It’s an AI-fueled “audio enhancer” with one key goal in mind: to provide speech restoration for uploaded audio and video recordings, including interviews recorded in locations with lots of noise. Distorted phone interviews and archival material in need of a boost is also a selling point for Diffio, the creation of Nathan Harmon.
A web application, Diffio uses diffusion neural networks to process speech and separate it from background sound. The service operates on an existing recording after capture.
For radio and television producers, that includes material recorded outside their own microphone setup, such as a viewer-submitted video or a remote interview. It complements microphone placement and recording practices rather than replacing them. Users upload a file through the browser; processing takes place in the cloud. The player lets an editor compare the original and restored speech and adjust the amount of background sound in the mix. Restored speech, background sound and a mix can be downloaded separately. Video processing retains the picture while restoring the speech soundtrack.
As Harmon sees it, Diffio focuses on speech, not music restoration. “Crowd voices and other speech-like sounds can sometimes be assigned inconsistently between the speech and background outputs,” he says. As such, results depend on the source recording, and editors should compare the output with the original before broadcast or publication. “A cleaner-sounding result does not by itself establish that words have been preserved accurately,” he notes.
The application includes Diffio 4.5 Flash and Pro model options, and Diffio also offers a developer API for integrating speech restoration into other workflows.
Independently funded Diffio grew from work on restoring older sermon recordings and now addresses spoken-word audio and video. It is a service provided by Ozni Apps LLC.
Product information and original/restored audio examples are available at diffio.ai.
