Soniox

60+ languages with translation-grade accuracy; realtime bills wall-clock session time.

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
`soniox/stt-rt-v5`	streaming	$0.12/session-hr	✓	✓
`soniox/stt-async-v5`	batch	$0.1/hr	✓	✓

Provider options

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

Param	Type	Default	Applies to	What it does
`language_hints_strict`	boolean	`false`	streaming · batch	Restrict recognition to the hinted languages
`enable_language_identification`	boolean	`false`	streaming · batch	Tag tokens with the detected language
`max_endpoint_delay_ms`	integer	`2000`	streaming	Upper bound on endpoint latency
`endpoint_sensitivity`	number	—	streaming	Endpoint eagerness; docs suggest 0.3 for voice agents
`endpoint_latency_adjustment_level`	integer	—	streaming	Latency/accuracy trade for endpoint detection (2 = voice-agent preset)
`context`	object	—	streaming · batch	Domain context: general text, terms, translation_terms (≤8k tokens)
`translation`	object	—	streaming	Live translation config (one_way or two_way)
`client_reference_id`	string	—	streaming · batch	Your correlation id, echoed in Soniox logs

Try it

curl -s https://api.speechrouter.ai/v1/audio/transcriptions \
  -H "Authorization: Bearer $SPEECHROUTER_API_KEY" \
  -F model=soniox/stt-async-v5 \
  -F file=@audio.wav

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