Select a representative clip
Choose a recording with the language, microphone quality and speaker overlap you expect in production. Make a human-checked reference transcript; preserve uncertain words rather than silently guessing them.
Model selection
How to evaluate Deepgram speech-to-text models
Deepgram speech-to-text models are easier to assess when you start with your audio and the transcript you need. This guide covers prerequisites, a sample test and the trade-offs to check before committing to a workflow.
Prerequisites
Gather representative audio before comparing Deepgram results. A clean studio clip alone will not reveal how a model handles the conditions in your actual workflow.
One full run-through
Treat a model comparison as a repeatable test, not a judgment based on the first transcript that looks readable.
Run the same audio through each candidate
Record the Deepgram model and request settings used for each result. Keep preprocessing consistent, and do not compare a cleaned file against an untouched version of the same recording.
Inspect consequential errors
Compare names, numbers, negations and speaker turns against the reference. Note whether the transcript needs editing, then repeat the test with several clips before drawing a conclusion.
Options table
These are two evaluation paths, not claims that every Deepgram model supports identical settings. Confirm current model availability and request options in the official documentation.
Recorded-audio evaluation
Live-audio evaluation
1
Source
Recorded-audio evaluation
A saved file or recording
Live-audio evaluation
Audio arriving during a session
2
Primary question
Recorded-audio evaluation
How accurate is the completed transcript?
Live-audio evaluation
How useful are results while speech is still arriving?
3
Reference
Recorded-audio evaluation
Human-checked transcript of the full clip
Live-audio evaluation
Human-checked transcript plus timing notes
4
Errors to inspect
Recorded-audio evaluation
Missing words, names and speaker turns
Live-audio evaluation
Revisions, missed words and delayed final text
5
Network conditions
Recorded-audio evaluation
Document transfer and processing conditions
Live-audio evaluation
Test interruptions and variable connectivity
6
Decision criterion
Recorded-audio evaluation
Editing effort on the finished result
Live-audio evaluation
Usability of interim and final results
What fails
Avoid a misleading model comparison
A single clean clip, inconsistent request settings or an unverified reference can make Deepgram results look better or worse than they are. Test the audio that matters to you, and verify current capabilities before designing around a particular model. The link below leads to a separate voice product, not a Deepgram model selector.
- A polished example cannot stand in for noisy production audio.
- A readable transcript can still get names or negations wrong.
- A recorded-audio result does not establish live-audio performance.
FAQ
Begin with the kind of audio you need to transcribe and the language it contains. Check the current Deepgram documentation for eligible models, then test representative clips against a human-checked reference using consistent settings.
Not necessarily for every recording or workflow. Compare the candidates available for your use case on your own audio, paying particular attention to errors that would change the meaning of a transcript.
You can reuse source material and a reference transcript, but the evaluation questions differ. For live audio, also inspect interim results, revisions and when usable final text arrives.
Review names, numbers, negations, speaker changes and the amount of human correction required. If timing matters, assess it separately rather than relying on a single accuracy impression.
No. This is an evaluation guide, and its voice-tool link leads to a separate product. Consult Deepgram's current documentation and supported interfaces to confirm and run a specific model.