AI software can detect the accent of a person on one end of an online or telephone conversation and modify the accent of a person responding through a phone or computer microphone to match it in real time. It is hoped that the technology will enable greater understanding and clarity in a range of interactions, including customer support, education and telemedicine.
US-based startup Sanas developed software that runs locally on a smartphone or laptop rather than on a central server, as virtual assistants like Amazon Alexa do, which it says minimises lag and improves security. The company says that audio can be manipulated to match an accent with a delay of just 200 milliseconds.
The software intercepts sounds between the microphone and whichever software is being used to communicate, whether that is Zoom, Microsoft Teams or Skype, and alters what is broadcast. It converts accents using a neural network that has been trained on a range of audio clips, including files found online, recordings made of strangers on the street in cities around the world and recordings made by call centre staff at partner companies.
Advertisement
at Sanas says that some elements of the AI system are a 鈥渂lack box鈥� and that the way it alters voices isn鈥檛 known. He says the team quickly realised during development that the AI couldn鈥檛 be designed to transcribe speech to text and then generate an artificial voice to read it in a new accent, as errors in transcription would occur and there would be a lag in computation that rendered it useless for conversation.
The team eventually settled on a model that is based on聽individual phonemes 鈥� the distinct sounds that make up words 鈥� and therefore was able to begin changing the stream of audio before each word was finished.
P茅rez Soderi says that Sanas is in talks with several businesses to roll out the technology in coming months.
