Gyan

Topic · updated Monday, 12 October 2026

Who leads in audio AI: Deepgram or ElevenLabs?

Neither one leads the whole field. ElevenLabs is the bigger company and wins on how voices sound. Deepgram is the faster, cheaper engine that developers wire into phone bots. And the giants are now on the same leaderboards as both.

What the question means

“Audio AI” is really two jobs, done in opposite directions.

Speech-to-text (STT) turns sound into words. The jargon for it is ASR, automatic speech recognition. A call centre uses it to write down what a caller said.

Text-to-speech (TTS) turns words into sound. A reading app uses it to speak a book aloud.

Put the two together with a language model in the middle and you get a voice agent: software that listens, thinks, and talks back, like a phone assistant that never puts you on hold.

Deepgram started on the first job, listening. ElevenLabs started on the second, speaking. Over 2025 and 2026 each one built the other half, so now they compete head on. “Who leads” depends on which job you mean, and on whether you measure by money, by accuracy, or by speed.

The answer, in short

As of 12 October 2026:

So: ElevenLabs is the brand people hear. Deepgram is the plumbing people do not notice. Both are real leaders in their lane, and neither is safe.

Do the numbers with your own hands

Take one hour of recorded speech, about 9,000 words.

The names that matter

An analogy

Think of a restaurant kitchen. Deepgram is the industrial stove: fast, reliable, cheap to run, bought by chefs who build their own menus. ElevenLabs is the celebrity chef: the food tastes best, the name draws the crowd, and it now sells stoves too. Which “leads” depends on whether you are a diner or a restaurant owner. And a few supermarket chains, OpenAI, Google and Microsoft, have started selling both stoves and ready meals.

Where the picture is uncertain

Self-test

  1. Which company started with listening, and which with speaking?
  2. A hospital wants to transcribe ten years of recorded consultations overnight. Which measure matters most, and who wins it?
  3. A startup wants the most natural-sounding voice for a bedtime-story app. Who wins, by which evidence?
  4. Why does “within 1-2 points on clean English” not settle the accuracy question?

Answers: 1. Deepgram listening, ElevenLabs speaking. 2. Speed per hour of audio, Deepgram, at 608 times real time. 3. ElevenLabs, because its Eleven v4 Turbo tops the blind-listening arena. 4. Because real calls are not clean English, and the gap widens on noisy, accented or domain-specific audio, which is where each vendor’s own test set favours itself.

Sources