Best speech-to-text for voice agents in 2026

Streaming models compared

Word error rates come from public benchmarks on conversational audio and are not directly comparable between vendors; treat a two-point gap as noise and a ten-point gap as real.

Why end-of-turn detection is the shortcut

A cascaded agent normally waits for silence, then runs a turn detector, then waits for the STT model to finalise. Models with native end-of-turn detection, such as Deepgram Flux (English), fuse those steps: the transcript is final at the moment the model decides the caller is done, which removes 200 to 400 ms from every turn. That is why the balanced and lowest-latency stacks on Voiceflint start from Flux even though it is not the cheapest line.

Accuracy first

For calls with names, addresses and order numbers, keyterm prompting matters more than a point of WER. Deepgram Flux (English), Deepgram Flux (Multilingual), Deepgram Nova-3, AssemblyAI Universal-3.5 Pro Realtime accept a keyterm list, and Voiceflint passes the agent's keyterms through automatically. Medical and financial vocabularies deserve a model tuned for them; the receipt shows exactly what that costs per minute.

Multilingual calls

The multilingual stack uses Soniox Real-time v5 at $0.0020 per minute. Code-switching mid-sentence is still the hardest case; if callers switch between two known languages, set both on the agent so the model can bias toward them.

Cheapest that still works

AssemblyAI Universal-Streaming anchors the lowest-cost stack at $0.0025 per minute. It has no native end-of-turn detection, so the stack pairs it with a separate turn detector and accepts roughly 150 ms more per turn in exchange for the saving.

Questions

Can I bring my own speech-to-text key?

Yes. Add the provider key under Keys, switch the agent to your keys, and the STT line on the receipt drops to $0 while the provider bills you directly.

Does Voiceflint pick the STT model for me?

Only if you ask. Presets are measured defaults you can change per agent; the model picker shows price, latency and accuracy side by side so the trade-off is explicit.