Google has announced the rollout of "an end-to-end, all-neural, on-device speech recognizer to power speech input in Gboard", the company's keyboard with Google Search baked in.
The technology could give Google an edge over Siri and Alexa in convincing people to talk to machines through phones and home speakers that can deliver answers faster, by cutting down the latency that comes with sending a request from a device to a remote server and waiting for a response.
The company has enabled on-device voice recognition by miniaturizing a machine-learning model that can do the task on a phone rather than handing off the job to a server in the cloud.
According to Google researchers, the model works at the character level, so as the user enunciates a word, the machine repeats it one character at a time, exactly how an expert human transcriber would type.
Beyond supreme low-latency speech recognition, Google wanted its system to exploit "on-device user context", such as the user's list of contacts, music apps to provide a list of song names they might be referring to, and location.
To achieve the on-device intelligence, Google employed a Recurrent Neural Networks (RNN) transducer aided by a recent innovation called 'Connectionist temporal classification' that's used for training neural networks. The technique allowed for a more efficient manner for machines to interpret speech.
Google explains that the speech-recognition engine would normally depend on a search graph that can be 2GB in size, which would be onerous if stored on a device.
Instead, it trained a neural network that provides the same accuracy as a client-server setup that was just 450MB in size. Not happy with that, the Google researchers shrunk the model to just 80MB.
"Our new all-neural, on-device Gboard speech recognizer is initially being launched to all Pixel phones in American English only," Google researchers said.
"Given the trends in the industry, with the convergence of specialized hardware and algorithmic improvements, we are hopeful that the techniques presented here can soon be adopted in more languages and across broader domains of application."
Researchers at Google Brain and DeepMind go in quest of better "representations" of the world by AI, through exploration of the polytope, a Euclidean geometric form that represents the possible solutions to a game of strategy.