Same voice?
Does it still sound like them?
We built 36 test voicemails where we also have the clean recording of the same words, then measured how close each cleaned copy sounds to the real voice.
- In the part the phone carried
- 28 of 36
- cleaned copies closer to the real voice (0.70 → 0.79)
- Over the whole file
- 14 of 36
- measured less like the real voice than the phone file
- Older men talking
- 6 of 6
- measured less like themselves (0.68 → 0.61)
All 36 as a table
| Clip | Voice | Situation | Phone file | Diffio |
|---|---|---|---|---|
| vm03 | Reader | room | 0.52 | 0.70 |
| vm05 | Reader | handling | 0.63 | 0.66 |
| vm06 | Reader | distance | 0.50 | 0.56 |
| vm07 | Reader | quiet | 0.66 | 0.70 |
| vm10 | Reader | car | 0.57 | 0.70 |
| vm11 | Reader | handling | 0.44 | 0.61 |
| vm12 | Reader | distance | 0.52 | 0.73 |
| vm13 | Reader | quiet | 0.56 | 0.66 |
| vm16 | Reader | car | 0.66 | 0.69 |
| vm17 | Reader | handling | 0.61 | 0.67 |
| vm18 | Reader | distance | 0.44 | 0.45 |
| vm20 | Reader | quiet | 0.47 | 0.51 |
| vm21 | Reader | room | 0.43 | 0.61 |
| vm23 | Reader | handling | 0.65 | 0.74 |
| vm24 | Reader | distance | 0.37 | 0.59 |
| vm25 | Reader | quiet | 0.51 | 0.57 |
| vm26 | Reader | quiet | 0.60 | 0.65 |
| vm27 | Reader | room | 0.38 | 0.54 |
| vm28 | Reader | car | 0.42 | 0.51 |
| vm32 | Reader | quiet | 0.45 | 0.68 |
| vm34 | Reader | car | 0.51 | 0.69 |
| vm35 | Reader | handling | 0.76 | 0.83 |
| vm01 | Older man talking | quiet | 0.79 | 0.74 ↓ |
| vm02 | Reader | quiet | 0.81 | 0.66 ↓ |
| vm04 | Reader | car | 0.47 | 0.42 ↓ |
| vm08 | Older man talking | quiet | 0.63 | 0.57 ↓ |
| vm09 | Reader | room | 0.57 | 0.51 ↓ |
| vm14 | Reader | quiet | 0.56 | 0.56 ↓ |
| vm15 | Older man talking | room | 0.74 | 0.63 ↓ |
| vm19 | Reader | quiet | 0.61 | 0.53 ↓ |
| vm22 | Older man talking | car | 0.56 | 0.47 ↓ |
| vm29 | Older man talking | handling | 0.76 | 0.68 ↓ |
| vm30 | Reader | distance | 0.51 | 0.43 ↓ |
| vm31 | Reader | quiet | 0.70 | 0.66 ↓ |
| vm33 | Reader | room | 0.51 | 0.50 ↓ |
| vm36 | Older man talking | distance | 0.63 | 0.55 ↓ |
A phone keeps only the middle of a voice, about 300 to 3,400 Hz. Diffio rebuilds the rest so it sounds full again. That rebuilt part is a good guess, not a recording of the person, and on some voices it measured less like them. Keep both copies.
Hear the two that measured less like the speaker
Dashed outline: a simulated mix, not a real recording.
Landline call with dropoutsSimulated mix
An older woman's voice, with 9% of the call cut out in bursts.
Clearer (2.11 → 4.31) and fewer word errors (1.8% → 0.9%), but the biggest drop of all 36: 0.81 → 0.65 like her real voice. Keep the original.
Read what's said
Which should proceed uniformly in a straight line, that is, which should preserve the same direction and the same velocity, we would say that the cause of this continuation of motion is not to be found out of the body, but that it is founded in its very nature, and that in virtue of its inertia, it remains always in the same state.
Source: elizabethhousworth, LibriVox (librivox.org), "Letters on Different Subjects in Natural Philosophy", public domain.
Source & method
License: Public domain. Source recording. Full credits. Made by sending a clean recording through a real phone codec; not a real voicemail. Cleaned once with Diffio 4.7 Pro on November 19, 2026. All three versions play at the same loudness; "Real voice" is the clean recording the test voicemail was made from.
Same words?
Does it change what they said?
7words in 3,187really changed, in 4 of 36 clips. All 4 had car or room noise. None of the 12 quiet clips changed a word.
- Real word change 6
- Repeats they really said 9
- Spelling or number format 10
- Our reference transcript was wrong 3
- Place name, old word or unclear 4
| Test voicemail | What was said | Cleaned copy reads |
|---|---|---|
| Car noise, mobile call | "the parched mouth" | "the parts to mouth" |
| Car noise, mobile call | "he has passed suffering" | "he is past suffering" |
| Car noise, weak signal | "which exited into the ocean" | "which edged it into the ocean" |
| Kitchen noise, Wi-Fi call | "fortune" | "fortunes" |
| Car noise, Wi-Fi call | "then the mass" | "then if the mass" |
A speech recognizer read these; a listener may hear them differently, which is why the audio is published. The car-noise clip above has the first two.
Clearer
How much clearer it gets
- Sounds clear (SCOREQ, out of 5)
- 2.19 → 4.74
- the clean recordings scored 4.21
- Fuller voice
- 4.5 → 19.9 kHz
- how high the voice reaches (median); a phone stops near 4 kHz
- Word error rate
- 5.5% → 3.4%
- weak-signal AMR calls: 11.8% → 3.1%
Word error rate counts the words a speech recognizer gets wrong against the clean recording's transcript. In quiet rooms the phone file was already easy to follow (2.1%); the cleaned copy came out slightly worse (2.9%).
On real phone and Zoom calls
Our test voicemails are simulated. Our speech enhancer test also has 3 real call recordings, with every tool run the same way.
| 3 real calls | Clear (1–5) | Word errors | Same voice (0–1) |
|---|---|---|---|
| Original calls | 2.97 | 3.8% | 0.93 |
| Diffio | 4.58 | 5.0% | 0.72 |
| Cleanvoice | 4.48 | 3.5% | 0.66 |
| Adobe Podcast | 4.43 | 5.3% | 0.70 |
| Auphonic | 3.52 | 4.2% | 0.89 |
| ElevenLabs | 3.13 | 4.2% | 0.91 |
| Descript | 3.20 | 5.3% | 0.83 |
Tools: Diffio 4.7 Pro, Cleanvoice, Adobe Podcast Enhance Speech v2, Auphonic Speech Isolation, ElevenLabs Voice Isolator, Descript Studio Sound. Diffio sounded clearest on these calls but got slightly more words wrong than the untouched calls (5.0% vs 3.8%), and Auphonic and ElevenLabs kept the voice much closer. 3 clips only.
Podcast guest on the phone (real call)
Source & method
A guest calling in to a NOAA podcast in December 2015: a real phone line, not a simulation. Source: "Diving Deeper: Exploring Hawaii's Coral Reefs" (December 2015), NOAA Ocean Service podcast, public domain. No endorsement implied.
License: Public domain. Source recording. Restored with Diffio 4.7 Pro on November 3, 2026; Cleanvoice and ElevenLabs were run on the same clip with the settings listed in the test, for our 59-clip speech enhancer test. Every version plays at matched loudness.
Options
Which option fits
A free cleaned copy you can download today
Adobe Podcast Enhance Speech (free)
Free download. In our test it changed more words: word errors 5.5% → 13.1%, worse than the phone file on 30 of 36. The free plan refuses .amr and treats .3gp as video; convert to WAV first.
Nothing rebuilt, just a little less hiss
Audacity noise reduction or ffmpeg's afftdn
Adds nothing the phone didn't carry, so it can't change the voice, but barely sounds clearer: 2.19 → 2.16 out of 5 (afftdn).
The clearest copy that changes the fewest words
Diffio
3 free previews to listen; the download needs Basic ($8 a month). 4.74 out of 5; 7 changed words in 3,187.
Avoid
Open-source denoisers at default settings on noisy phone files
DeepFilterNet3 returned silence on 7 of our 36 voicemails, including all 6 with car noise.
Limits
What cleaning can't do
It rebuilds, it doesn't recover. A phone drops everything above about 3,400 Hz. Diffio rebuilds it; that part is a good guess, not a recording of the person.
Some voices come out less like themselves. All 6 older men talking naturally (0.68 → 0.61); 14 of 36 clips overall, by up to 0.16.
A few words can change in noisy messages. 7 of 3,187 words, all in clips with car or room noise. None in quiet clips.
Lost audio stays lost. 5 clips with dropouts: clearer (4.65 out of 5), but less like the voice on 3 of the 5.
Quiet, clear messages gain little. Word errors in quiet rooms went from 2.1% to 2.9%.
Our test voicemails are made, not real. Real codecs (G.711, GSM, AMR, Opus) on a simulated network; real carriers add processing we couldn't test.
English works best. Other languages run, with varying quality.
We make Diffio. Every test file, both versions and the scores are listed in the results file, so you can check.
How to
Clean up a voicemail in 3 steps
- 01
Get the voicemail off your phone
- iPhone: open the voicemail, tap Share, then Save to Files.
- Android (Phone by Google): open Voicemail, expand the message, tap Send to.
- Samsung: Archive saves it to My Files > Internal Storage > VisualVoicemail.
- Google Fi: download voicemails as MP3s at takeout.google.com.
- 02
Upload a copy to Diffio
- Keep the original where it is. M4A, AMR, 3GP, MP3 and WAV all work.
- 03
Listen side by side, keep both
- Play the phone file and the cleaned copy back to back. Let the people who knew the voice decide.
Full export steps for each phone: how to save a loved one's voice.
Plans
Free to listen, paid to download
Free
$0
Try Diffio on your own recordings.
- 3 uploads in total (not monthly), audio or video
- Preview the first 2 minutes of each result
- Diffio 4.7 Flash model
- No full playback, download or export
Basic
$8per month
For hobbyists and small teams.
- 6 hours (360 minutes) of audio or video a month
- Full playback, download and export
- Diffio 4.7 Flash and 4.7 Pro. Developer model docs
Pro
$29per month
For professionals and businesses.
- 30 hours (1,800 minutes) of audio or video a month
- Full playback, download and export
- Diffio 4.7 Flash and 4.7 Pro. Developer model docs
Diffio costs $1.60 an hour on Basic at 5 hours a month ($8 a month for up to 6 hours). A voicemail counts by its length.
Questions
Questions about cleaning up voicemails
Can you make an old voicemail clearer?
Usually, yes. On 36 test voicemails, Diffio raised the clarity score from 2.19 to 4.74 out of 5 and cut speech recognizer errors from 5.5% to 3.4%. It rebuilds the part of the voice the phone dropped, so keep the original too.
Will it still sound like my mom or dad?
In the frequencies the phone carried, the cleaned copy measured closer to the real voice on 28 of 36 clips. Over the whole file, 14 of 36 measured slightly less like the person, including all 6 recordings of older men talking. Only the people who knew the voice can judge, so keep both versions.
Can AI add words my loved one never said?
It can change a word now and then in noisy messages. We read every changed word in our test: 7 of 3,187 words really changed, all in clips with car or room noise, none in quiet ones.
What file format are voicemails?
It depends on the phone and carrier. Google Fi documents MP3 downloads; Google Voice MP3 downloads are reported by users but not documented by Google. iPhone shares voicemails as audio files (commonly .m4a, though Apple doesn't state it). Visual voicemail systems can use AMR or WAV. Diffio accepts MP3, M4A, WAV, AAC, AMR and 3GP.
Is it free?
You can upload 3 voicemails free and listen to a 2-minute preview of each cleaned copy, made with Diffio 4.7 Flash. Downloading needs a paid plan: Basic is $8 a month for 6 hours, $1.60 an hour at 5 hours a month, and includes Diffio 4.7 Pro, which this page's tests used. Adobe Podcast's free Enhance Speech gives a free download. See plans.
3 more questions
Will cleaning the voicemail change the original file?
No. You upload a copy; the original stays on your phone or computer. Projects stay in your Diffio account until you delete them.
Can you fix parts where the call cut out?
Partly. Diffio smooths short dropouts and makes the rest clearer, but it can't bring back the exact sound that was lost; on our 5 dropout clips the cleaned copy measured less like the real voice on 3.
How do I save a voicemail forever?
See our step-by-step guide: how to save a loved one's voice. For tapes and cassettes, see restoring old recordings.