Skip to main content
Updated
Real-time transcription models turn speech into text as you talk, trading off accuracy, latency, price, and how fast they detect when a speaker finishes. Vendor latency claims rarely mean the same thing, so we ranked 15 streaming models on one independent benchmark. Our primary score is final-transcript accuracy from the Artificial Analysis streaming speech-to-text benchmark, normalized to a 0-100 scale where higher is better. It covers English-language audio only, so treat it as a strong baseline, not the last word: telephony, accents, and other languages can shift the order.

​
Best Real-Time Transcription Models

#ModelBest for

​
How to Choose

When choosing between these models, consider:
  • Access: Decide first whether you need a managed API, a first-party app, or a self-hosted model, because that choice drives cost, privacy, latency, and setup work. Most models here are API-only. Only Voxtral Mini and Nemotron 3.5 offer a real self-host route, and Soniox and Azure add a first-party app or portal for trying the model without code.
  • Quality: We use final-transcript accuracy from the Artificial Analysis streaming benchmark, scored 0-100 where higher is better. Watch the partial-versus-final split: a few models (Qwen3, Grok) produce excellent final transcripts but rough live partials, which is invisible in a single accuracy number and matters if users watch text appear as they speak.
  • Price: We compare on price per hour of streaming audio. Streaming costs more than batch, committed and volume tiers swing prices widely, and features like diarization are often billed on top - so confirm the tier and add-ons before you budget.
  • Time to final transcript: This is seconds from the end of speech to the final transcript; lower matters most for voice agents, where the practical target is a sub-500ms end-to-end response. Raw latency is only half of it - how well a model detects that a speaker has finished (its endpointing) shapes the felt responsiveness just as much.
For most teams building voice agents, start with Soniox v5 for value, Cartesia Ink 2 or Deepgram Flux when turn-taking is the hard part, and ElevenLabs Scribe v2 when transcript quality outweighs everything. For private or offline deployments, Voxtral Mini and Nemotron 3.5 are the two open-weight picks worth real testing.

​
Other Models We Considered

​
Frequently Asked Questions

For pure transcript quality, ElevenLabs Scribe v2 Realtime and Cartesia Ink 2 lead on accuracy. But the model most teams should try first is Soniox v5, which pairs near-top accuracy with among the fastest finals at the lowest price on this list.
Soniox v5 Real-Time. It is the rare model that is accurate, fast, and cheap at the same time, and there is a first-party app if you want to try it before writing any code. Move to ElevenLabs or Cartesia only when transcript quality has to be the best available.
Not really. OpenAI’s original Whisper is batch-native - it processes fixed audio chunks, so “streaming” wrappers repeatedly recompute overlapping windows, which is slow and jittery. If you want real streaming, use a streaming-native model such as Voxtral Mini Transcribe Realtime, Nemotron 3.5 ASR Streaming, or Kyutai STT.
Voxtral Mini Transcribe Realtime is the strongest open-weight pick, though it needs a capable GPU. Nemotron 3.5 ASR Streaming is smaller but still needs the supported high-end NVIDIA stack. For ordinary-machine or phone deployments, Kyutai STT and Moonshine v2 are the more practical options.
It depends on the hard part. If turn-taking and endpointing are what break your agent, Deepgram Flux fuses transcription with end-of-turn detection and delivers the fastest finals. If you want accuracy plus native endpointing, Cartesia Ink 2 is excellent. For the best overall value, Soniox v5.
They are built for different jobs. Flux is tuned for conversational turn-taking and the fastest possible finals, which is what voice agents need. Nova-3 is the general-purpose streaming workhorse and scores higher on final-transcript accuracy in our benchmark. Choose Flux for responsiveness, Nova-3 for broader transcription.
Treat them as a strong starting point, not a guarantee. The benchmark covers English-language audio and does not directly measure 8kHz telephony, multilingual quality, diarization, or full voice-agent latency, so your own workload can reorder the results. Always test the shortlist on your own audio before committing.