OpenAI released a live speech translation tool on May 7, 2026, that does something the company's own general-purpose voice models could not do reliably: stay in translation mode during a real, flowing conversation without answering questions, following instructions, or needing the speaker to pause.
The difference matters more than it sounds. Earlier attempts at AI interpretation used models that were also designed to hold conversations. In practice, those models would sometimes respond to what someone said rather than translate it, or they would wait for a speaker to finish a whole thought before outputting anything. Both behaviors make them feel clunky during a live earnings call or a customer support conversation.
The new model was trained specifically on professional interpreter recordings. It waits for enough context before speaking, which is important because some languages place the verb at the end of a sentence. It also streams translated audio back while it is still receiving the original, meaning both sides of a conversation can keep talking at a natural pace.
The price point is the number that reshapes the industry. At $0.034 per minute, a full hour of live translated conversation costs around $2. Professional human simultaneous interpreters at conferences or on specialist phone lines typically cost between $50 and several hundred dollars per hour. The translation services market globally is estimated at around $65 billion this year and growing. The portion of that market that handles live spoken interpretation, covering events, support centers, medical consultations, and legal proceedings, is where this tool makes its first serious move.
Some early adopters are already in production. Deutsche Telekom is building multilingual customer support where callers speak in whichever language they prefer. Vimeo is using the model to translate product videos live as they play, removing the need for a separately recorded version in each language. BolnaAI, which builds voice tools for the Indian market, reported meaningfully lower error rates on Hindi, Tamil, and Telugu compared to previous options.
The practical limits are equally important to understand. The model currently produces output in only 13 languages: Spanish, Portuguese, French, Japanese, Russian, Chinese, German, Korean, Hindi, Indonesian, Vietnamese, Italian, and English. It can detect over 70 input languages, but if your customers or partners primarily speak a language not on the output list, this tool does not help them yet. The model also cannot be given a glossary. If your industry uses specific product names, legal terms, medical procedures, or internal shorthand, the model may substitute something plausible but wrong. OpenAI recommends testing those terms manually before any live deployment.
There is also a structural limitation for mixed-language speakers. If a customer in a Spanish-language support call switches into English mid-sentence, the model may go quiet during the English portion rather than pass it through. This is a known behavior, not a bug, but it creates a rough experience if your callers naturally mix languages.
For non-technical businesses thinking about where this matters operationally, the clearest starting points are customer support lines that currently route non-primary-language callers to a small pool of bilingual agents or third-party interpretation services, live events and webinars that require translation tracks, and any cross-border sales or procurement call where conversations currently happen through intermediaries. The model handles phone calls, browser audio, and video conferencing, so the integration path is shorter than it has historically been.
The human interpretation profession is not disappearing overnight. High-stakes settings like courtrooms, complex medical decisions, and diplomatic negotiations will continue to require certified professionals for accuracy, accountability, and legal compliance reasons. But the large, unglamorous middle of the market, the daily support calls, the regional sales meetings, the product webinars, the training sessions, moves quickly once cost per interaction drops by roughly 99 percent.