Google AI researchers working with the ALS Therapy Development Institute right now shared particulars about Project Euphonia, a speech-to-text transcription service for folks with talking impairments. They additionally say their strategy can enhance computerized speech recognition for folks with non-native English accents as properly.
People with amyotrophic lateral sclerosis (ALS) usually have slurred speech, however current AI techniques are sometimes skilled on voice information with none affliction or accent.
The new strategy is profitable primarily as a result of introduction of small quantities of information that represents folks with accents and ALS.
“We show that 71% of the improvement comes from only 5 minutes of training data,” based on a paper published on arXiv July 31 titled “Personalizing ASR for Dysarthric and Accented Speech with Limited Data.”
Personalized fashions have been capable of obtain 62% and 35% relative phrase error fee (WER) enchancment for ALS and accents respectively.
The ALS speech information set consists of 36 hours of audio from 67 folks with ALS, working with the ALS Therapy Development Institute.
The non-native English speaker information set is known as L2 Arctic and has 20 recordings of utterances that final one hour every.
Project Euphonia additionally makes use of strategies from Parrotron, an AI device for folks with speech impediments launched in July, in addition to fine-tuning strategies.
Written by 12 coauthors, the work is being offered at International Speech Communication Association, or Interspeech 2019, which takes place September 15-19 in Graz, Austria.
“This paper’s approach overcomes data scarcity by beginning with a base model trained on thousands of hours of standard speech. It gets around sub-group heterogeneity by training personalized models,” the paper reads.
The analysis, which a Google AI weblog submit highlighted right now, follows the introduction of Project Euphonia and different initiatives in May, akin to Live Relay, a characteristic to make telephone calls simpler for deaf folks, and Project Diva, an effort to make Google Assistant accessible for nonverbal folks.
Google is soliciting data from people with ALS to enhance its mannequin’s accuracy and is engaged on subsequent steps for Project Euphonia, akin to utilizing phoneme errors to scale back phrase error charges.