OpenAI

Whisper and the GPT-4o transcribe family — LLM-grade accuracy.

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
openai/gpt-realtime-whisper	streaming	$0.01698/min	—	—
openai/whisper-1	batch	$0.006/min	—	✓
openai/gpt-4o-transcribe	streaming · batch	$0.006/min	—	—
openai/gpt-4o-mini-transcribe	streaming · batch	$0.003/min	—	—
openai/gpt-4o-transcribe-diarize	batch	$0.006/min	✓	—

Provider options

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

Param	Type	Default	Applies to	What it does
prompt	string	—	batch	Context/spelling hints for the transcription
temperature	number	—	batch	Sampling temperature
timestamp_granularities[]	word · segment	—	batch	whisper-1 with verbose output only
chunking_strategy	string	—	batch	Required by the diarize model for audio over 30s (“auto”)
known_speaker_names[]	array	—	batch	Diarize model: enroll up to 4 named speakers
known_speaker_references[]	array	—	batch	Reference audio for enrolled speakers
delay	minimal · low · medium · high · xhigh	—	streaming	gpt-realtime-whisper: latency/quality knob
turn_detection	object	—	streaming	server_vad {threshold, silence_duration_ms} or semantic_vad {eagerness}

Try it

curl -s https://api.speechrouter.ai/v1/audio/transcriptions \
  -H "Authorization: Bearer $SPEECHROUTER_API_KEY" \
  -F model=openai/whisper-1 \
  -F file=@audio.wav

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