IBM speech recognition creates a new world record, Huawei's European patent application jumps second

Editor's note: According to foreign media reports, Microsoft's speech recognition has made a major breakthrough, and this technology has reached the human level in the level of recognition of words in dialogue. The latest news, now IBM announced the creation of a new industry record: 5.5% error rate. This is a very difficult speech recognition task: record the daily conversations between people, such as "buy a car", and calculate the results. At the same time, in the field of Internet of Things and mobile communications, Samsung, Intel, Apple, and Huawei are all deploying different patents. The latest Huawei patent application in Europe has reached a new high. Xiaobian organizes the latest reports for everyone to share.


March 10th news, according to IBM's official website, when people talk, each other will miss or misread 1-2 words for every 20 words. In a 5-minute conversation, you may get 80 words wrong. But most of us have no problem in understanding and speaking. However, the computer is different. Last year, IBM announced a major achievement in speech recognition in a natural conversation environment: the development of a system with a word error rate of 6.9%.

Since then, the company has continued to make progress. Now IBM is announcing a new industry record: 5.5% error rate. This is a very difficult speech recognition task: record the daily conversations between people, such as "buy a car", and calculate the results. This corpus of records is called "SWITCHBOARD" and has been used to test speech recognition systems for more than 20 years.

In realizing this breakthrough, IBM researchers focused on applying deep learning techniques to combine LSTM (long-short-term memory) and WaveNet language models with three powerful acoustic models. Among the three acoustic models used, the first two are bidirectional 6-layer LSTMs, one of which is multi-feature input and the other has conversational multi-task learning capabilities. The last model has a unique place that not only learns from positive examples, but also uses negative examples, so it becomes more and more clever and performs better when repeated similar styles of speech.

Achieving the same level of humanity – the error rate is comparable to that of two people – has long been the ultimate goal of the industry. Others in the industry are also struggling to catch up with IBM's record, with some recently claiming to reach 5.9%. In the process of achieving today's achievements, IBM found that the same level of human error should be 5.1%. In determining this number, IBM worked with partner Appen to reproduce the results of the human level. Although IBM achieved a 5.5% error rate is a big breakthrough, but found that human equivalent level is 5.1% to prove that technology has to reach the same level as humans.

In the study, IBM contacted different industry experts to let them comment on the matter. Yoshua Bengio, director of the University of Montreal's MILA lab, agrees that IBM still has a lot of work to do to achieve the same level of humanity. IBM realized that the standard of finding humans at the same level was more complicated than originally thought.

DC Support Capacitors

China leading manufacturers and suppliers of DC Support Capacitors,DC Capacitor, and we are specialize in Electrolytic capacitor,High Voltage Capacitor, etc.DC Support Capacitors

DC Support Capacitors,DC Capacitor,Electrolytic Capacitor,High Voltage Capacitor

YANGZHOU POSITIONING TECH CO., LTD. , https://www.cnpositioning.com

Posted on