-service bots, language-learning apps, voice assistants — without engineering the audio pipeline themselves. OpenAI also added audio input and output to its ordinary Chat Completions API for use cases that did not need the Realtime API’s low latency.

The release mattered less for any single capability than for what it signalled about where OpenAI expected developer demand to go: voice as a first-class interface rather than an add-on. It followed roughly a year after ChatGPT’s own voice mode launched to consumers, and set up a wave of voice-agent products from third parties over the following year.