Keyword Boosting

General-purpose speech recognition can miss words it rarely hears, such as brand names, product names, or industry terms. Keyword boosting tells the model to favour specific words and phrases for one connection.

Usage

Pass a comma-separated list in keywords, and optionally set the boost strength with alpha:

wss://in.api.murf.ai/v1/transcription/stream-input
    ?api_key=YOUR_API_KEY
    &language=hi
    &keywords=Murf,Falcon,voice%20clone
    &alpha=0.6

How boosting works

• Boosting biases recognition; it does not force words into the transcript. A keyword appears only when the audio supports it.
• Each keyword is boosted only in its own language. A Hindi term is not boosted against English audio, and vice versa.
• alpha applies to every keyword in the list. Values below 0.0 or above 1.0 are treated as 0.0 and 1.0.

Choosing keywords

• Boost what the model is likely to miss. Brand names, product names, plan names, and proper nouns benefit most. Common words don’t need boosting.
• Write keywords the way they should appear in the transcript.
• Keep the list short and specific to the call. A focused list for each use case works better than one long list for everything.
• Use multi-word phrases for terms that are always said together, for example voice%20clone.

Choosing alpha

Start with the default of 0.5. Very high values can make boosted words appear where the caller said something similar but different. Increase it if boosted terms are still being missed. Lower it if boosted terms start appearing where the caller said something else.

Test with recordings of real calls before changing alpha in production.

Example: a collections voice agent

?language=hi&keywords=EMI,NACH,autopay,Murf%20Finance&alpha=0.6

Keywords are set when you connect and apply to every turn on that connection, in both partial and final transcripts. To change them, open a new connection.