deepgram
Deepgram — Speech-to-Text and Voice API
Overview
Deepgram is a hosted speech platform with four parts: speech-to-text (REST for recorded audio, WebSocket for live audio), text-to-speech (Aura-2), text intelligence (sentiment, topics, summaries, intents) and a Voice Agent API that chains listen, think and speak. It needs an API key from the Deepgram console; there is no offline mode in the SDKs.
Choose the speech-to-text model by the job:
| Model | Use it for |
|---|---|
nova-3 |
General transcription of recordings or live audio: meetings, captions, noisy or multilingual audio. Highest accuracy, no turn detection. Variants: nova-3-medical, nova-3-pharma. |
flux-general-en, flux-general-multi |
Voice agents and live conversation: the model itself detects end of turn. Streaming only, on the /v2/listen endpoint. multi covers English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian, Dutch. |
nova-2 |
Languages Nova-3 does not support yet, and filler-word detection. Older, kept for compatibility. |
All models default to language=en; pass language explicitly (fr, es, ...), detect_language=true for one unknown language, or language=multi for code-switching audio on Nova-3.
The SDKs were rewritten: JavaScript @deepgram/sdk 5.x (class DeepgramClient; the old createClient and LiveTranscriptionEvents are gone) and Python deepgram-sdk 7.x (the old PrerecordedOptions and dg.listen.rest.v("1") are gone).