This is where you measure how accurately Deepgram transcribes your own recordings. For each recording, you provide an accurate ground truth transcript of what was said. Deepgram transcribes the audio and reports a semantic word error rate (SWER) for each file and for the full set. Word error rate (WER) is included alongside, so you can see how much of the difference comes from wording or formatting that does not change what was said.

You can sign in with a personal or work Google account. After you sign in, starting a run requires a Deepgram project ID and an API key for that project.

A run you start is visible to you, to anyone signed in with the same company Google Workspace account, and to Deepgram.

If you do not have a Google account, or you would rather run this on your own machine, dg-bench-cli reports the same metrics from the command line or as a Python library.

If something is not working, contact your Deepgram account team.