AirCaps Live 1
The fastest and most accurate speech-to-text for voice agents.
Real-time speech-to-text built for voice agents and phone calls. It knows when the caller has finished speaking and hands your agent the final transcript in 38 ms. It is faster, more accurate and lower-priced than the leading speech-to-text APIs.
Most accurate. Fastest. Lowest price.
Speech-to-text usually makes you trade accuracy, speed and price against each other. AirCaps Live 1 leads on all three, so it sits alone on the Pareto frontier: the leading models from Deepgram, AssemblyAI, ElevenLabs and Microsoft each make more errors, take longer and cost more.
All 9,205 voice-agent clips
Accuracy
3.65% WER
Others make 12–105% more errors
Time to final transcript, p95
Speed
52 ms at p95
Others take 2.3–4.1× as long
Per hour of streaming audio
Price
$0.20 per hour
Others charge 10–170% more
Word error rate and time to final transcript over all 9,205 clips of the November 19, 2026 run described below. Prices are each provider's list price per hour of English streaming audio.
The fastest final transcript.
TTFS, time to final segment, is the time from the moment a caller stops speaking to the moment the final transcript of what they said arrives. It is the speech-to-text share of the silence a caller hears before your agent answers, and AirCaps Live 1 keeps it lower than every leading speech-to-text API.
median time to final transcript, and 52 ms at p95. On those slower turns, models from Deepgram, AssemblyAI, ElevenLabs and Microsoft take 2.3 to 4.1 times as long.
faster than AssemblyAI Universal-3.6 Pro and Microsoft MAI-Transcribe-2 at p95: 52 ms against 211 and 208 ms. Live 1's p95 beats the median of every other model in the chart.
Live 1's p99 is 65 ms. 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. TTFS runs from the end of speech, when the last audio and the finalize message have been sent, to the arrival of the final transcript, network included.
Time to final transcript, p95
95% of turns are final within this time after the end of speech; the tick marks the median. Lower is faster.
The fewest errors on agent calls.
word error rate on 9,205 clips and 154,661 words from five public voice-agent benchmarks, lower than the leading models from Deepgram, AssemblyAI, ElevenLabs and Microsoft.
on accuracy, speed and price against the leading speech-to-text APIs: none makes fewer errors, returns a final transcript sooner or costs less.
Accuracy against speed
Word error rate and p95 time to final transcript, all 9,205 clips.
Lowest error on the public voice-agent benchmarks.
1,000 voice-agent clips
Pipecat STT benchmark
1.52% WER
Others make 9–134% more errors
877 clips
Coval public STT set
1.12% WER
Others make 14–173% more errors
400 human-to-agent turns
LiveKit EOT Bench
7.02% WER
Others make 8–28% more errors
798 turns, accented English
LiveKit EOT Bench, accented
5.23% WER
Others make 12–72% more errors
The leading models, next to Live 1.
| Model | More word errors than Live 1 | Slower to a final transcript | Price per audio hour | Turn detection | ||||
|---|---|---|---|---|---|---|---|---|
| All clips | Coval | Pipecat | LiveKit | LiveKit, accented | TTFS at p95 | List price | Built in | |
| AirCaps Live 1 | 3.65% | 1.12% | 1.52% | 7.02% | 5.23% | 52 ms | $0.20 | Yes |
| ElevenLabs Scribe v3 Live | +12% | +17% | +12% | +28% | +12% | 2.3× | $0.22 | No |
| Microsoft MAI-Transcribe-2 | +13% | +24% | +20% | +8% | +51% | 4.0× | $0.54 | No |
| AssemblyAI Universal-3.6 Pro | +18% | +14% | +9% | +19% | +56% | 4.1× | $0.45 | Yes |
| AssemblyAI Universal-3.5 Pro | +21% | +28% | +23% | +22% | +62% | 3.0× | $0.45 | Yes |
| ElevenLabs Scribe v2 Realtime | +23% | +25% | +22% | +26% | +17% | 2.8× | $0.22 | No |
| Deepgram Nova-3 | +72% | 2.4× | 2.1× | +15% | +33% | 2.5× | $0.29 | No |
| Deepgram Flux | 2.0× | 2.7× | 2.3× | +17% | +72% | 3.3× | $0.39 | Yes |
The first row is AirCaps Live 1's own result; every other row is measured against it. Pooled word error rate on each benchmark, and p95 time from the end of speech to the final transcript over all 9,205 clips. Prices are list prices per hour of English streaming audio. Turn detection means the model ends turns itself, in the same stream. LiveKit is EOT Bench's general set; LiveKit, accented is its accented-English set.
Built for the turn.
Turn detection is built into AirCaps Live 1: it ends the turn when the caller has finished the thought, judged from what they said as well as how long they paused. Of the other models on this page, only AssemblyAI Universal-3.5 Pro and 3.6 Pro and Deepgram Flux detect turns themselves, so they are the ones compared here.
fewer turns ended while the caller was still mid-thought than Deepgram Flux, and 6 to 7× fewer than AssemblyAI Universal-3.6 Pro and 3.5 Pro.
Turns ended too early · incomplete turns, lower is better
Turn endings that were right · endpoint precision, higher is better
Time to call the end of a turn · median, after the caller stops, lower is better
Turns ended too early and endpoint precision: Pipecat's Smart Turn v3.2 English test set, 5,367 natural clips (2,718 of them incomplete), an endpoint counted within 600 ms of the end of speech. Time to call the end of a turn: LiveKit EOT Bench English, 7,655 turns, each model's own turn detection, timed where the audio is sent. Same recorded streams for every model.
One stream, two answers
The same stream returns the words and the end of the turn, each final transcript flagged with an end-of-turn confidence. No separate turn model to run.
Pick how eager it is
Choose min_latency, balanced or max_accuracy, or tune the threshold and pauses yourself.
Or bring your own
Already run VAD or a turn model? Send finalize when it fires and the final transcript comes back at once. Both work in the same session.
Add it to your agent in a few lines.
- A speech-to-text service for Pipecat and LiveKit Agents
- Or 16 kHz PCM over a WebSocket, from any language
- Let the model end turns, or send
finalizefrom your own VAD - Partial text as people speak, a final transcript at every turn
- LiveKit Agents
- Pipecat
- WebSocket API
import json, os, requests, websockets
H = {"Authorization": f"Bearer {os.environ['AIRCAPS_API_KEY']}"}
url = requests.get("https://api.aircaps.com/v1/realtime/endpoint",
params={"model": "aircaps-live-1"}, headers=H).json()["websocket_url"]
async with websockets.connect(url, additional_headers=H) as ws:
await ws.send(json.dumps({"type": "start", "model": "aircaps-live-1",
"encoding": "pcm_s16le", "sample_rate": 16000, "channels": 1,
"end_of_turn": True}))
# after "ready": send 20 ms PCM frames, read "transcript" messages;
# a final with "end_of_turn": true means the caller has finished
Where it fits.
Support, scheduling and sales calls
Callers are heard right the first time, so the agent books the right slot and reads back the right number.
Assistants that answer sooner
Turns end when people finish their thought, so replies come fast without cutting anyone off.
One model your customers switch on
Plug it into the LiveKit or Pipecat stack you already run and give every agent on your platform faster turns.
Specifications.
More in the AirCaps Live 1 guide.
- Model ID
aircaps-live-1- Interface
- WebSocket; look up the closest endpoint before each session
- Audio
- 16 kHz, mono, 16-bit PCM (
pcm_s16le), at real-time pace - Turn endings
- Detected by the model (three modes, tunable), or sent by you with
finalize - Output
- Live text as people speak, the part that will not change marked, and a final transcript at every turn
- Latency
- 38 ms median time to final transcript; 52 ms at p95, 65 ms at p99
- Frameworks
- Pipecat, LiveKit Agents, or the WebSocket API directly
- Language
- English; multilingual support coming soon
- Price
- $0.20 per audio hour, less than every leading speech-to-text API; on every account
Questions.
What is TTFS?
Time to final segment: the time from the moment a speaker stops talking to the moment the final transcript of what they said arrives. For a voice agent it is the speech-to-text share of the silence a caller hears before the agent answers. We start the clock when the last of the caller's audio has been sent and stop it when the final transcript arrives, the same way for every model.
How does turn detection work?
With end_of_turn on, the model decides when the caller has finished, from what they said as well as how long they paused, and sends the final transcript with an end-of-turn flag and a confidence. Choose min_latency, balanced or max_accuracy, or set the threshold and pause limits yourself, and change them mid-session.
Can I use my own VAD or turn detector?
Yes. Send finalize when your detector says the turn is over, and the final transcript comes back right away. Both ways work in the same session.
Does it work with LiveKit and Pipecat?
Yes. Use it as the speech-to-text service in a Pipecat pipeline or a LiveKit agent, or connect to the WebSocket directly.
How fast is it?
38 ms median from the end of speech to the final transcript, 52 ms at p95 and 65 ms at p99: faster than the leading models from Deepgram, AssemblyAI, ElevenLabs and Microsoft, on 9,205 clips from public voice-agent benchmarks.
How accurate is it?
3.65% word error rate across 9,205 clips from the Pipecat, Coval, LiveKit EOT Bench, WildASR and PriMock57 benchmarks, lower than the leading models from Deepgram, AssemblyAI, ElevenLabs and Microsoft. On Pipecat's STT benchmark it is 1.52%.
Which languages does it support?
English, with multilingual support coming soon.
Can I send phone audio?
Yes. Send 16 kHz, mono, 16-bit PCM: resample 8 kHz call audio first.
How much does it cost?
$0.20 per hour of audio you stream, less than any of the leading speech-to-text APIs, which list at $0.22 to $0.54 an hour. Every account includes 5 free hours. See pricing.
How do I get access?
Every account has AirCaps Live 1. Create an API key in the console and start streaming; every account includes 5 free hours.
Do you train on my audio?
By default, customer data may be used to train and benchmark our models. Paid accounts can opt out in the console (Settings, Data controls) and choose how long data is kept; Enterprise customers can get zero data retention. See Data and privacy in the docs.
More models.
Answer every caller sooner.
AirCaps Live 1 is on every account. Plug it into your agent with 5 free hours.