Ultra had the highest perceived sound quality of the speech models we tested, with noise and with echo. Plus took noisy recordings from 1.72 to 3.04 on a 5-point scale, while keeping the voice as recorded. Those are the results of 480 noisy test recordings and a small echo trial we ran in November 2026. This post shows the method, the numbers and the limits, so you can judge them yourself.

What we tested

We took 120 ten-second clips of nine people reading in English, German and Dutch, plus two bilingual readers (Arabic and English, Korean and English). We mixed each clip with one noise recording at four levels, from -5 dB, where the noise is louder than the voice, to +10 dB, where it's a quiet background. That made 480 recordings per system.

Every system cleaned the same files. We compared two modes of clir Voice with four other speech models, which we've anonymized as Model A to D, and with the unprocessed noisy recording.

We measured perceived quality with DNSMOS, a neural network trained to predict how people would rate speech, from 1 to 5. It needs no clean reference. It's a prediction, not a listening test.

For echo, we took two clips and added reverb from two rooms, one of them large, without noise. That made four recordings per system.

The results

System Perceived quality with noise (1-5) Perceived quality with echo (1-5)
Clean reference 3.29 3.31
Unprocessed recording 1.72 2.29
clir Voice Ultra 3.36 3.39
clir Voice Plus 3.04 2.66
Model A 3.30 3.30
Model B 2.64 2.81
Model C 3.16 3.17
Model D 1.72 2.62

What the numbers mean

Plus lifts noisy speech and keeps the voice. It took noisy recordings from 1.72 to 3.04 and echoey ones from 2.29 to 2.66. It's meant to remove noise and leave the voice as recorded, so it does less for echo. Models A and C scored higher here.

Ultra sounds the cleanest. Its 3.36 with noise is well above the unprocessed 1.72 and above the clean recordings' 3.29, because it also cleans up flaws in the source. Model A came close at 3.30, so this lead is small. In the hardest condition, with noise louder than speech, Ultra and Model A tied at about 3.27, while Plus scored 2.74.

Echo

Ultra had the highest average on echo, at 3.39, and the highest score in three of the four recordings. Model A beat it in the smaller room on one clip, and the gap between models widened in the large room. Four recordings can show a direction, but they don't prove one.

On a phone

Lab scores don't matter if a model can't run where you need it. On an iPhone 13 Pro on iOS 17, a 60-second file took about 3 seconds in Plus and about 8 seconds in Ultra, without any audio leaving the phone. Ultra's first load after installing took 41 to 50 seconds, because the phone prepares the model once. Ultra needs more memory, so it may be unavailable on some devices.

Our earlier research

We've worked on separating speech from noise since before these models existed. In a 2022 internal study, we built a two-microphone separation network and tested it on a phone on a table, with voices and noise played through measured room responses, and compared it with five commercial noise-cancellation tools as they were in June 2022.

One voice with noise Signal accuracy (dB) Speech quality (PESQ) Intelligibility (STOI) Transcript errors
Our research model 25.5 3.22 0.93 12.5%
Best of the five tools on each measure 9.72 1.66 0.81 35.5%

The study also tested something harder: two people talking at once, each to be recovered separately.

Two voices Signal accuracy (dB) Speech quality (PESQ) Intelligibility (STOI) Transcript errors
Without noise: our research model 24.09 2.53 0.93 4.0%
Without noise: best of four tools on each measure -18.97 1.10 0.16 81.6%
With noise: our research model 10.63 1.66 0.79 38.2%
With noise: best of five tools on each measure -19.25 1.12 0.13 89.9%

The commercial tools are built to keep voices and remove noise, not to pull two speakers apart, so this shows what the two-microphone approach made possible more than a failure on their part. One tool's results without noise weren't reliable in our records, so that row covers four, and the transcript figure in the last row covers four as well.

Three things limit this result. The study is internal and hasn't been peer-reviewed. Our network had two microphones and was built for this setup, while the commercial tools are general-purpose and those versions are years old. And it's a research model, not what clir Voice uses today.

It also taught us how fast this field moves: in our 2026 benchmark, the models clir Voice shipped until 2026 scored 2.57 to 2.65 on perceived quality, against 3.04 for Plus and 3.36 for Ultra. So we replaced them.

Limits of this test

  • One measure. The results table shows perceived quality only. A recording can sound clean and still change a word, because Ultra rebuilds the voice. Check names, numbers and quotes against the original.
  • Mixed noise isn't real life. Added noise and reverb are repeatable, but real rooms and microphones behave differently.
  • Scores don't replace ears. DNSMOS predicts opinion, and a person may hear things it misses. Listen for yourself.
  • We ran this ourselves. We have an interest in the result, so treat it accordingly.
  • Other products weren't included. This compares speech models, not other apps, so it doesn't say how clir Voice compares with any app you might be using.
  • Small differences mean little. Treat close scores as ties.

Try it on your own voice

You can record and preview Plus and Ultra for free in clir Voice. Your own room and microphone are a better test than ours. Features, prices and limits are on the clir Voice facts page.

Tested on 3 November 2026 on an RTX 4090 GPU. iPhone timings from 18 October 2026.