How can I enable NER formatting in my self-hosted deployment?

Deepgram's NER formatting provides a sophisticated, context-rich method of formatting transcripts so they can be displayed to end users or passed to downstream consumers.

NER stands for Named Entity Recognition. Deepgram's NER model recognizes entities in transcripts and formats them how a human may write them.

For instance, if someone says "my phone number is six zero eight four one five two zero nine seven", NER formatting would yield a transcript of "My phone number is [phone]."

Nova-3 requires NER formatting, since formatted models do not exist.

Nova-2 can use NER formatting when combined with its non-formatted models.

NER formatting can be enabled in your self-hosted deployment by following a few steps below.

Updating your self-hosted deployment

1. Update your container versions to the March 2025 self-hosted release (release-250307).

2. Add feature flags in your TOML files to enable NER formatting.

   1. For streaming NER, edit engine.toml:

      [features]
      streaming_ner = true

   2. For pre-recorded NER, edit api.toml:

      [features]
      format_entity_tags = true

3. In your models directory, add non-formatted Nova-3 or Nova-2 models, plus the "entity detector" model(s) for your transcription mode(s) (streaming and/or pre-recorded audio). Ask your Deepgram account representative to provide you with these new models.

   1. For streaming:

      1. Add the streaming entity detector model: entity-detector.en.streaming.ade60a4c.dg.

      2. Add non-formatted streaming transcription models. For Nova-3, add nova-3-general.en.non-formatted.streaming.bf05427e.dg.

   2. For pre-recorded:

      1. Add the batch entity detector model: entity-detector.batch.06bc8f36.dg.

      2. Add non-formattted batch transcription models. For Nova-3, add nova-3-general.en.non-formatted.batch.3b3aabe4.dg.

4. Start or restart your containers so that the changes take effect. e.g. for Docker: Restart (if your containers are running): sudo docker compose restart Start (if your containers are stopped): sudo docker compose up -d