Test results · October 22, 2026
AI Speech Enhancer Test (2026): Diffio vs Adobe, Descript, Auphonic, ElevenLabs & Cleanvoice
The largest quality gain and the lowest word error rate on 59 real recordings.
Our verdict
Diffio 4.6 Pro had the largest average quality gain of the seven runs on 59 real recordings (SCOREQ +2.31, 95% interval 2.15 to 2.47) and the lowest word error rate (10.5%). Of the other tools, Cleanvoice (+1.89) and Adobe Podcast (+1.87) came closest on quality, and Descript Studio Sound improved it least (+0.46). Diffio did not win everywhere: other tools beat Diffio 4.6 Pro on word error rate for phone and remote calls, and on quality gain and word error rate for narration over music.
Highest quality
Diffio 4.6 Pro
+2.31 quality gain
10.5% word error rate
Paid plans · API $10/h
Every plan
Diffio 4.6 Flash
+2.26 quality gain
10.9% word error rate
Every plan, including free · API $5/h
Closest other tool
Cleanvoice
+1.89 quality gain
16.4% word error rate
$2.20/h at 5 hours a month (Subscription 10 hours plan)
Diffio is not affiliated with or endorsed by Adobe, Descript, Auphonic, ElevenLabs or Cleanvoice. Product names are trademarks of their owners.
Results across all 59 clips
Each tool processed the same 59 recordings once, in its strongest speech cleanup mode at default amounts. Diffio ran both of its models, so there are 7 runs of 6 products. Only Diffio 4.6 Pro and Diffio 4.6 Flash lowered the word error rate below the untouched recordings' 13.2%; every other run made the speech harder for a recognizer to transcribe.
All recordings
Bars are averages over 59 clips; the thin lines show 95% intervals. SCOREQ (no-reference speech quality model; higher is better). Gain = output minus the original clip. Word error rate of a small speech recognizer (faster-whisper small.en) on each output, against MAI-Transcribe-2 on the original; lower is better. On the word error rate view, the dashed line is the untouched original recording.
Show the numbers, including similarity to the original recording
| Tool | Quality gain (SCOREQ) | Word error rate | Similarity to the original recording |
|---|---|---|---|
| Diffio 4.6 Pro | +2.3195% interval 2.15 to 2.47 | 10.5%95% interval 7.8% to 13.7% | 0.58 |
| Diffio 4.6 Flash | +2.2695% interval 2.09 to 2.44 | 10.9%95% interval 7.6% to 15.1% | 0.56 |
| Adobe Podcast | +1.8795% interval 1.74 to 2.00 | 18.2%95% interval 13.6% to 23.3% | 0.60 |
| Cleanvoice | +1.8995% interval 1.73 to 2.05 | 16.4%95% interval 11.8% to 21.3% | 0.58 |
| Auphonic | +0.8295% interval 0.65 to 1.00 | 13.3%95% interval 10.0% to 16.8% | 0.84 |
| ElevenLabs | +0.7995% interval 0.60 to 0.98 | 18.1%95% interval 13.0% to 23.6% | 0.83 |
| Descript Studio Sound | +0.4695% interval 0.39 to 0.53 | 22.6%95% interval 16.6% to 28.9% | 0.72 |
| Original recording | No processing (average SCOREQ 2.31) | 13.2%95% interval 9.6% to 17.3% | 1.00 |
Averages over 59 clips of real recordings, run October 22, 2026. A higher SCOREQ gain and a lower word error rate are better. Similarity to the original recording is measured against the noisy original, so a tool that changes less scores higher; it is not a measure of voice quality.
Similarity to the original recording. ECAPA speaker-embedding similarity to the original recording, noise included; tools that change less score higher. A tool that removes more noise and echo moves further from the original, so this is not a measure of how natural the voice sounds.
Quality and cost
At 5 hours of audio a month, Diffio's Basic plan is the cheapest of the tested tools ($1.60/h), and Diffio 4.6 Pro has the largest quality gain (+2.31). Tools nearer the top left give more quality per dollar.
Quality gain against cost per hour
| Tool | Cost per hour at 5 hours a month | Quality gain (SCOREQ) |
|---|---|---|
| Diffio 4.6 Pro | $1.60/h on Basic | +2.31 |
| Diffio 4.6 Flash | $1.60/h on Basic | +2.26 |
| Adobe Podcast | $4/h on Firefly Pro (includes Podcast Premium) | +1.87 |
| Cleanvoice | $2.20/h on Subscription 10 hours | +1.89 |
| Auphonic | $2.40/h on One-time 5 hours | +0.82 |
| ElevenLabs | $19.80/h on Pro | +0.79 |
| Descript Studio Sound | $4.80/h on Hobbyist | +0.46 |
Cost is each tool's cheapest published plan for 5 hours of audio a month, from the vendors' own price pages; both Diffio models use the same web plan. Free tiers are not counted: Adobe Podcast's free tier also covers 5 hours a month, so its point shows the paid plan. Full prices, free tiers and other usage levels: AI audio enhancer pricing.
Results by recording type
Each group below charts the same two measures for its clips, then plays them: pick a clip to hear the same recording through each tool. Players start on the clip with Diffio 4.6 Pro's median result, not its best one; archived sermons are numbered from the noisiest original to the cleanest. The credit and license for every recording are under its player.
Archived sermons
Decades-old church recordings with tape hiss, hum and room sound. In our test's archived sermons group (29 clips), Diffio 4.6 Pro raised the SCOREQ speech quality score by 2.37 on average, against 1.74 for Cleanvoice, the highest of the other tools. Its word error rate was 12.1%, against 16.4% for the original recordings.
Archived sermons: 29 clips
Starts on the clip with Diffio 4.6 Pro's median result. Pick any of the 29. Sermons are numbered from the noisiest original to the cleanest.
Sermon 17. Decades-old church recordings with tape hiss, hum and room sound.
Archived sermon recording, used with permission. Excerpts are loudness-matched for fair listening.
| Version | SCOREQ | Word error rate | Similarity to the original |
|---|---|---|---|
| Diffio 4.6 Pro | 4.86 | 6.7% | 0.55 |
| Diffio 4.6 Flash | 4.85 | 2.7% | 0.55 |
| Adobe Podcast | 4.44 | 8.0% | 0.52 |
| Cleanvoice | 4.62 | 12.0% | 0.58 |
| Auphonic | 2.87 | 6.7% | 0.86 |
| ElevenLabs | 2.95 | 6.7% | 0.92 |
| Descript Studio Sound | 3.36 | 12.0% | 0.69 |
| Original recording | 2.40 | 5.3% | 1.00 |
Show the numbers for archived sermons
| Tool | Quality gain (SCOREQ) | Word error rate | Similarity to the original recording |
|---|---|---|---|
| Diffio 4.6 Pro | +2.3795% interval 2.18 to 2.57 | 12.1%95% interval 8.2% to 16.5% | 0.45 |
| Diffio 4.6 Flash | +2.2795% interval 2.05 to 2.49 | 13.9%95% interval 8.2% to 21.3% | 0.43 |
| Adobe Podcast | +1.6895% interval 1.53 to 1.82 | 25.0%95% interval 17.3% to 33.4% | 0.55 |
| Cleanvoice | +1.7495% interval 1.53 to 1.94 | 18.5%95% interval 12.8% to 25.0% | 0.56 |
| Auphonic | +0.3695% interval 0.25 to 0.48 | 16.2%95% interval 11.0% to 21.9% | 0.88 |
| ElevenLabs | +0.3295% interval 0.18 to 0.49 | 20.9%95% interval 14.5% to 28.1% | 0.83 |
| Descript Studio Sound | +0.5395% interval 0.43 to 0.63 | 32.0%95% interval 21.6% to 42.8% | 0.60 |
| Original recording | No processing (average SCOREQ 2.29) | 16.4%95% interval 10.6% to 23.3% | 1.00 |
Averages over 29 clips of real recordings, run October 22, 2026. A higher SCOREQ gain and a lower word error rate are better. Similarity to the original recording is measured against the noisy original, so a tool that changes less scores higher; it is not a measure of voice quality.
Echoey rooms
Lectures, sermons and talks recorded in halls, chapels and galleries. In our test's echoey rooms group (5 clips), Diffio 4.6 Pro raised the SCOREQ speech quality score by 2.25 on average, against 2.16 for Adobe Podcast, the highest of the other tools. Its word error rate was 3.4%, against 4.8% for the original recordings.
Echoey rooms: 5 clips
Starts on the clip with Diffio 4.6 Pro's median result. Pick any of the 5.
Sermon at Calvary, Jerusalem. Lectures, sermons and talks recorded in halls, chapels and galleries.
"Sermon at Calvary, Church of the Holy Sepulchre" by Fabricio E. Lomanto, CC BY 3.0, via Wikimedia Commons. License: CC BY 3.0. Source recording. Excerpts are loudness-matched for fair listening.
| Version | SCOREQ | Word error rate | Similarity to the original |
|---|---|---|---|
| Diffio 4.6 Pro | 4.65 | 0.0% | 0.74 |
| Diffio 4.6 Flash | 4.62 | 0.0% | 0.73 |
| Adobe Podcast | 4.64 | 11.9% | 0.58 |
| Cleanvoice | 4.42 | 1.7% | 0.56 |
| Auphonic | 4.06 | 3.4% | 0.71 |
| ElevenLabs | 4.26 | 3.4% | 0.80 |
| Descript Studio Sound | 2.93 | 1.7% | 0.81 |
| Original recording | 2.37 | 1.7% | 1.00 |
Show the numbers for echoey rooms
| Tool | Quality gain (SCOREQ) | Word error rate | Similarity to the original recording |
|---|---|---|---|
| Diffio 4.6 Pro | +2.2595% interval 1.94 to 2.60 | 3.4%95% interval 1.3% to 5.5% | 0.75 |
| Diffio 4.6 Flash | +2.2495% interval 1.88 to 2.61 | 2.9%95% interval 1.0% to 4.7% | 0.74 |
| Adobe Podcast | +2.1695% interval 1.96 to 2.36 | 8.2%95% interval 4.3% to 12.1% | 0.67 |
| Cleanvoice | +2.0195% interval 1.83 to 2.20 | 3.9%95% interval 2.1% to 5.6% | 0.57 |
| Auphonic | +1.0395% interval 0.53 to 1.55 | 4.2%95% interval 2.2% to 6.8% | 0.77 |
| ElevenLabs | +1.6695% interval 1.27 to 1.95 | 5.9%95% interval 3.1% to 9.4% | 0.81 |
| Descript Studio Sound | +0.2695% interval 0.05 to 0.47 | 7.2%95% interval 2.7% to 12.2% | 0.83 |
| Original recording | No processing (average SCOREQ 2.41) | 4.8%95% interval 1.8% to 9.7% | 1.00 |
Averages over 5 clips of real recordings, run October 22, 2026. A higher SCOREQ gain and a lower word error rate are better. Similarity to the original recording is measured against the noisy original, so a tool that changes less scores higher; it is not a measure of voice quality.
Phone and remote calls
A phone-in podcast guest, a Zoom lecture and a Skype interview. In our test's phone and remote calls group (3 clips), Diffio 4.6 Pro raised the SCOREQ speech quality score by 1.69 on average, against 1.52 for Cleanvoice, the highest of the other tools. On phone and remote calls (3 clips), Cleanvoice (3.5%), Auphonic (4.2%), ElevenLabs (4.2%), Adobe Podcast (5.3%) and Descript Studio Sound (5.3%) had a lower word error rate than Diffio 4.6 Pro (6.5%). Diffio 4.6 Pro's word error rate (6.5%) was higher than the original recordings' (3.8%), so check that every word survives.
Phone and remote calls: 3 clips
Starts on the clip with Diffio 4.6 Pro's median result. Pick any of the 3.
NOAA podcast guest on the phone. A phone-in podcast guest, a Zoom lecture and a Skype interview.
"Diving Deeper: Exploring Hawaii's Coral Reefs" (December 2015), NOAA Ocean Service podcast, public domain. No endorsement implied.. Source recording. Excerpts are loudness-matched for fair listening.
| Version | SCOREQ | Word error rate | Similarity to the original |
|---|---|---|---|
| Diffio 4.6 Pro | 4.58 | 5.1% | 0.47 |
| Diffio 4.6 Flash | 4.54 | 5.1% | 0.45 |
| Adobe Podcast | 4.47 | 4.0% | 0.51 |
| Cleanvoice | 4.53 | 2.0% | 0.46 |
| Auphonic | 3.34 | 3.0% | 0.87 |
| ElevenLabs | 3.00 | 3.0% | 0.85 |
| Descript Studio Sound | 3.30 | 5.1% | 0.81 |
| Original recording | 2.96 | 3.0% | 1.00 |
Show the numbers for phone and remote calls
| Tool | Quality gain (SCOREQ) | Word error rate | Similarity to the original recording |
|---|---|---|---|
| Diffio 4.6 Pro | +1.6995% interval 1.47 to 1.98 | 6.5%95% interval 5.1% to 7.2% | 0.66 |
| Diffio 4.6 Flash | +1.6295% interval 1.39 to 1.89 | 5.7%95% interval 4.8% to 7.2% | 0.65 |
| Adobe Podcast | +1.4695% interval 1.27 to 1.61 | 5.3%95% interval 4.0% to 7.2% | 0.67 |
| Cleanvoice | +1.5295% interval 1.35 to 1.63 | 3.5%95% interval 2.0% to 6.0% | 0.62 |
| Auphonic | +0.5595% interval 0.38 to 0.76 | 4.2%95% interval 2.4% to 7.2% | 0.87 |
| ElevenLabs | +0.1695% interval 0.03 to 0.42 | 4.2%95% interval 2.4% to 7.2% | 0.91 |
| Descript Studio Sound | +0.2495% interval 0.13 to 0.34 | 5.3%95% interval 4.8% to 6.0% | 0.78 |
| Original recording | No processing (average SCOREQ 2.97) | 3.8%95% interval 2.4% to 6.0% | 1.00 |
Averages over 3 clips of real recordings, run October 22, 2026. A higher SCOREQ gain and a lower word error rate are better. Similarity to the original recording is measured against the noisy original, so a tool that changes less scores higher; it is not a measure of voice quality.
Wind and traffic
Interviews and narration outdoors: wind, surf, streets and aircraft. In our test's wind and traffic group (7 clips), Diffio 4.6 Pro raised the SCOREQ speech quality score by 2.05 on average, against 1.96 for Adobe Podcast, the highest of the other tools. Its word error rate was 6.9%, against 13.6% for the original recordings.
Wind and traffic: 7 clips
Starts on the clip with Diffio 4.6 Pro's median result. Pick any of the 7.
Beach cleanup interview in wind and surf. Interviews and narration outdoors: wind, surf, streets and aircraft.
"Ocean pollution in Ghana" by Esthee2010, CC BY-SA 4.0, via Wikimedia Commons; outputs shared under the same license. License: CC BY-SA 4.0. Source recording. Every processed version of this clip is shared under the same license. Excerpts are loudness-matched for fair listening.
| Version | SCOREQ | Word error rate | Similarity to the original |
|---|---|---|---|
| Diffio 4.6 Pro | 4.60 | 8.9% | 0.79 |
| Diffio 4.6 Flash | 4.64 | 10.0% | 0.77 |
| Adobe Podcast | 4.63 | 10.0% | 0.73 |
| Cleanvoice | 4.49 | 12.2% | 0.66 |
| Auphonic | 4.28 | 13.3% | 0.85 |
| ElevenLabs | 3.45 | 6.7% | 0.89 |
| Descript Studio Sound | 2.92 | 11.1% | 0.93 |
| Original recording | 2.46 | 12.2% | 1.00 |
Show the numbers for wind and traffic
| Tool | Quality gain (SCOREQ) | Word error rate | Similarity to the original recording |
|---|---|---|---|
| Diffio 4.6 Pro | +2.0595% interval 1.57 to 2.54 | 6.9%95% interval 3.7% to 10.4% | 0.72 |
| Diffio 4.6 Flash | +2.0495% interval 1.55 to 2.51 | 7.3%95% interval 3.8% to 10.7% | 0.71 |
| Adobe Podcast | +1.9695% interval 1.50 to 2.38 | 12.6%95% interval 5.5% to 23.0% | 0.65 |
| Cleanvoice | +1.8495% interval 1.27 to 2.37 | 17.4%95% interval 5.0% to 38.4% | 0.64 |
| Auphonic | +1.6095% interval 1.29 to 1.88 | 9.2%95% interval 4.8% to 13.9% | 0.78 |
| ElevenLabs | +1.5595% interval 1.05 to 2.07 | 22.7%95% interval 5.3% to 52.2% | 0.80 |
| Descript Studio Sound | +0.4695% interval 0.33 to 0.64 | 11.7%95% interval 5.4% to 20.7% | 0.93 |
| Original recording | No processing (average SCOREQ 2.43) | 13.6%95% interval 4.8% to 27.1% | 1.00 |
Averages over 7 clips of real recordings, run October 22, 2026. A higher SCOREQ gain and a lower word error rate are better. Similarity to the original recording is measured against the noisy original, so a tool that changes less scores higher; it is not a measure of voice quality.
Crowds and chatter
A diner, a media scrum, a stadium speech, a fair and a zoo tour. In our test's crowds and chatter group (5 clips), Diffio 4.6 Pro raised the SCOREQ speech quality score by 2.59 on average, against 2.35 for Adobe Podcast, the highest of the other tools. Diffio 4.6 Pro's word error rate (19.5%) was higher than the original recordings' (17.8%), so check that every word survives.
Crowds and chatter: 5 clips
Starts on the clip with Diffio 4.6 Pro's median result. Pick any of the 5.
John F. Kennedy at Rice Stadium, 1962. A diner, a media scrum, a stadium speech, a fair and a zoo tour.
President John F. Kennedy at Rice University, September 12, 1962, NASA film, public domain. No endorsement implied.. Source recording. Excerpts are loudness-matched for fair listening.
| Version | SCOREQ | Word error rate | Similarity to the original |
|---|---|---|---|
| Diffio 4.6 Pro | 4.61 | 6.6% | 0.74 |
| Diffio 4.6 Flash | 4.55 | 11.5% | 0.77 |
| Adobe Podcast | 4.35 | 3.3% | 0.72 |
| Cleanvoice | 4.40 | 9.8% | 0.77 |
| Auphonic | 3.86 | 11.5% | 0.88 |
| ElevenLabs | 3.05 | 1.6% | 0.88 |
| Descript Studio Sound | 2.53 | 6.6% | 0.94 |
| Original recording | 2.21 | 4.9% | 1.00 |
Show the numbers for crowds and chatter
| Tool | Quality gain (SCOREQ) | Word error rate | Similarity to the original recording |
|---|---|---|---|
| Diffio 4.6 Pro | +2.5995% interval 2.32 to 2.98 | 19.5%95% interval 8.6% to 33.7% | 0.72 |
| Diffio 4.6 Flash | +2.6695% interval 2.37 to 3.01 | 16.3%95% interval 10.9% to 22.3% | 0.73 |
| Adobe Podcast | +2.3595% interval 1.96 to 2.78 | 20.3%95% interval 10.2% to 29.9% | 0.65 |
| Cleanvoice | +2.3095% interval 1.72 to 2.79 | 34.2%95% interval 14.5% to 56.5% | 0.57 |
| Auphonic | +1.2495% interval 0.47 to 1.87 | 23.1%95% interval 13.9% to 32.4% | 0.80 |
| ElevenLabs | +1.4795% interval 0.89 to 2.06 | 33.2%95% interval 12.4% to 54.6% | 0.83 |
| Descript Studio Sound | +0.4595% interval 0.20 to 0.71 | 28.3%95% interval 14.8% to 40.2% | 0.87 |
| Original recording | No processing (average SCOREQ 1.93) | 17.8%95% interval 10.1% to 24.5% | 1.00 |
Averages over 5 clips of real recordings, run October 22, 2026. A higher SCOREQ gain and a lower word error rate are better. Similarity to the original recording is measured against the noisy original, so a tool that changes less scores higher; it is not a measure of voice quality.
Hum, machines and distortion
Mains hum, fans, machinery and an overloaded recorder. In our test's hum, machines and distortion group (4 clips), Diffio 4.6 Pro raised the SCOREQ speech quality score by 2.02 on average, against 1.96 for Adobe Podcast, the highest of the other tools. Its word error rate was 5.5%, against 5.6% for the original recordings.
Hum, machines and distortion: 4 clips
Starts on the clip with Diffio 4.6 Pro's median result. Pick any of the 4.
Space Station crew quarters with fan noise. Mains hum, fans, machinery and an overloaded recorder.
"Crew Quarters Tour Inside the Space Station", NASA Johnson Space Center, public domain. Source recording. Excerpts are loudness-matched for fair listening.
| Version | SCOREQ | Word error rate | Similarity to the original |
|---|---|---|---|
| Diffio 4.6 Pro | 4.60 | 9.9% | 0.84 |
| Diffio 4.6 Flash | 4.60 | 8.5% | 0.86 |
| Adobe Podcast | 4.55 | 8.5% | 0.76 |
| Cleanvoice | 4.29 | 9.9% | 0.84 |
| Auphonic | 4.12 | 8.5% | 0.92 |
| ElevenLabs | 3.58 | 9.9% | 0.91 |
| Descript Studio Sound | 3.23 | 8.5% | 0.93 |
| Original recording | 3.02 | 8.5% | 1.00 |
Show the numbers for hum, machines and distortion
| Tool | Quality gain (SCOREQ) | Word error rate | Similarity to the original recording |
|---|---|---|---|
| Diffio 4.6 Pro | +2.0295% interval 1.18 to 3.05 | 5.5%95% interval 3.4% to 8.4% | 0.81 |
| Diffio 4.6 Flash | +1.9295% interval 0.90 to 3.06 | 5.1%95% interval 3.4% to 7.4% | 0.82 |
| Adobe Podcast | +1.9695% interval 1.30 to 2.87 | 7.0%95% interval 5.2% to 8.3% | 0.72 |
| Cleanvoice | +1.8895% interval 1.13 to 2.78 | 7.8%95% interval 6.4% to 9.2% | 0.72 |
| Auphonic | +1.7195% interval 1.00 to 2.43 | 7.0%95% interval 5.6% to 8.5% | 0.84 |
| ElevenLabs | +1.3295% interval 0.54 to 2.10 | 7.9%95% interval 3.5% to 12.3% | 0.89 |
| Descript Studio Sound | +0.4895% interval 0.31 to 0.64 | 6.2%95% interval 4.8% to 7.8% | 0.92 |
| Original recording | No processing (average SCOREQ 2.67) | 5.6%95% interval 3.7% to 7.6% | 1.00 |
Averages over 4 clips of real recordings, run October 22, 2026. A higher SCOREQ gain and a lower word error rate are better. Similarity to the original recording is measured against the noisy original, so a tool that changes less scores higher; it is not a measure of voice quality.
Narration over music
An explainer voice-over with a music bed underneath. In our test's narration over music group (1 clip), Diffio 4.6 Pro raised the SCOREQ speech quality score by 1.73 on average, against 1.74 for Adobe Podcast, the highest of the other tools. On narration over music (1 clip), Adobe Podcast (+1.74) had a higher quality gain than Diffio 4.6 Pro (+1.73). On narration over music (1 clip), ElevenLabs (1.3%), Cleanvoice (3.7%), Auphonic (3.7%) and Descript Studio Sound (5.0%) had a lower word error rate than Diffio 4.6 Pro (6.3%). Diffio 4.6 Pro's word error rate (6.3%) was higher than the original recordings' (1.3%), so check that every word survives.
Narration over music: 1 clip
A single clip: treat these differences as indications, not rankings.
Weather explainer over music. An explainer voice-over with a music bed underneath.
"El Niño – WMO animated explainer video", World Meteorological Organization, CC BY 3.0, via Wikimedia Commons. License: CC BY 3.0. Source recording. Excerpts are loudness-matched for fair listening.
| Version | SCOREQ | Word error rate | Similarity to the original |
|---|---|---|---|
| Diffio 4.6 Pro | 4.73 | 6.3% | 0.87 |
| Diffio 4.6 Flash | 4.73 | 6.3% | 0.87 |
| Adobe Podcast | 4.73 | 7.5% | 0.80 |
| Cleanvoice | 4.47 | 3.7% | 0.83 |
| Auphonic | 4.59 | 3.7% | 0.91 |
| ElevenLabs | 4.44 | 1.3% | 0.90 |
| Descript Studio Sound | 3.03 | 5.0% | 0.85 |
| Original recording | 3.00 | 1.3% | 1.00 |
Show the numbers for narration over music
| Tool | Quality gain (SCOREQ) | Word error rate | Similarity to the original recording |
|---|---|---|---|
| Diffio 4.6 Pro | +1.73 | 6.3% | 0.87 |
| Diffio 4.6 Flash | +1.74 | 6.3% | 0.87 |
| Adobe Podcast | +1.74 | 7.5% | 0.80 |
| Cleanvoice | +1.47 | 3.7% | 0.83 |
| Auphonic | +1.60 | 3.7% | 0.91 |
| ElevenLabs | +1.45 | 1.3% | 0.90 |
| Descript Studio Sound | +0.04 | 5.0% | 0.85 |
| Original recording | No processing (average SCOREQ 3.00) | 1.3% | 1.00 |
Averages over 1 clip of real recordings, run October 22, 2026. A higher SCOREQ gain and a lower word error rate are better. Similarity to the original recording is measured against the noisy original, so a tool that changes less scores higher; it is not a measure of voice quality.
Historic recordings
Earhart, Apollo 9, 1890s–1910s cylinders and an oral history. In our test's historic recordings group (5 clips), Diffio 4.6 Pro raised the SCOREQ speech quality score by 2.86 on average, against 2.63 for Cleanvoice, the highest of the other tools. Its word error rate was 11.9%, against 12.3% for the original recordings.
Historic recordings: 5 clips
Starts on the clip with Diffio 4.6 Pro's median result. Pick any of the 5.
Apollo 9 air-to-ground. Earhart, Apollo 9, 1890s–1910s cylinders and an oral history.
Apollo 9 air-to-ground audio, NASA / Smithsonian National Air and Space Museum, public domain. Excerpts are loudness-matched for fair listening.
| Version | SCOREQ | Word error rate | Similarity to the original |
|---|---|---|---|
| Diffio 4.6 Pro | 4.39 | 46.0% | 0.42 |
| Diffio 4.6 Flash | 4.39 | 39.7% | 0.39 |
| Adobe Podcast | 4.33 | 39.7% | 0.37 |
| Cleanvoice | 4.28 | 41.3% | 0.52 |
| Auphonic | 2.26 | 36.5% | 0.77 |
| ElevenLabs | 1.71 | 22.2% | 0.85 |
| Descript Studio Sound | 2.22 | 36.5% | 0.39 |
| Original recording | 1.48 | 30.2% | 1.00 |
Show the numbers for historic recordings
| Tool | Quality gain (SCOREQ) | Word error rate | Similarity to the original recording |
|---|---|---|---|
| Diffio 4.6 Pro | +2.8695% interval 2.54 to 3.19 | 11.9%95% interval 0.7% to 29.6% | 0.50 |
| Diffio 4.6 Flash | +2.8695% interval 2.51 to 3.21 | 10.2%95% interval 1.1% to 25.0% | 0.49 |
| Adobe Podcast | +2.3695% interval 1.99 to 2.73 | 13.4%95% interval 3.7% to 27.2% | 0.51 |
| Cleanvoice | +2.6395% interval 2.37 to 2.88 | 14.5%95% interval 6.0% to 28.2% | 0.47 |
| Auphonic | +1.0695% interval 0.77 to 1.32 | 13.3%95% interval 5.4% to 25.5% | 0.74 |
| ElevenLabs | +0.6895% interval 0.41 to 0.95 | 12.7%95% interval 5.7% to 19.7% | 0.80 |
| Descript Studio Sound | +0.4195% interval 0.23 to 0.61 | 20.7%95% interval 14.1% to 29.7% | 0.68 |
| Original recording | No processing (average SCOREQ 1.67) | 12.3%95% interval 4.5% to 21.7% | 1.00 |
Averages over 5 clips of real recordings, run October 22, 2026. A higher SCOREQ gain and a lower word error rate are better. Similarity to the original recording is measured against the noisy original, so a tool that changes less scores higher; it is not a measure of voice quality.
Where Diffio did not win
We publish every result, including the ones that favor another tool. Single-clip groups are noted; one recording is not enough to rank tools.
- Phone and remote calls (3 clips): word error rate. Cleanvoice 3.5%, Auphonic 4.2%, ElevenLabs 4.2%, Adobe Podcast 5.3% and Descript Studio Sound 5.3%, against 6.5% for Diffio 4.6 Pro and 5.7% for Diffio 4.6 Flash.
- Narration over music (1 clip): quality gain. Adobe Podcast +1.74, against +1.73 for Diffio 4.6 Pro and +1.74 for Diffio 4.6 Flash.
- Narration over music (1 clip): word error rate. ElevenLabs 1.3%, Cleanvoice 3.7%, Auphonic 3.7% and Descript Studio Sound 5.0%, against 6.3% for Diffio 4.6 Pro and 6.3% for Diffio 4.6 Flash.
- Word error rate higher than the untouched original. Diffio 4.6 Pro's output had a higher word error rate than the untouched original on phone and remote calls (6.5% against 3.8%), crowds and chatter (19.5% against 17.8%) and narration over music (6.3% against 1.3%).
- Similarity to the original recording. Auphonic (0.84), ElevenLabs (0.83), Descript Studio Sound (0.72) and Adobe Podcast (0.60) stayed closer to the original recording than Diffio 4.6 Pro (0.58). This score is measured against the noisy original, so it is highest for tools that change the recording least; it does not show which output sounds more like the speaker.
| Tool | SCOREQ | Word error rate |
|---|---|---|
| Diffio 4.6 Flash | Better on 23, worse on 36, tied on 0 | Better on 23, worse on 9, tied on 27 |
| Adobe Podcast | Better on 8, worse on 51, tied on 0 | Better on 11, worse on 42, tied on 6 |
| Cleanvoice | Better on 5, worse on 54, tied on 0 | Better on 15, worse on 36, tied on 8 |
| Auphonic | Better on 2, worse on 57, tied on 0 | Better on 17, worse on 31, tied on 11 |
| ElevenLabs | Better on 0, worse on 59, tied on 0 | Better on 16, worse on 35, tied on 8 |
| Descript Studio Sound | Better on 0, worse on 59, tied on 0 | Better on 11, worse on 43, tied on 5 |
"Better" means a higher SCOREQ or a lower word error rate than Diffio 4.6 Pro on the same clip. Word error rate ties are common on short, clear clips where both outputs were transcribed the same.
How we ran the test
Test set
59 English speech clips, about 29 minutes in total, each up to 30 seconds long. 29 are archived sermon recordings with tape hiss, hum and room sound, sampled evenly from poor to fair quality. The other 30 are real recordings, all public domain, openly licensed or cleared for Diffio use, in seven groups: echoey rooms (5), phone and remote calls (3), wind and traffic (7), crowds and chatter (5), hum, machines and distortion (4), narration over music (1) and historic recordings (5). There are no synthetic noise mixes. 60 clips were selected; one was excluded after review (below).
Excluded clip vfc-51622: Congregational singing, not speech. Both Diffio models returned silence, which is the correct speech-only result; the clip does not measure speech enhancement. Its input and every output are kept with the data.
Tools and settings
Each tool ran once in its strongest speech cleanup mode at default amounts, with anything that cuts or moves audio turned off so every output stays lined up with its original. Each tool is scored on the file it returns to a customer.
| Tool | How it ran |
|---|---|
| Diffio 4.6 Pro | Public API, model diffio-4.6-pro, restored speech only. Returns MP3 192 kb/s. |
| Diffio 4.6 Flash | Public API, model diffio-4.6-flash, restored speech only. Returns MP3 192 kb/s. |
| Adobe Podcast Enhance Speech v2 | Web app (Premium), Enhance Speech v2, speech stem at 100% enhancement. |
| Cleanvoice | API, noise removal and Studio Sound on, all cutting and normalization off. Returns WAV. |
| Auphonic (Speech Isolation) | API, Speech Isolation denoiser at default amounts; leveler, loudness, filtering and cutting off. Returns WAV. |
| ElevenLabs Voice Isolator | API audio isolation (no settings exist). Returns MP3 320 kb/s. |
| Descript Studio Sound | API, Studio Sound at default intensity applied by an Underlord edit, nothing else changed. Returns AAC 160 kb/s. |
Joined uploads for Cleanvoice and Descript
Cleanvoice bills each file rounded up to a whole minute, and Descript bills media minutes per import plus AI credits per edit. So for those two tools every clip was joined into longer files, separated by short gaps of digital silence, and the outputs were cut back at the exact offsets. Both tools returned files of exactly the input length. Every other tool received each clip as its own file. Descript's API applies Studio Sound through an agent edit and returns audio only after a private publish, so that is how it ran.
Metrics
All three metrics run on the same mono decode of each output, with no change in level. SCOREQ is reported as the gain over the original clip. Word error rate uses a deliberately small recognizer on the outputs so it stays sensitive to restoration quality; it is a proxy, not a human transcript. Speaker similarity is reported as similarity to the original recording, not as proof that a voice was preserved.
Repeatability
A handful of clips were run again for Auphonic and both Diffio models. Auphonic matched its first run every time. Diffio's repeats were very close but not always identical, so Diffio's output is not fully deterministic. Cleanvoice, ElevenLabs, Descript and Adobe were not repeated, to save credits or because the run was manual.
Limitations
- Every score is model-predicted. There was no human listening study.
- Each tool ran once, with its default amounts. A tool tuned by hand may do better on a given recording.
- Return formats differ (some tools return lossy MP3 or AAC), and each is scored as returned.
- Cleanvoice and Descript processed joined files; the other tools processed single clips.
- Some groups are small, and the music group is a single clip. Treat per-group differences with fewer clips as indications, not rankings.
- Diffio ran this test and makes one of the tools. The inputs, settings and per-clip scores are documented and available on request so the results can be checked.
Data
Every score on this page is in the tables above, and every clip plays with its credit. The inputs, each tool's exact settings, the transcripts and the per-clip scores are documented and available on request; contact us.
Diffio is not affiliated with or endorsed by Adobe, Descript, Auphonic, ElevenLabs or Cleanvoice. Product names are trademarks of their owners.
FAQ
Which AI speech enhancer is best?
In this test, Diffio 4.6 Pro had the largest average quality gain (SCOREQ +2.31) and the lowest word error rate (10.5%) across 59 real recordings. It did not win everywhere: other tools did better on word error rate for phone and remote calls, and on quality gain and word error rate for narration over music. Results depend on the recording, so test the file you need to fix.
Is Adobe Podcast Enhance good?
For sound quality, yes: Enhance Speech v2 raised SCOREQ by 1.87 on average, fourth of the seven runs. Its word error rate was 18.2%, higher than the original recordings' 13.2%, so on average the small recognizer understood its output less well than the untouched audio. Its free tier is also the most generous for short audio-only files.
How was speech quality measured?
With three automatic measures and no human listening study. SCOREQ (no-reference speech quality model; higher is better). Gain = output minus the original clip. Word error rate of a small speech recognizer (faster-whisper small.en) on each output, against MAI-Transcribe-2 on the original; lower is better. ECAPA speaker-embedding similarity to the original recording, noise included; tools that change less score higher.
Why does Diffio score low on similarity to the original?
Similarity is measured against the original recording with its noise and echo still in it. A tool that removes more of the room and background moves further from that recording, so it scores lower. It is reported so you can see how much each tool changes the audio, not as a measure of how natural the voice sounds.
Who ran this test?
Diffio ran it on October 22, 2026, so it is not an independent study. Every tool ran once with the settings listed in the method, every clip plays on this page, and the inputs, settings and per-clip scores are documented and available on request.
Can I hear every tool's output?
Yes. Every clip on this page plays the original and each tool's output, loudness-matched so you can compare them fairly.
Your recording is the test that matters
Upload the file you need to fix and compare Diffio's result with the original before you pay.
Test Diffio free on your recording →