Introducing gpt-transcribe and gpt-live-transcribe
Summary
OpenAI released two transcription models designed for different workflows: gpt-transcribe for batch processing (processes 30-min files in under 1 minute) and gpt-live-transcribe for real-time streaming. Both support 57 languages, auto-detect multilingual speech, and excel at traditionally difficult areas like accents, short answers, proper nouns, and numbers through custom vocabulary hints.
Key Takeaways
- Use custom vocabulary lists to improve accuracy on domain-specific terms (cybersecurity: phishing, sales: ARR, healthcare: A1C). Developers can provide proper nouns and code terms the model should prioritize.
- gpt-transcribe optimizes for accuracy and throughput on completed files (ideal for call archives, podcasts, batch jobs); gpt-live-transcribe prioritizes low latency for real-time captions, dictation, and voice interfaces.
- Both models automatically filter background noise and side conversations, keeping transcripts focused on primary speaker. Useful for low-quality recording environments like loud cafes or conference halls.
- Automatic multilingual detection allows seamless language switching within a single session without additional configuration or language hints. Model follows speech across language boundaries.
- gpt-transcribe can process 30-minute meeting recordings in under 1 minute. Use for asynchronous workflows where processing latency is acceptable but accuracy is critical.
Related topics
Transcript Excerpt
Today we're introducing two new transcription models, gpt-transcribe and gpt-live-transcribe. These models give you two different ways to build with transcription. gpt-transcribe takes a completed file and returns the full transcript. gpt-live-transcribe keeps a live open connection and returns text as the audio arrives. Both models support 57 languages and they're much better at the parts of transcription that tend to break accents, multilingual speech, very short answers, names, and numbers. Developers can also give the models a list of words, proper nouns, or code terms that are especially important to get right. For this session, I provided a prompt mentioning transcription models plus a vocabulary with terms like phishing, ARR, and A1C from cybersecurity, sales, and healthcare domains…