What is the fastest speech-to-text API?

For live audio, the fastest speech-to-text API in the 2026 Artificial Analysis streaming benchmark was Deepgram Flux. It returned its first partial transcript 0.019 seconds after the end of speech, and its final transcript at 0.020 seconds ^{1 2}. That is a benchmark result on one set of audio, and speed is only one of three things to weigh. The fastest service was not the most accurate, and "fastest" means something different for captions, voice agents and batch files.

Deepgram Flux led on speed, but not on accuracy

The same benchmark shows the trade-off. Deepgram Flux had a word error rate of 7.4 percent on its first partial transcript, while ElevenLabs Scribe v2 Realtime had 3.6 percent ³. ElevenLabs took 0.13 seconds to its first partial ⁴.

So the choice depends on the job. For a voice agent that must reply quickly to a short phrase, a fast first word may matter more. For a transcript that people will read, a lower error rate usually matters more, and a slower first word costs little.

The ranking changes with the measure and the audio

The Artificial Analysis test covers several audio sets, and its authors say that "performance varies meaningfully across the three datasets with different audio lengths, accents, vocabulary, and background noise" ⁵. A service that leads on one dataset can fall behind on another, so a single ranking is a starting point, not a verdict.

Speed also means different things for different products:

For batch work, speed means throughput. On the Open ASR Leaderboard, NVIDIA's Parakeet TDT 0.6B v2 had an inverse real-time factor of 3,390 ⁶. That number describes how quickly a model works through long audio, and it is measured on the leaderboard's setup, not on a live call. A model that processes a file quickly can still take a while to produce its first word, which is why the live and batch numbers should not be mixed.

Vendor claims are not the same as an independent test

Vendors publish their own speed figures. Deepgram says its Nova-3 model delivers streaming latency below 300 milliseconds ⁷. That is a claim from the company that sells the service. It may be true on the company's own test setup, but it does not show how the service performs on your microphone, your room or your accent. The Artificial Analysis test is independent, but it is also one test. Treat both as leads to check, not as answers.

How to choose the fastest API for your use

1. Name the job. Captions need the first partial. A voice agent needs the time to end-of-turn. Dictation needs the time to final text, plus the errors in it.
2. Use your own audio. Record a few minutes in the room and with the microphone you will use, including the names and terms you need. Benchmarks use their recordings, not yours.
3. Time the wait, not the server. Measure from when you stop speaking to when the final text is on screen, on your own network. The Artificial Analysis clock starts when its voice activity detector reports the end of speech ⁸, so its numbers leave out the time that detector takes to decide. Your test should include that time.
4. Score the errors. Count word errors with word error rate on the same clips for each service, and compare the fast services with the accurate ones.
5. Check the price and data terms. A fast service that bills per minute, or keeps audio, may cost more than a slightly slower one. The guide to speech-to-text APIs compared covers price and data terms side by side.

If you would rather not send audio to a cloud API, look at on-device recognition. Yap, which we make, is an on-device dictation app for Mac and Windows. See the download page for current availability.

The short answer is Deepgram Flux for first-word speed in this 2026 test, with a cost in accuracy. The useful answer is the one you get by timing two or three services on the audio you actually use.

For how a live system decides that you have stopped talking, which affects the wait you feel, read the question on how speech recognition knows when you stop talking. For the Whisper side of live use, read whether Whisper can transcribe in real time.

For the wider field, read our overview of speech recognition.