OpenAI ships two transcription models

OpenAI released two new transcription models through its Audio and Realtime APIs: GPT Live Transcribe for low-latency streams and GPT Transcribe for completed files and batches. In OpenAI’s tests across nine languages, GPT Live Transcribe posted a 9.60% error rate, down from GPT Realtime Whisper 1’s 11.65%, while both new models accept context, keyword hints, and expected languages. Builders choosing among AI transcription workflows should file away the context controls: domain vocabulary is becoming an API input rather than a post-processing repair job.