OpenAI Launches New Transcription API Models, Cuts Speech Error Rate to 9.6% on Real-World Audio Benchmark
OpenAI launched two new speech transcription models to its developer API, GPT-Live-Transcribe for low-latency audio and GPT-Transcribe for batch processing. The models incorporate free-form context prompts to improve recognition accuracy, with the live version achieving a 9.60% error rate on a nine-language real-world audio benchmark.
Providing contextual keywords and language tags raised semantic accuracy by 6.1 percentage points for the live model and 3.6 points for the batch version. Performance data places both new models ahead of OpenAI's prior Whisper-based alternatives, which registered error rates ranging from 11.65% to 40.37% across the tested language sets.
From the sources (4 posts)
@openaidevsWe're introducing two new transcription models in the API: • GPT-Live-Transcribe: built for low-latency live transcription. • GPT-Transcribe: optimized for asynchronous transcription of completed audio files and batch workloads. Both mode
@openaidevsBuilders can improve live audio transcription by providing: • Free-form context about the recording • Keywords for names and domain-specific terms • Expected input languages • Earlier transcribed turns as context On our new Context Aware
@openaidevsGPT-Transcribe also improved on the Context Aware ASR benchmark, with semantic accuracy increasing from 41.6% without free-form context to 45.2% with it. Across 22 languages on Common Voice, GPT-Transcribe achieved a 19.27% transcription e
@openaidevs