How to Choose the Right Speech-to-Text 3.7 Model (Full Transcript)
[00:00:00] Speaker 1: You wouldn't wear a scuba diving suit for a sprinting race, or a football jersey for your wedding dance, or a three-piece suit to play tennis. And it's the same way with Speech AI. With the release of our Speech-to-Text 3.7 models, each model serves its own purpose and it's crucial that you use the right model for the right job. Speech-to-Text 3.7 is our asynchronous model for pre-recorded audio. It catches every detail in your audio recording, and with top speaker diarization performance, it's as close to perfect as it gets. 3.7 Realtime is built for live conversations, think meetings, calls, and captions. It comes with live speaker labels and a competitive word error rate without sacrificing latency. But what it doesn't do is turn-taking. If anything on your end needs to tell the model when a turn is over, then what you're looking for is 3.7 Realtime Agent. 3.7 Realtime Agent is built for voice agent use cases that require ultra-low latency while not sacrificing accuracy on numbers, short replies, even in scenarios with background noise or accented speech. And so it's the perfect foundation for your cascaded voice agent architecture. So those are our three new models released today. Remember to pick the right one for your use case, and let us know if you have any questions. Bye.
- Use Speech-to-Text 3.7 for detailed transcription of pre-recorded audio with strong speaker diarization.
- Use Speech-to-Text 3.7 Realtime for live meetings, calls, and captions requiring low latency and live speaker labels.
- Realtime does not provide turn-taking support for systems that need to know when a speaker has finished.
- Use Speech-to-Text 3.7 Realtime Agent for voice agents requiring turn-taking, ultra-low latency, and strong performance on short replies, numbers, noise, and accents.
- Model selection should be driven by the specific speech application and interaction requirements.