Which model should I use?

Twelve providers is a lot of menu. Start from the job:

Live voice agents

Pick	Why
soniox/stt-rt-v5	Semantic endpointing (pause + intonation), 60+ languages, $0.12/session-hr
deepgram/nova-3	Lowest cost with tunable endpointing, huge encoding support
assemblyai/universal-3-5-pro	Turn model built for agents, min_latency mode
cartesia/ink-2	Server-driven turns with an eager-end signal for LLM head-start

Always add a fallback from a different provider — agents must not go deaf.

Live captions / dictation

deepgram/nova-3 for English speed and price; soniox/stt-rt-v5 when the room might switch languages mid-sentence; elevenlabs/scribe_v2_realtime for 90+ languages at ~150 ms.

Batch: accuracy first

openai/gpt-4o-transcribe (LLM-grade), assemblyai/universal-2, or speechmatics/enhanced (strong on accents; medical tier).

Batch: price first

Pick	Price	Catch
groq/whisper-large-v3-turbo	$0.04/hr	10s minimum/request, no diarization
speechmatics/melia-1	$0.129/hr	batch only
openai/gpt-4o-mini-transcribe	$0.003/min	no word timestamps
mistral/voxtral-mini	$0.003/min	13 languages; 3h files in one request

Multilingual / code-switching

speechmatics/melia-1 (56 languages, switches mid-utterance, auto-detect), soniox/stt-rt-v5 (streaming), assemblyai/universal-streaming-multilingual (streaming, en/es/fr/de/it/pt).

Speaker labels

Word-level: deepgram/nova-3, soniox/stt-rt-v5, assemblyai/universal-2, speechmatics/*, elevenlabs/scribe_v2. Turn-level with enrollment (name known speakers): openai/gpt-4o-transcribe-diarize. Streaming diarization billed separately on Azure: azure/conversation-transcription.

Telephony (8 kHz mulaw)

deepgram/nova-3, soniox/stt-rt-v5, assemblyai/universal-3-5-pro, aws/transcribe-streaming — all take mulaw natively. Recipe: Transcribe phone calls.

When in doubt

Start with deepgram/nova-3 + fallbacks=soniox/stt-rt-v5, ship, and let real traffic decide. Switching later is a one-line change — that’s the point.