Deepgram: artificial intelligence features and setup

Overview

When adding voice to a product, transcribing a file after a call ends is different from producing a response that needs to be heard in real time. Deepgram offers components for both types of work, so choosing an interface and model should start with the desired outcome. In a conversational product, it's especially important to check when the system decides that the user has finished speaking and how a transcription error carries into the response layer.

Features

• Real-time and file transcription
• Speech generation from text
• Voice Agent API for connecting conversation components

Business uses

• Transcribing calls to classify service topics
• A voice layer in a product that takes action based on a conversation

Getting started

• Submit a familiar audio sample and compare the transcription with a manual transcript.
• Run the same speech sample as a file and as a real-time stream, and compare numbers, technical terms, and the time until usable text is available.
• If also using voice generation, interrupt the agent mid-sentence and check that an old response does not continue after the user corrects the request.

What to check

• Check Hebrew support separately for each STT and TTS model you selected.
• Measure latency, end-of-speech detection, and interruptions in conversation mode, not only in file processing.