How to Use Deepgram Speaker Diarization

Before you start

Deepgram speaker diarization identifies which person is speaking at each point in an audio or video recording. The process runs as a speech service rather than locally on your computer, so you need an account, an API key, and a supported audio or video file. You do not need a specific input device: a file recorded with a camera, phone, USB microphone, or meeting app can be used. For the clearest labels, start with a recording where only one person talks at a time and keep background noise low.

Check the audio properties before uploading. A higher sample rate, such as 16 kHz or above, is suitable for speech, but Deepgram accepts common recording formats regardless of the original sample rate. Set the input level so speech is clear but not clipped. Diarization does not recognize names; it returns speaker markers such as Speaker 0, Speaker 1, and Speaker 2.

Step-by-step instructions

The exact steps depend on whether you use the hosted console, a client library, or an API request. The following procedure covers a standard API call, which gives the most direct control over diarization.

• Prepare an audio or video file with speech from two or more people.
• Enable speaker diarization in the transcription request. In the API, set the diarize parameter to true.
• Include the speech model and any other options you want, such as language or punctuation.
• Send the request with your API key in the authorization header. Keep the key private and do not place it in a shared file or screenshot.
• Inspect the response for a results section containing channels, alternatives, words, and speaker identifiers.
• Use the start and end offsets for each word or phrase to assign the speaker labels over time.

Some client libraries expose diarization as a boolean option when configuring the transcription. Version differences can affect parameter names or where the option is placed, so check the documentation for the version you have installed rather than relying on an older example.

Check the output

Confirm that the transcript contains speaker labels and that each label is attached to the correct word or utterance range. If the recording has two clear voices, the result should normally include at least Speaker 0 and Speaker 1. A short recording or one with long periods of silence may produce a single speaker.

Review a section where people interrupt each other or speak quickly. Diarization marks changes at detected speaker boundaries, but overlapping speech and crosstalk may cause frequent or uncertain switches. The transcript quality depends on the same factors as regular speech recognition: clear speech, a suitable sample rate, low background noise, and consistent input level.

Common problems and next steps

If speaker labels are missing, confirm that the diarize parameter was sent as true and that the file contains more than one voice. If the response returns text but no speaker field, the request may be using an older endpoint or a model configuration that does not include diarization.

If the same speaker is split into multiple labels, try a cleaner recording with less background noise or reduce crosstalk. A quiet input device and consistent distance from the microphone help the service track a voice. If a recording appears to merge two people into one speaker, check whether the voices are similar, whether one person is much louder than the other, or whether the file is mono rather than stereo. The supported fixes are to improve the source recording and rerun the request; real names and speaker identities must be assigned afterward using the returned time ranges.