Access Bolna playground from https://platform.bolna.ai/.
Transcriber configuration tab in Bolna Playground showing speech-to-text provider selection and language settings for Voice AI agents

Transcriber Tab on Bolna Playground

  1. Choose your Transcriber Provider and model
    • Deepgram (Default transcriber, most tried and tested)
    • Whisper (open-source, cheapest)
  2. Language - By default the agent can only transcribe English language. By choosing any other language, the agent will be able to transcribe sentences spoken in chosen language + English
  3. Endponting - Number of milliseconds your agent will wait before generating response. Lower endpointing reduces latency could lead to agent interrupting mid-sentence. If you want quick short responses, keep a low (100ms) endpoint.
    If you are expecting users to speak longer sentences, keep a higher (500ms) endpoint.
  4. Linear Delay - Linear delay accounts for long pauses mid-sentence. If the recipient is expected to speak long sentences, increase value of linear delay.
  5. Interruption settings - Agent will not consider interruption until human speaks these number of words. Ideal to prevent Agent pausing when human is actively listening by saying Oh, yes etc.
    (If the user says a Stopword, such as stop, wait, etc., the agent will automatically pause regardless of the settings)
  6. Backchanneling - Switch on only if user is expected to speak long sentences. Agent will show they are listening by give soft verbal nudges of acknowledgement.
    You can change the time to wait before the agent gives the first filler, as well as the time between subsequent fillers.