AI speech enhancer benchmark 2026

We ran 100 noisy recordings through eight speech enhancers and counted the word errors.

Model under test
Cleanvoice Studio Sound 4.1
Evaluation
Cleanvoice AI Research
Dataset
Dolby UNIVERSE, 100 clips
Other tools
Their own app or API, default settings
Snapshot
2 November 2026

Cleanvoice Studio Sound 4.1 made the fewest word errors: 15.5%, against an average of 21.8% for Adobe Podcast, Auphonic, Dolby UNIVERSE, ElevenLabs and Descript. It dropped 56% fewer words than that average, and got the top SCOREQ, DNSMOS and speaker similarity scores.

At a glance

  • Word error rate15.5%1st of 8next best: Adobe Podcast 17.1%
  • Dropped words3.0%1st of 8next best: Dolby UNIVERSE 5.1%
  • SCOREQ4.541st of 8next best: Dolby UNIVERSE 4.33
  • DNSMOS OVRL3.351st of 8next best: Descript 3.24
  • Speaker similarity0.9551st of 8next best: Dolby UNIVERSE 0.953

01Overview

Each cell is coloured by the tool's place in that column; solid blue is first. Lower is better for the word columns. Higher is better for the rest.

Bar chart of word error rate on the UNIVERSE benchmark: Cleanvoice 4.1 15.5%, Cleanvoice 4.0 16.7%, Adobe Podcast 17.1%, Auphonic Studio Voice 20.3%, Dolby UNIVERSE 22.3%, DeepFilterNet3 22.6%, ElevenLabs 23.8%, Descript Studio Sound 25.4%, against 17.0% for the unprocessed input. Side panels: SCOREQ, DNSMOS overall, speaker similarity and bandwidth, where Cleanvoice 4.1 leads on SCOREQ (4.54), DNSMOS (3.35) and speaker similarity (0.955) and ties Descript Studio Sound on bandwidth (17.3 kHz).
Word error rate per tool, with speech quality, overall quality, speaker similarity and bandwidth beside it.
Every metric for every tool on the UNIVERSE 100-clip benchmark, 1 November 2026. Each cell is coloured by the tool's place in its column.
ToolWord error rate (lower is better)Dropped words (lower is better)Changed words (lower is better)SCOREQ (higher is better)DNSMOS OVRL (higher is better)Speaker similarity (higher is better)ESTOI (higher is better)PESQ (higher is better)Bandwidth (higher is better)
Cleanvoice Studio Sound 4.115.5% (best)3.0% (best)11.2%4.54 (best)3.35 (best)0.955 (best)0.8401.8817.3 kHz (best)
Cleanvoice Studio Sound 4.016.7%3.2%12.4%4.513.320.9450.8321.9414.0 kHz
Adobe Podcast (Enhance Speech)17.1%6.8%9.8% (best)3.573.080.9450.8242.079.5 kHz
Auphonic Studio Voice20.3%6.2%13.8%4.163.210.9490.7931.6016.5 kHz
Dolby UNIVERSE (research outputs)22.3%5.1%15.7%4.333.220.9530.863 (best)2.64 (best)8.0 kHz
DeepFilterNet3 (open source)22.6%9.6%12.4%3.182.840.9340.7531.857.5 kHz
ElevenLabs Voice Isolator23.8%7.1%15.3%3.723.110.9280.7541.7811.8 kHz
Descript Studio Sound25.4%8.1%16.1%3.993.240.9330.8031.8517.3 kHz (best)
Unprocessed input17.0%5.9%10.8%2.642.270.9160.7141.558.0 kHz

The raw input is there for reference and isn't ranked. DeepFilterNet3 is a free open-source model. Dropped and changed words are per 100 words. Word error rate also counts words the transcriber added, so those two don't add up to it. Cleanvoice 4.1, 4.0 and DeepFilterNet3 processed each clip on its own; the other tools processed all 100 clips joined into one file.

02Word errors

A speech enhancer should clean the sound and leave the words alone. Here are the two main kinds of error: words dropped and words changed.

Stacked bars of word errors by type for each tool. Cleanvoice 4.1 drops the fewest words (3.0 per 100), then Cleanvoice 4.0 (3.2); DeepFilterNet3 drops the most (9.6). Adobe Podcast changes the fewest words (9.8 per 100); Descript Studio Sound changes the most (16.1).
Dropped and changed words per 100 reference words, with each tool's word error rate.
Word error rate, with dropped and changed words per 100 reference words, for each tool.
ToolWord error rate (lower is better)Dropped words (lower is better)Changed words (lower is better)
Cleanvoice Studio Sound 4.115.5% (best)3.0% (best)11.2%
Cleanvoice Studio Sound 4.016.7%3.2%12.4%
Adobe Podcast (Enhance Speech)17.1%6.8%9.8% (best)
Auphonic Studio Voice20.3%6.2%13.8%
Dolby UNIVERSE (research outputs)22.3%5.1%15.7%
DeepFilterNet3 (open source)22.6%9.6%12.4%
ElevenLabs Voice Isolator23.8%7.1%15.3%
Descript Studio Sound25.4%8.1%16.1%
Unprocessed input17.0%5.9%10.8%

03Clip by clip

On each of the 100 clips, who made fewer word errors?

Clip-by-clip word errors, Cleanvoice 4.1 against each tool. Cleanvoice 4.0: fewer on 21 clips, same on 65, more on 14. Adobe Podcast: 26, 52, 22. Auphonic: 41, 45, 14. Dolby UNIVERSE: 39, 42, 19. DeepFilterNet3: 49, 39, 12. ElevenLabs: 52, 39, 9. Descript Studio Sound: 50, 40, 10.
Clips where Cleanvoice made fewer, the same or more word errors than each tool, and the gap in word error rate.
Cleanvoice vs.Cleanvoice fewer errorsSameCleanvoice more errorsWER gap (points)
Studio Sound 4.021 clips6514+1.2
Adobe Podcast26 clips5222+1.6
Auphonic41 clips4514+4.8
Dolby UNIVERSE39 clips4219+6.8
DeepFilterNet349 clips3912+7.1
ElevenLabs52 clips399+8.3
Descript50 clips4010+9.9

WER gap: the other tool's word error rate minus Cleanvoice's.

04Words and sound

Word errors on one axis, overall sound quality on the other.

Scatter plot of word error rate against DNSMOS overall quality. Cleanvoice 4.1 sits at the top right (15.5% word errors, DNSMOS 3.35), with Cleanvoice 4.0 next to it (16.7%, 3.32). Adobe Podcast is at 17.1% and 3.08, Auphonic 20.3% and 3.21, Dolby UNIVERSE 22.3% and 3.22, DeepFilterNet3 22.6% and 2.84, ElevenLabs 23.8% and 3.11, Descript Studio Sound 25.4% and 3.24. The unprocessed input is at 17.0% and 2.27.

Every tool scored well above the raw input on overall quality (DNSMOS 2.27). Cleanvoice was the only one that also clearly cut word errors: 15.5%, against 17.0% for the raw input. Adobe Podcast stayed level at 17.1%, and the rest made more errors than the raw input.

05Studio Sound 4 vs. 4.1

Five recordings, each with a different problem, through the old model and the new one. These are for listening and aren't part of the score.

8%

fewer word errors than Studio Sound 4

On the same 100 clips, word errors went from 16.7% with Studio Sound 4 to 15.5% with 4.1. Both models processed each clip on its own.

  • Restoration

    Old, damaged audio from a 1960s film.

    Selected: Original

    0:00

  • Reverb

    A voice with room echo.

    Selected: Original

    0:00

  • Distortion

    Distorted speech.

    Selected: Original

    0:00

  • Thin

    A thin-sounding voice.

    Selected: Original

    0:00

  • Muffled

    A muffled voice.

    Selected: Original

    0:00

06How we tested

Dataset
The 100-clip validation set from Dolby's UNIVERSE paper (Serrà et al., CC BY 4.0): noisy, echoey 16 kHz speech with clean originals to compare against. Some of the noise comes from DEMAND (CC BY-SA 3.0).
Settings
Each tool's main speech enhancement mode at default strength, with anything that removes silence or trims the file turned off. For Auphonic that's Studio Voice (Beta), its speech enhancement model, and not Speech Isolation, which is a noise remover. Dolby's row uses the outputs published with its paper.
Word errors
We transcribed every output and every clean original with Omnilingual ASR CTC 7B v2 from Meta, a multilingual model. CTC models generally don't hallucinate the way models with a language model do, which makes the comparison fairer and more predictable. Errors are per 100 words of the clean transcript (1,423 words in total).
Sound quality
SCOREQ and DNSMOS P.835 rate the sound without a reference. ESTOI and PESQ compare it with the clean original. Speaker similarity compares speaker embeddings of the output and the original. All at 16 kHz.
Bandwidth
The median highest frequency each tool's output reaches. The input is 16 kHz audio, so it stops at 8 kHz.
Reruns
Want the files and scripts to rerun this, or want a tool added? Email support@cleanvoice.ai.

07Limitations

  • One dataset, 100 clips.

    UNIVERSE is built mostly from VCTK speech, all in English. Podcasts, phone calls, music beds and other languages could give different results.

  • 16 kHz input only.

    We didn't test full-band recordings here, and the ranking could change on them.

  • Default settings, one day.

    Tools ran on 1 November 2026. Products change, and other settings could give other results.

  • Test it on your own recordings.

    Run your audio through Cleanvoice Studio Sound 4.1 and compare. New accounts get 30 free minutes, no credit card.

    Try Cleanvoice

Questions about the benchmark

Still have a question?

Send it over and a real person will answer.

support@cleanvoice.ai
Ask Now

In this test, Cleanvoice Studio Sound 4.1 made the fewest word errors: 15.5%, with 3.0 dropped words per 100. Next came Adobe Podcast (17.1%), Auphonic Studio Voice (20.3%), Dolby UNIVERSE (22.3%), ElevenLabs Voice Isolator (23.8%) and Descript Studio Sound (25.4%). Cleanvoice AI Research ran the test on 100 noisy clips from Dolby's UNIVERSE set.

On these 100 clips it made fewer word errors (15.5% vs. 17.1%), dropped fewer words (3.0 vs. 6.8 per 100), and scored higher on SCOREQ (4.54 vs. 3.57) and DNSMOS (3.35 vs. 3.08). Adobe did better on PESQ (2.07 vs. 1.88) and changed fewer words (9.8 vs. 11.2 per 100). Clip by clip it was close: Cleanvoice had fewer errors on 26 clips, the same on 52 and more on 22.

If you care about word accuracy, Cleanvoice Studio Sound 4.1 had the lowest word error rate in this test (15.5%, vs. 17.1% for Adobe). It dropped less than half as many words and returned 17.3 kHz audio, where Adobe returned 9.5 kHz. Cleanvoice also has an API, which Adobe Podcast doesn't.

ElevenLabs Voice Isolator had a 23.8% word error rate and dropped 7.1 words per 100. Cleanvoice Studio Sound 4.1 had 15.5% and dropped 3.0. Cleanvoice also scored higher on SCOREQ (4.54 vs. 3.72), DNSMOS (3.35 vs. 3.11) and bandwidth (17.3 vs. 11.8 kHz). Clip by clip, Cleanvoice had fewer errors on 52 clips, the same on 39 and more on 9.

It made 8% fewer word errors on the same 100 clips: 15.5% against 16.7%. You can hear both models on five example recordings on this page.

No. They're different products from different companies that happen to share a name. In this test, Descript Studio Sound had a 25.4% word error rate and Cleanvoice Studio Sound 4.1 had 15.5%. Both reached 17.3 kHz of bandwidth.

Cleanvoice Studio Sound 4.1 made fewer word errors (15.5% vs. 20.3%), dropped fewer words (3.0 vs. 6.2 per 100), and scored higher on SCOREQ (4.54 vs. 4.16), DNSMOS (3.35 vs. 3.21) and bandwidth (17.3 vs. 16.5 kHz). Clip by clip, Cleanvoice had fewer errors on 41 clips, the same on 45 and more on 14.

We used the 100 noisy clips from Dolby's public UNIVERSE validation set. Adobe Podcast, Auphonic Studio Voice, Descript and ElevenLabs processed them joined into one 8.4-minute file, at default settings, on 1 November 2026. Cleanvoice processed each clip on its own. Omnilingual ASR, a CTC model, transcribed every output and the clean originals to count word errors. SCOREQ, DNSMOS, PESQ, ESTOI and speaker similarity scored the sound.

No. Cleanvoice AI Research ran it, and Cleanvoice makes one of the tools. The dataset is public, and you can email us for the files and scripts to rerun it.

No, every score is automatic. You can listen to five before-and-after examples on this page.

No. Studio Sound only works on the sound. Filler-word, stutter and silence removal are a separate Cleanvoice feature you turn on yourself, and they were off for this test.

Hear it on your own audio.

Run your recordings through Cleanvoice Studio Sound 4.1 in the app, or from your code with the API.

Try CleanvoiceSpeech enhancement API