Mobile client "Today's last class from Beijing to Shanghai when the high-speed rail departure?" Mr. A once again encountered a flight canceled, desperation took out his cell phone, a friend recommended a name "go out to ask" WeChat public account issued the above instructions for help, just a few seconds, "go out to ask" to give the alternative answer.

This is just one of the trivial daily scenarios in the language wave that Mobile Internet has opened.

Every day, hundreds of millions of voice messages are converted into words by machines using speech recognition technology; some of them are parsed out in a machine-learning manner for specific meanings. In human-computer interaction, voice search for users Provide the answer

According to statistics, Google 25% of the mobile search results from the voice, this ratio also exceeded 10% in Baidu; Siri before and after the emergence of a number of third-party voice assistant is rapidly occupying the domestic intelligent terminal to provide users with all kinds of information query service and Siri simple entertainment features; and in the smart TV, navigation, language learning and other fields away from the public voice companies are providing the most basic technical support.

However, for half a century's idea of ​​artificial intelligence, the arrival of intelligent voice has actually come a bit late.

The most crucial boost comes from cloud formation. "In the past, the huge amount of computation constituted the threshold for increasing the accuracy of recognition and also restricted the identification of application scenarios (often only dedicated areas). Nowadays, the convenient access to terminals such as cloud computing and mobile Internet enables voice recognition Increasingly becoming a universal service capability, "Liu Wei, executive director of Legend Star, which focuses on artificial intelligence technologies such as face recognition and voice analysis, told reporters.

In the next three to five years of disclosure disclosed by China's Ministry of Industry and Information Technology, intelligent voice technology and industrial promotion are the focus of work. "Smart Voice has truly become a mobile Internet portal" was officially put forward by the government.

However, for the industry's entrepreneurs, how to bring innovative technology to the market, it is a protracted battle. From the earliest boarding in the capital market, Itexamworld, to later distribution in speech recognition and semantic analysis of the part of the company, it is inevitable in the 2B and 2C business model between the struggle. Is to do horizontal technology service providers, or do in-depth product provider? Entrance Currently, the voice of the various companies on the chain how to crack the industrialization problems?

Identification of the storm

Walking in a campus in a university in Beijing's Haidian, you may be tempted to stop crying and invite you to read the daily conversation with your hometown's accent on the mobile device in your hand.

The same scene, for a different text, repeated hundreds of times. This seemingly "bulky" work, but it is the beginning of intelligent voice pipeline.

In the first six months of the official establishment of the cloud technology company Voice Technology, the accumulation of basic voice data has been quietly carried out by outsourcing companies while recruiting troops and horses. These precious data collected offline can provide machines with more simulated learning samples.

Liang Jien-yun, co-founder and CEO of Cloud Knowing Voice, describes the process of voice recognition as: "capturing the user's voice through a microphone, converting the acoustic signal into a machine-processable 'pronunciation feature', combining the pronunciation dictionary with various vocabulary Combination of language models, compared to search out the closest sound waveform sentences. " Briefly speaking, the machine does not have to understand the meaning of the sentence, you can automatically convert the speech into accurate text.

This is the first pass that voice technology needs to break. In China, entrepreneurs engaged in speech recognition technology can be roughly divided into two "martial art", one from Tsinghua University, the other from the Chinese Academy of Sciences. Liang Jiagen where the CAS Institute of Automation from the 80s of last century to focus on the field of voice research, and Tsinghua University at the same time and started. According to an insider's article estimates, the country engaged in voice technology professionals no more than one hundred people.

Liang Jia-en witnessed in the university stage "with the door" - the rise of IFT. Before the outbreak of the mobile Internet, IETF and Jietong Huasheng focused on the field of speech synthesis. This technology, which was widely used after World War II, allowed the machine to read out texts. However, it later focused on speech recognition.

Not afraid of IDF and many other voice search class rival, cloud knowing with a set of known as the core technology of deep neural network, quickly gain a firm foothold. This technology enhances recognition in accent and noise environments, and can individually reduce the recognition error rate by more than 30%. And Si Bi Chi also uses deep neural network technology to achieve the performance of voice recognition, Baidu also set up a special deep nerve institute earlier this year to carry out research and development.

In Liang's opinion, the advantage of speech recognition lies in the integrity of the statistical framework. "Algorithms and frameworks are open to the public in academic circles and are not much different from them." However, in this case, , "The same system architecture, read 90% of the laboratory environment to achieve 90% recognition rate is easy, but in the mass user and practical environment to achieve 90% of the difficulty is still quite high," Leung told reporters.

Understand the confusion

"If we had only speech recognition, we could chat at most," said Leung, "adding semantic understanding can be tied to the real business."

In the downstream of voice industry, semantic analysis can be the relay of voice recognition. Briefly speaking, semantic analysis is to analyze the input sentence, understand the logical relationship of the sentence, and construct the feedback results needed by the user according to the logical relationship. The classic form of semantic analysis is quizzes or conversations - you need to understand the user's input and then generate the answer or generate a question that needs to be supplemented by the user.

Travel vertical search where to fill the form by the user, the natural semantic analysis is for users to fill in the form directly, go out and ask the founder Li Zhifei called for example. Semantic analysis translates text into standardized tables, leverages open API data support, and docks vertical searches.

Li graduated from the Johns Hopkins Language and Speech Processing Laboratory (CLSP). Before gaining investment in Sequoia Capital and True Fund to determine his career in China, he developed Google Translate Products at Google Research Institute. His Ph.D. The field of machine translation, one of the branches of artificial intelligence.

Li Zhifei pointed out that the sound is understood and identified technical problems are very different. For speech recognition, the biggest problem is noise, surround sound in different scenes and dialects spoken by different people, and the sound signal is ever-changing. The difficulty of semantic analysis is that the same meaning of the sentence, with a variety of terms and word order, "such as China Southern Airlines and China Southern, Shanghai and magic are."

Teacher from the domestic semantic analysis, Mr. He Zhongxiong Beijing Jiao Tong University worm CEO

ACRO® Parts for DEUTZ®

Acro Parts For Deutz,Deutz Engine Parts,Deutz Engine Tcd6.1/ 4.1,Coolant Pump 04507212

ACRO (TIANJIN) INTERNATIONAL TRADE CO., LTD , https://www.acrospareparts.com