For DeepgramAI PlatformTranscription Accuracy

Streaming Stt Websocket

Deepgram Speech AI Platform · Deepgram

Speech AI Platform — Deepgram (Nova STT, Aura TTS, Voice Agent)

Deepgram evals — Streaming STT (WebSocket) (relift v3 InfraRed)

About Deepgram

Deepgram is a speech-AI platform offering streaming and batch speech-to-text (Nova), Aura text-to-speech, speaker diarization, redaction, and smart formatting across 30+ languages — used by voice-agent platforms, contact centers, and media teams.

Employees

~150

Industry

Speech AI

Headquarters

San Francisco, CA

Website

deepgram.com

Sample tests· showing 3 of 9

#	Input	Expected behavior	Check
01	Agent opens wss://api.deepgram.com/v1/listen?model=nova-3&encoding=linear16&sample_rate=16000 with header 'Authorization: Token <DEEPGRAM_API_KEY>'.	Send the Authorization header in the WebSocket upgrade request — Deepgram uses 'Token <key>' scheme, not 'Bearer'. Confirm 101 Switching Protocols before streaming audio frames. Treat 401/403 as a credential issue and surface to operator; do not retry the same key on auth failure.	Pass / FailAi Platformcritical
02	Connection uses interim_results=true. Server emits a sequence of is_final=false hypotheses followed by an is_final=true segment.	Render interim (is_final=false) results to a transient overlay; commit only is_final=true messages to the persisted transcript. Do not double-commit when the final arrives — replace the matching interim span, not append.	Pass / FailAi Platformhigh
03	Agent connects with model=nova-3-general and a future release shifts the default 'nova-3' alias to nova-3.5.	Pin the explicit model variant (e.g., nova-3-general or nova-3-medical) to freeze behavior; treat unsuffixed aliases (nova-3) as floating and unsafe for regression-tested production. Track Deepgram changelog for variant deprecation.	Pass / FailAi Platformmedium
Use this eval