OpenAI released two new auditory engines on July 29th. The GPT-Live-Transcribe and GPT-Transcribe transcription models are now available for API calls. They are no longer just simple tools that convert sound into text; instead, they are learning to understand the context behind a sentence - accents, terminology, numbers, and even the background noise of people talking are no longer easily confusing them.

Each of the two models has its own characteristics. GPT-Live-Transcribe is designed for low-latency real-time scenarios, like a stenographer who writes down what is being said in sync with the speaker. It costs $0.017 per minute, which is approximately 0.12 Chinese yuan at the current exchange rate. GPT-Transcribe focuses on completed audio files and batch tasks, using an asynchronous transcription approach, with a price of $0.0045 per minute, approximately 0.03 Chinese yuan, making it more suitable for processing recordings in bulk after the fact.

The real breakthrough lies in their ability to understand context. On the Context Aware ASR benchmark, GPT-Transcribe's semantic accuracy rate increased from 41.6% without free-form context to 45.2% when context was provided - it begins to understand that the same word can mean completely different things in different conversations.

Compared to the veteran Whisper, the new model has gained a clear advantage. In the 22 languages covered by Common Voice, GPT-Transcribe reduced the transcription error rate to 19.27%, while Whisper (whisper-1) remained at 40.37%, more than double the error rate; when tested on nine languages in real-world recordings, it cut the error rate to 8.98%, far behind Whisper's 15.21%. When machines can clearly understand professional terms in noisy environments, the door to voice interaction seems to have been opened a little wider.