Neural Networks Helped Reduce by 50% the Google Voice Transcription Errors

The Google has always used some sort of artificial intelligence in the service, but in recent years, so-called neural networks have gained more space in the company. An example was given on Tuesday (11): the company explained how neural networks helped to significantly reduce the amount of Google Voice errors.

Can you sum up Google Voice as a service phone online. One of the features offered (at least in English – speaking countries) is the transcription of messages: you get a voice mail from a friend, for example, and the tool turns it into text.

This function has many uses, but came up against a problem: not infrequently, the transcripts had so many mistakes that it was simply impossible to understand the message. To alleviate the problem, Google had to redo the transcription system.

The first change was made in 2012 when the company started using neural networks of deep learning, a type that essentially works with several layers of “neurons” and therefore can make associations between different parameters. This technique improved speech recognition engines enough.

But Google says that things only improved the truth some time later with the implementation of recurrent neural networks in LSTM standard (Long Short Term Memory). This type basically has connections in cycles and memory cells that allow the network to “remember” the data analyzed previously.

Neural networks require a considerable amount of data to learn. Google could even have used the transcripts made with the old algorithm, but that database was already “contaminated” with several recognition errors. The company then had to rely on the help of Voice users: thousands of them agreed to disclose your messages based on Google’s promise that this content would be used solely for system training.

With this volume of messages, Google engineers were able to submit the algorithms to various models of acoustic recognition and language. One score was employed to mitigate failures, for example. Before, Google Voice did things like transcribing the phrase “I received the message you left me” and “I received the message. You left me”. Can you imagine the confusion?

Fortunately, neural networks have paid off. After several “rounds” of training, many of which again (performed with the same data), learning made Google Voice decrease the number of errors in nearly 50%.

Advances should not end there. The algorithms are still learning and, of course, Google maintains the efforts to improve the tool. This is the kind of service that requires continuous work.

Note that the Voice is not the only recent example of how the neural networks are making a difference in Google. Late last month, the company revealed how the idea became the Google translator smarter in translations from images.