Skip to content

Deepgram

The realtime workhorse — fast, cheap, excellent English.

All prices are the vendor’s public list price — 0% markup. Extra provider knobs pass through untouched via provider_params.

Model Modes List price Diarization Word timings
deepgram/nova-3 streaming · batch $0.0048/min
deepgram/nova-3-medical streaming · batch $0.0048/min
deepgram/nova-2 streaming · batch $0.0048/min
deepgram/flux-general-en streaming $0.00648/min
deepgram/flux-general-multi streaming $0.0078/min

Reach past the unified surface with provider_params — forwarded streaming → WebSocket query parameters; batch → query parameters. Typed in the SDKs as {provider}Params interfaces (providerParams option / provider_params= kwarg).

Param Type Default Applies to What it does
smart_format boolean false streaming · batch Format numbers, dates, currency in the transcript
punctuate boolean false streaming · batch Add punctuation and capitalization
endpointing integer 10 streaming Silence (ms) before the endpointer fires
utterance_end_ms integer streaming Gap (ms) that triggers UtteranceEnd; needs interim results on
vad_events boolean false streaming Emit speech-started events from the voice-activity detector
diarize_model string streaming · batch Diarization model variant (replaces the deprecated diarize flag)
multichannel boolean false streaming · batch Transcribe each audio channel independently
keywords string streaming · batch word:boost pairs — nova-2 family only (nova-3 uses keyterms)
eot_threshold number 0.7 streaming Flux models: end-of-turn confidence threshold
eager_eot_threshold number streaming Flux models: earlier, lower-confidence end-of-turn signal
eot_timeout_ms integer 5000 streaming Flux models: force end-of-turn after this silence
Terminal window
curl -s https://api.speechrouter.ai/v1/audio/transcriptions \
-H "Authorization: Bearer $SPEECHROUTER_API_KEY" \
-F model=deepgram/nova-3 \
-F file=@audio.wav

Generated from the gateway catalog — the billing engine’s own source of truth.