Google reduced number of speech recognition errors by 50%

Recently, Google has posted a message that said that they were able to improve the speech recognition system, reducing by half the number of recognition errors. Today, Google explained how the developers managed to get such a result.

Google reduced number of speech recognition errors by 50%

In short, Google rebuilt transcription system. The old system was based on a general model of machine learning, also known as a Gaussian mixture model. The new version uses a type of artificial intelligence with the so-called «deep learning»Built on short-term memory of recurrent neural networks.

Artificial neural network can be trained using more data, such as voice mail. In this case, learning artificial intelligence by Google received a lot of voice messages from their users.

Google uses a method «deep learning»To improve its other services, including the Google Translate. At the conference, Google I / O, the current CEO of Google Sundar Pichai announced that thanks to advances in «deep learning», Speech recognition error is only 8 percent.

Now that Google Voice has become more accurate speech recognition, Siri in services and Cortana were serious problems.

Bookmark the permalink.

Leave a Reply