Skip to main content

OpenAI's New Transcription Models Leave Whisper in the Dust

OpenAI has released two new transcription models that promise to make machines better listeners. On July 29, the company made GPT-Live-Transcribe and GPT-Transcribe available via API, and they're not just about turning sound into text anymore. These models are learning to grasp the context behind words—accents, technical jargon, numbers, and even background chatter no longer trip them up as easily.

Two Models, Two Strengths

Each model has its own niche. GPT-Live-Transcribe is built for real-time, low-latency scenarios—think of it as a stenographer that keeps up with the speaker in real time. It costs $0.017 per minute, roughly 12 cents in Chinese yuan. GPT-Transcribe, on the other hand, focuses on completed audio files and batch processing. It uses an asynchronous approach and costs just $0.0045 per minute, about 3 cents, making it ideal for post-recording bulk transcription.

The Context Breakthrough

The real leap forward is in context understanding. On the Context Aware ASR benchmark, GPT-Transcribe's semantic accuracy jumped from 41.6% without context to 45.2% when given free-form context. That means the model is starting to understand that the same word can mean different things in different conversations—a crucial step toward more natural voice interaction.

Leaving Whisper Behind

Compared to OpenAI's earlier Whisper model, the new models are in a different league. In tests across 22 languages on the Common Voice dataset, GPT-Transcribe reduced the word error rate to 19.27%, while Whisper (whisper-1) lagged at 40.37%—more than double the errors. In real-world recordings across nine languages, GPT-Transcribe achieved an error rate of just 8.98%, far better than Whisper's 15.21%.

What This Means

When machines can clearly understand professional terms in noisy environments, the door to voice interaction swings open a little wider. These improvements could make voice assistants, transcription services, and real-time translation more reliable and useful in everyday life.

Key Points

  • OpenAI released two new transcription models: GPT-Live-Transcribe (real-time, $0.017/min) and GPT-Transcribe (async, $0.0045/min).
  • The models understand context, accents, and background noise better than Whisper.
  • GPT-Transcribe cut error rates to 19.27% on Common Voice (vs. Whisper's 40.37%) and 8.98% on real-world recordings (vs. Whisper's 15.21%).
  • Context-aware accuracy improved from 41.6% to 45.2% when context was provided.