In a café
Without Clarity
Chatter keeps turn detection busy, so your agent waits or interrupts.
With Clarity
Your agent answers when the caller stops talking.
ClarityYour voice agent hears the caller. Not the room.
Clarity removes background noise in real time and, with a short voice sample, every voice but the caller’s.
Press play, then switch Clarity on.
Each sample: a caller, a quieter second talker and the noise of the place.
Your agent talks to one person.
In a station hall
Without Clarity
Voices from the crowd end up in the transcript.
With Clarity
The transcript holds what the caller said.
In an open-plan office
Without Clarity
A colleague’s “yes” counts as the caller’s answer.
With Clarity
With a 2 to 8 second voice sample, your agent hears only the caller.
Ahead of every streaming model we compared.
On both public benchmarks, Clarity 1 gives your agent the cleanest audio of any streaming model in the test.
Put to the test.
Against the best.
Target-speaker quality
Libri2Mix · DNSMOS, higher is better
| Metric | Clarity 1240 ms | StarTSE[1]560 ms |
|---|---|---|
| Speech quality SIG | 3.56 (best) | 3.54 |
| Background quality BAK | 4.03 (best) | 3.75 |
| Overall quality OVRL | 3.27 (best) | 3.12 |
Noise removal
DNS 2020 Challenge · DNSMOS, higher is better
| Metric | Clarity 1240 ms | UnprocessedOriginal audio | AI Acoustics15 ms | DeepFilterNet310 ms | GTCRN16 ms |
|---|---|---|---|---|---|
| Speech quality SIG | 3.62 (best) | 3.39 | 3.46 | 3.51 | 3.34 |
| Background quality BAK | 4.11 (best) | 2.62 | 3.60 | 4.11 (best) | 3.98 |
| Overall quality OVRL | 3.37 (best) | 2.48 | 2.98 | 3.25 | 3.05 |
DeepFilterNet3 ties on background quality without reverb; offline models, which need the whole recording first, score up to 0.08 higher on speaker extraction.
Works on both ends of the line.
Voice agents
Your agent hears only the caller, even in a station hall.
Call centers
Customers hear your agent, without the noise, echo and colleagues around her.
One line in LiveKit or Pipecat.
Clarity sits on your agent’s audio input and streams, so everything after it works from the cleaned caller.
- Caller
- ClarityNoise and other voices out
- Turn detection
- Speech-to-text
- Language model
- Text-to-speech
from livekit.agents import room_io
from kugelaudio import load_audio
from kugelaudio.livekit import enhancement
await session.start(
room=ctx.room,
agent=agent,
room_options=room_io.RoomOptions(
audio_input=room_io.AudioInputOptions(
noise_cancellation=enhancement(), # reads KUGELAUDIO_API_KEY
# keep one voice: enhancement(speaker=load_audio("caller.wav"))
),
),
)Every spec of Clarity 1.
- Model ID
clarity-1- Tasks
- Noise removal. With a speaker sample: only that voice (the Python and JavaScript SDKs pick the task)
- Position
- Before voice activity detection and speech-to-text
- Added latency
- 50 ms at 240 ms chunks
- Input
- WAV, 16, 24 or 32 bit integer or 32 bit float, mono or stereo, 8 to 48 kHz
- Output
- Mono 16 bit PCM, 24 kHz, same length as the input
- Speaker sample
- 2 to 8 seconds
- Recordings
- Up to 300 seconds per request
POST /v1/audio/enhance - Live audio
- WebSocket stream, enhanced as it arrives
WS /v1/audio/enhance/stream
Pay per minute, billed by the second.
Pay as you go
€0.0015per audio minute, excl. VAT
Billed from your credits per second of input audio, rounded up, one second minimum. Failed requests are not billed.
Get startedEnterprise
Self-hosting and on-premise deployment, on Enterprise plans.
The short answers.
What is Clarity?
KugelAudio’s line of speech enhancement models. Clarity 1 (clarity-1) removes background noise and, with a speaker sample, keeps only that voice, on recordings and live audio.
How much latency does Clarity add?
About 50 ms if your pipeline already reads audio in 240 ms chunks. With smaller frames, a sound can come out up to 290 ms after it went in.
How does it know which voice to keep?
From a 2 to 8 second sample of that person, ideally alone and with little noise. Without a sample, Clarity removes noise and keeps the speech.
Which audio can I send?
WAV at 8 to 48 kHz, up to 300 seconds per recording, or live 16 bit PCM over a WebSocket. The output is mono 16 bit PCM at 24 kHz.
Can I run Clarity on my own infrastructure?
Yes. Self-hosting and on-premise deployment are available on Enterprise plans.