OpenAI launches two new transcription models for better audio processing

OpenAI has released two new transcription models, GPT-Live-Transcribe and GPT-Transcribe, available via API. The first is for low-latency real-time transcription at $0.017 per minute, and the second for async audio files at $0.0045 per minute.

The company says these models better understand context, accents, numbers, jargon, and noisy audio. In the Context Aware ASR test, GPT-Transcribe's semantic accuracy improved from 41.6% without context to 45.2% with context.

On Common Voice with 22 languages, GPT-Transcribe had a 19.27% word error rate, much lower than Whisper's 40.37%. In real-world audio tests across nine languages, it achieved an 8.98% error rate versus Whisper's 15.21%.