Skip to main content
This guide only applies to cascading agents, if you are using speech to speech models, this feature does not apply.
Real time transcription is a trade off between latency and accuracy. Two concepts drive this trade off:
  • Interim results: early, partial transcripts produced while the user is still speaking. They arrive fast but are generated with less context, so they contain more errors.
  • Endpointing: how long the transcriber waits after the user stops speaking before finalizing the transcript. Waiting longer gives the model more context and produces more accurate transcripts, at the cost of a slower response.

​
Transcription modes

Retell offers two transcription modes:
  • Optimize for speed: uses the latest interim results with a low endpointing setting for downstream processing. This gives the lowest latency.
  • Optimize for accuracy: uses results generated with a higher endpointing setting for downstream processing. The transcriber waits longer with more context to generate more accurate transcripts. This adds roughly 200ms of latency.

​
Which mode to use?

From our benchmarking, the two modes have a similar overall WER (Word Error Rate). The difference mainly lies in capturing entities like numbers and dates.
  • Use optimize for accuracy if your use case relies heavily on capturing entities like phone numbers, account numbers, or dates correctly.
  • Use optimize for speed for the best latency in all other cases.