OpenAI Replaces Whisper With gpt-transcribe, Slashing Errors by 52%
OpenAI's new gpt-transcribe and gpt-live-transcribe models bring context-aware speech recognition to the API, cutting error rates by up to 41% over Whisper
- Two new models: gpt-transcribe (file/batch) and gpt-live-transcribe (realtime streams) are now available in the OpenAI API.
- Big accuracy gains: gpt-transcribe cuts transcription error rate from 40.37% to 19.27% vs. whisper-1 on Common Voice across 22 languages.
- Context-aware: Both models accept a free-form prompt, keyword hints, and language hints to improve accuracy at inference time.
- Cheaper or same price: gpt-transcribe costs $0.0045/min (vs. $0.006/min for gpt-4o-transcribe); gpt-live-transcribe matches gpt-realtime-whisper at $0.017/min.
- Key gaps: No word timestamps, SRT/VTT output, speaker diarization, or English translation — those still require whisper-1 or gpt-4o-transcribe-diarize.
- New benchmark: OpenAI introduced a Context Aware ASR benchmark measuring semantic accuracy, where context injection boosted scores by 3-6 percentage points.
OpenAI just shipped two new transcription models in the API: gpt-live-transcribe for low-latency live audio streams and gpt-transcribe for completed file uploads and batch workloads. Both are positioned as the new recommended starting points for any transcription integration, replacing the previous defaults of gpt-realtime-whisper-1 and whisper-1.
The headline improvement is not just raw accuracy, it is contextual awareness. Traditional ASR (Automatic Speech Recognition) models treat every audio chunk in isolation. These new models can take in free-form context about the recording, a list of domain-specific keywords, expected languages, and in the case of gpt-transcribe, prior conversation turns. That shift from stateless transcription to context-informed transcription is what makes these models meaningfully different.
The numbers that matter
OpenAI benchmarked both models across several datasets. Here is how they stack up against their predecessors:
| Model | Benchmark | New model TER | Old model TER |
|---|---|---|---|
| gpt-live-transcribe | Common Voice (22 langs) | 19.70% | 20.33% (gpt-realtime-whisper-1) |
| gpt-live-transcribe | Real-World Audio (9 langs) | 9.60% | 11.65% (gpt-realtime-whisper-1) |
| gpt-transcribe | Common Voice (22 langs) | 19.27% | 40.37% (whisper-1) |
| gpt-transcribe | Real-World Audio (9 langs) | 8.98% | 15.21% (whisper-1) |
The jump for gpt-transcribe on Common Voice is particularly striking: a drop from 40.37% to 19.27% transcription error rate compared to whisper-1. That is roughly a 52% relative reduction in errors. For gpt-live-transcribe