Transcriber
Voice Agents use STT (Speech to Text), to transcribe what the user speaks. This transcribed speech as text goes into an LLM to generate the response that gets played out to the user.
Paladin platform ships with Deepgram, Cartesia, OpenAI and Paladin transcribers by default. You can take a look at the providers documentation of which language to select for your language requirements.
Example: Deepgram has their language support documentation at https://developers.deepgram.com/docs/models-languages-overview#nova-3
What this controls#
The transcriber converts user speech into text that the agent can reason over. A poor transcription setup can make a correct workflow behave incorrectly, so this setting should be tested with real accents, background noise, and call conditions.
Setup checklist#
- Select the provider and model that best match your expected audio quality and languages.
- Test with short answers, interruptions, names, numbers, and domain-specific words that matter to the workflow.
- Review transcripts after test runs to identify repeated misheard words or missing context.
- Tune prompts and variables only after confirming the speech-to-text layer is capturing the caller correctly.
Common issues to watch#
- Names, addresses, and account identifiers may need confirmation prompts in the workflow.
- Very noisy calls can reduce recognition quality even when the workflow instructions are correct.
- If a caller switches languages, test the full path rather than assuming every model handles the switch the same way.