Diffio Charges $8 a Month to Rebuild Phone Audio Near Studio Quality

An AI speech restoration tool called Diffio charges $8 a month for six hours of audio, and its founder claims it beats Adobe Podcast, ElevenLabs and Audacity on the hardest recordings. The company, founded in Colorado Springs in 2025 under Ozni Apps, LLC, positions itself as the cheap middle ground between a phone recording and a professional studio.

"Diffio probably doesn't get you all the way there, but it gets you pretty close, and it's very cheap," founder Nathan Harmon said. "You can just use whatever equipment you've got, including your phone." Harmon runs the company alone and spent about a decade processing electromagnetic signals, much of it as a machine learning researcher at Northrop Grumman, before turning those methods on audio after struggling with old sermons recorded on cheap equipment.

What does the tool actually change for creators?

Unlike the noise removal built into most editing software, Diffio attempts to rebuild speech that noise has destroyed. "It'll actually rebuild what they said based on the other areas of the audio that you can hear them speaking correctly in, and it'll handle packet loss," Harmon said. Standard removal strips static or clapping but usually leaves echo, which he calls harder to fix because the voice reaches the microphone several times. "The algorithm has to determine which of those was actually the primary source and then remove the other ones," he explained.

The tool returns restored speech plus a separate background track, which also holds any music. Harmon recommends mixing some background back in for outdoor footage. "It's very disconcerting to someone watching your video, because it's like, 'Well, I can see that there's wind and I can't hear any of it,'" he said.

Rebuilt words are partly generated, and the share depends on how much of the original survived. Clean audio passes through unchanged. Dropped packets on a call usually last under a second, so Diffio can rebuild them with high confidence. "It's not an absolutely guaranteed accurate reconstruction, but it is a probabilistically correct reconstruction," Harmon said. When nothing in the recording reveals the speaker's voice, the limit arrives: "It probably will reconstruct some speech, but it'll probably not sound like the speaker."

Early customers include small churches cleaning up sermons, restored wedding videos shot on guests' phones, and one archivist who keeps a restored copy beside each original. Harmon sees starting-out creators as a fit, since they often cannot control where they record and good equipment is expensive early on. He does not yet know how far the product has reached into the Creator Economy.

What should creators check before posting?

Harmon recommends listening to every restored file, especially the noisiest passage in the middle, because that is the hardest part to reconstruct. Language is the second check: English tolerates the most noise because Diffio has the most training data for it, followed by Spanish and Portuguese. Hindi works only on cleaner recordings. A Spanish-language clip that Net Influencer sent ahead of the interview failed on the current production model and came through on an internal model Harmon says is close to release.

Failures can go to support, which usually responds right away and will try internal models, offering extra copies or free credits. Users must opt in before uploads can train models, though Harmon says the company has not done so, because it needs high-quality audio and users upload low-quality audio. Opting out shuts off training and metrics; deleted files leave all servers with no recovery. Outputs carry no restrictions, including for benchmarking.

How solid are the benchmark claims?

Harmon's claims are his own. By his count, Diffio ran real clips through about five competitors, measuring predicted speech quality and word accuracy, and published comparative results covering 59 recordings. At the time of the interview he could not publish benchmarks because of competitors' terms, and he suggested creators run free trials side by side. The free trial covers three uploads, previewing the first two minutes of each.

He concedes limits. "In general, if your audio is not really distorted, you probably can get away with using anyone," he said. Adobe Podcast adds a dedicated music track Diffio lacks: "That is something that Adobe Podcast currently does better than us," he acknowledged. ElevenLabs separates background noise without reconstructing speech, in his telling.

The pricing rests on unusual hardware. GPU costs kept climbing as major AI labs absorbed supply, so Diffio shrank its model to run inference on used, self-hosted Mac minis in Colorado Springs. The switch lets it charge current prices at a margin: Basic at $8 a month for six hours, Pro at $29 for 30 hours. The same hardware limits the recently launched API, which has seen little use. "Somebody throws a thousand jobs at us and, because we're hosting our own hardware, we cannot just scale our compute to absorb a giant job like that," Harmon said.

Next on the front burner is a new model with better multilingual support, plus the option for users to supply a room environment to emulate. Harmon's five-year view: recording to publishing in a few steps at a computer. "That process is what stops me from creating content," he said. "It's just so much work, so much time."