AI-powered service wants to make manual transcription a thing of the past

Baidu Research Labs announced yesterday that it is launching SwiftScribe, its first AI-powered transcription software.

SwiftScribe aims to make the job easier for professional transcribers who spend hours manually transcribing audio. By pairing speech recognition software with a simple editing suite, the service claims to increase productivity and streamline workflow.

The software takes Baidu’s speech recognition engine, Deep Speech 2, and joins it with editing features that tackle the “biggest obstacles users face when transcribing”.

“We designed intuitive shortcut keys and innovative human-computer interaction to solve the problem of discontinuity,” writes Tian Wu, SwiftScribe’s project manager.

The software’s biggest asset is its ability to link text with audio, so that whenever the cursor moves in the text, the audio adjusts with it.

SwiftScribe also adjusts according to what the transcriber has edited — meaning the system should eventually understand how to spell tricky names and the like.

Baidu Research launched SwiftScribe, which uses artificial intelligence to transcribe audio snippets

“Its neural network, which is trained on thousands of hours labeled audio data, learns to associate sounds with certain words and phrases,” writes Wu. “One big advantage of this approach is as users transcribe and make edits, the system can learn and improve along the way.”

Baidu claims that transcribers who used SwiftScribe have seen a significant cut to their work hours.

“It typically takes between four to six hours to transcribe one hour of audio data,” Wu writes. “Using SwiftScribe, the time a transcriptionist spent on a project is on average cut down by 40 percent.”

The product is designed for professionals for whom transcriptions are part of the job, and Baidu believes it could impact several industries including healthcare, law enforcement, business and media.

The research lab is accepting requests from transcribers to be part of the beta testing, but in the meantime SwiftScribe’s website hosts a small trial for those interested.

Featured image: ian munroe via Flickr (CC 2.0, resized)