GPT-4o Released: A 'Sci-Fi Moment' for Real-Time Speech Translation

On May 13, 2024, OpenAI released its flagship model GPT-4o ("o" for omni). Among the most impressive moments of the launch demo was near-real-time spoken translation: the presenter spoke English, the AI instantly rendered Italian, the counterpart replied in Italian, and the AI translated back — with latency close to natural conversation and even tone and emotion preserved in the voice.

The clip replayed endlessly across global media, and the question "will AI replace interpreters?" heated up again. Professional judgment inside the industry stayed characteristically sober: the demo was a pace-controlled two-person dialogue, while real conferences involve multiple speakers, specialist terminology, accents, noise and cultural nuance — a different order of difficulty. Speech translation will keep improving as a "communication aid", but simultaneous interpreting for high-stakes meetings remains professional interpreters' territory.

The truly noteworthy change is architectural: GPT-4o processes speech-to-speech end to end, no longer chaining "speech-to-text → translate → text-to-speech" — meaning less error compounding and more natural prosody. As this technical route matures, it will progressively improve remote-meeting assistance, exhibition services and similar scenarios.

For interpreters, the GPT-4o demo is a mirror: the market for simple communication will keep moving to technology, while the value of professional interpreting — accuracy, accountability, cultural judgment — stands out all the clearer by contrast.