Soniox has released Soniox v5 Real-Time: a low latency streaming Speech to Text model on the Pareto frontier for accuracy and latency, at the lowest price of any proprietary model tested

Soniox v5 Real-Time is @soniox_ai's latest streaming Speech to Text (STT) model, joining Soniox v5 Async, their non-streaming model released last week. On AA-WER Streaming it occupies the middle of the Pareto frontier: faster than the most accurate models (Cartesia Ink-2, ElevenLabs Scribe v2 Realtime) and more accurate than the fastest (Deepgram Flux, Nova-3), while at a lower price than all of them.

AA-WER Streaming Overview

AA-WER Streaming reports WER and latency as a pair, measured from Silero VAD-detected end of speech on the same ~8 hours of audio as our non-streaming STT benchmark, AA-WER v2.0. We report both at two points: First Final (first final-denoted transcript, best for accuracy) and First Partial (first transcript-bearing event, best for when speed matters most).

Key takeaways

➤ First Final Transcription: Soniox v5 Real-Time achieves a 4.5% WER at 0.05s after end of speech, more accurate than the faster Deepgram Flux (7.4%, 0.02s) and Deepgram Nova-3 Realtime (6.7%, 0.06s), and faster than the more accurate Cartesia Ink-2 external endpoints (3.7%, 0.09s) and ElevenLabs Scribe v2 Realtime (3.6%, 0.14s)

➤ First Partial Transcription: The model achieves a 4.7% WER at 0.05s after end of speech, behind only Cartesia Ink-2 external endpoints (4.3%, 0.07s) and ElevenLabs Scribe v2 Realtime (3.6%, 0.13s) on accuracy, while faster than both

➤ Price: The model costs $2 per 1,000 minutes representing the lowest of any proprietary streaming model tested, below Cartesia Ink-2 ($4), Deepgram Nova-3 Realtime ($4.80) and ElevenLabs Scribe v2 Realtime ($6.50)

➤ Language support: The model supports over 60 languages, providing language identification and real-time translation across multilingual conversation.

See more details below ⬇️