Skip to content

Rumble-2-Flash is now open weights, Apache 2.0.

Read the announcement

Custom Models

A custom model is a collection of phrases that biases transcription toward domain-specific vocabulary, such as personal names, product names, technical terms, or internal jargon.

Custom models are available on the Prerecorded, Realtime, and Turns APIs.

Start with a base model on its own, without a custom model or phrases. Rumble-2 handles everyday speech and a broad range of terminology well, and for many use cases it needs no customisation at all.

Consider a custom model when you notice that:

  • Specific phrases are consistently transcribed incorrectly, such as product names, technical terms, abbreviations, or jargon from your field.
  • Your use case relies heavily on names, such as people, companies, medications, or places, where an uncommon spelling matters.

If the base model already gets your terminology right, a custom model adds little, and an unfocused one can make results worse (see Choosing Phrases). For just a handful of terms, per-request phrases may be all you need.

A custom model can be created through the console or through the API. Models created in the console also show up in the API, so they can be inspected and managed there like any other model. There is no limit on the number of custom models you can create.

The console offers two options:

  • Enter the phrases yourself, if you already know which terms you need.
  • Generate them automatically. Provide documents, a description of your use case, or other context, and the relevant phrases are looked up in your documents or on the web to build the model for you.

Generating phrases is a good fit when you know transcription of your domain could be better, but don’t have a ready-made list of terms.

The API expects a complete, hand-picked list of phrases: you decide exactly which terms go into the model. This gives precise control and lets you keep the model up to date programmatically as your vocabulary evolves. See the API reference to create a model, add, list, and delete phrases, and to delete the model once you no longer need it.

For a small number of phrases, you don’t need a custom model at all: pass them directly with the request via the phrases query parameter, available on all three APIs. This is limited to 250 phrases per request - use a custom model when the vocabulary is larger or reused across requests.

A custom model can be applied in two ways:

  • Configured on an API client in the console, so every request made with that client uses the model by default.
  • Passed per request, via the custom_model_id query parameter on the Prerecorded, Realtime, and Turns APIs:
?custom_model_id=<id>

When both are set, the query parameter takes priority over the model configured on the API client.

The Prerecorded, Realtime, and Turns APIs accept an optional control:

  • bias_strength controls how strongly the selected phrases influence transcription. Its default is 0.45; larger values apply stronger biasing.

The default is suitable for most requests. Increase it only when the expected terminology is not being recovered, since stronger biasing can cause similar-sounding words to be transcribed as your phrases (see Choosing Phrases).

Good phrases are the terms the base model gets wrong: specialised vocabulary, uncommon names, brand and product names. Everyday words are already recognised well and only dilute the model. A phrase can be at most 255 characters long, and a model can currently hold at most 50,000 phrases; we expect to raise this limit over time.

For short structured tokens like order codes and licence plates, use Patterns instead; they describe the shape of a token rather than a fixed phrase. Note that patterns and phrase biasing are mutually exclusive - patterns cannot be combined with phrases or a custom model.