AirCaps AI

A Deepgram Flux alternative for voice agents.

AirCaps Live 1 returns the final transcript a median 38 ms after the caller stops speaking, against 98 ms for Deepgram Flux, and makes 51% fewer word errors (3.65% against 7.48% WER on 9,205 voice-agent clips). Its built-in turn detection ends 8× fewer turns while the caller is still mid-thought (3.6% against 28.8%). It costs $0.20 per audio hour; Deepgram Flux lists at $0.39.

38 msmedian time to final transcript; Flux takes 98 ms, and 3.2× as long at p95
3.65%word error rate on 9,205 voice-agent clips; Flux 7.48%
8×fewer callers cut off mid-thought than Flux
$0.20per audio hour; Flux lists at $0.39

Side by side.

The same audio, streamed in real time to every model. The better result in each row is highlighted.

What changes when you switch.

Fewer word errors

3.65% WER on 9,205 voice-agent clips, against 7.48% and 6.28% for Flux and Nova-3. Live 1 is also lower on each of the four public sets in the table.

A sooner final transcript

A median 38 ms from the end of speech to the final transcript, 52 ms at p95. Flux takes 98 ms, and 168 ms at p95.

Fewer callers cut off

On Pipecat's Smart Turn v3.2 set, Live 1 ended 3.6% of unfinished turns too early; Flux ended 28.8%. Live 1 also calls the end of a turn sooner: 111 ms against 383 ms.

A lower price

$0.20 per audio hour, with 5 free hours on every account. Flux lists at $0.39.

Questions.

Is AirCaps Live 1 a good alternative to Deepgram Flux?

AirCaps Live 1 returns the final transcript a median 38 ms after the caller stops speaking, against 98 ms for Deepgram Flux, and makes 51% fewer word errors (3.65% against 7.48% WER on 9,205 voice-agent clips). Its built-in turn detection ends 8× fewer turns while the caller is still mid-thought (3.6% against 28.8%). It costs $0.20 per audio hour; Deepgram Flux lists at $0.39.

How do I switch from Deepgram to AirCaps Live 1?

Stream 16 kHz mono PCM to the Live 1 WebSocket and read the transcript messages. Turn on end_of_turn to let the model end turns, or send finalize from your own VAD; both work in the same session. The AirCaps Live 1 guide at aircaps.ai/docs/guides/live-1 has the whole protocol.

How does its turn detection compare with Deepgram Flux?

On Pipecat's Smart Turn v3.2 English set (5,367 clips, 2,718 of them unfinished turns), Live 1 ended 3.6% of unfinished turns too early and 93.5% of its turn endings were right; Deepgram Flux ended 28.8% too early, with 64.0% right. On LiveKit EOT Bench (7,655 turns), Live 1 called the end of a turn a median 111 ms after the caller stopped; Deepgram Flux took 383 ms.

How was this measured?

9,205 English clips (154,661 words) from Pipecat's STT benchmark, Coval's public set, LiveKit EOT Bench, WildASR and PriMock57, run on November 19, 2026. Every model streamed the same audio in real time from one machine in AWS us-east-2. Time to final transcript runs from the end of speech, when the last audio and the finalize message have been sent, to the final transcript, network included.

How much does AirCaps Live 1 cost?

$0.20 per hour of audio you stream, billed per millisecond, with 5 free hours on every account. Deepgram Flux lists at $0.39; Deepgram Nova-3 lists at $0.29.

Does it work with Pipecat and LiveKit?

Yes. It runs over a WebSocket, so it fits a Pipecat pipeline or a LiveKit agent as the speech-to-text step.

Which languages does it support?

English, with multilingual support coming soon.

Try it on your calls.

AirCaps Live 1 is on every account, with 5 free hours. Read the AirCaps Live 1 guide.