Skip to content

Deepgram

Deepgram’s speech-to-text service turns a file or a live audio stream into words your app can use.

Visit Deepgram

Should your app send a file or stream audio while someone speaks?

Use a file request when the recording is already complete. Use streaming when captions need to appear while someone speaks. A streaming result can include provisional words that change as more audio arrives. Show those as draft captions and save finalized portions after your app has handled them. The app still needs the audio source and permission to capture or upload it.

What should you check before sending or saving audio?

Deepgram’s cloud service receives the audio you submit. Its data guide says participation in model improvement is the default, and documents a per-request opt-out for speech-to-text. The opt-out covers content retention after processing under the documented policy. Usage and request metadata can still be recorded. Confirm the setting before sending private material. Save the returned text and word times where your app needs them, with your own access and deletion rules. The normal transcription response is not an archive you can reopen later.

What should you count to estimate transcription cost?

The pricing page separates streaming from prerecorded speech-to-text and lists rates by model and language option. It counts processed audio duration across channels. Its example treats ten minutes of stereo audio as twenty processed minutes. Estimate from the actual recording format and the price for your model, plan and endpoint. Check the price impact of opting out before relying on a quote.

How would you check that a provisional caption can change?

Use paper and an invented recording. Do not connect a microphone, vendor account or real recording.

  1. Draw a draft and final line
    Pretend the recording says “room fifty.” Write a provisional “room fifteen” and a final “room fifty.” Show the first as draft text, then replace it with the final line and check it against the intended speech.
  2. Draw a missing final result
    Leave the final result blank for a second recording. Decide what the app should show while it has nothing confirmed to save. This does not test provider behavior.