XNews

OpenAI Releases GPT-Transcribe Speech-to-Text Models

OpenAI has released gpt-transcribe and gpt-live-transcribe, new speech-to-text models. - On a real-world recording benchmark, gpt-transcribe achieved an 8.98% word error rate (WER), compared to 15.21% for whisper-1. - This represents a relative error rate reduction of approximately 41%. - Pricing is $0.0045 per minute, 25% cheaper than whisper-1's $0.006 per minute. - Both accuracy and cost improved simultaneously.

Key insights

GPT-Transcribe Sets a New Price-Performance Standard for Speech-to-Text

OpenAI's new transcription models deliver a significant accuracy improvement over whisper-1, nearly halving the word error rate on real-world audio, while also reducing cost by 25%. The combination of higher accuracy and lower price makes the service more attractive for high-volume transcription use cases. The availability of a live transcription variant (gpt-live-transcribe) suggests OpenAI is targeting real-time applications such as meetings and live captioning.

Source posts