个推科普漫画，解读《女心理师》中的智能语音识别系统

▲ photo source: Youku online drama “female psychologist”

In the psychological rescue center where the woman works, there is not only a cool visual data screen, but also a very advanced intelligent speech recognition system. This intelligent speech recognition system can not only translate the voice content of both sides of the call into text in real time, but also make an early warning according to the content of the call, and help psychological counselors to make judgments and decisions, so as to better implement assistance.

Having studied physics in high school, we all know that the essence of sound is wave, and frequency and amplitude are two important attributes that describe sound waves. * We can get the sonogram by making a visual chart of the amplitude and frequency of the sound over a certain period of time.*

What do you think of when you see the sonogram above? Is it very similar to the image of sine function in mathematics?

The first step of intelligent speech recognition is to select the correct type of function to describe different sound waves, and then give the data to the machine to learn and calculate.

Understand this, your right foot has stepped into the door of intelligent speech recognition science!

But sound waves are very complicated. Sound waves are the superposition of sine waves of different frequencies and intensities. When we visualize the sound, all we get is the sound map after superposition and synthesis. In order to better understand the sound signal, the algorithm engineer also needs to decompose the sound map.

This process of decomposing (transforming) sound signals is the “Fourier transform” that is often mentioned in machine learning.

It is a very commonly used algorithm in the field of machine learning, and its function is to analyze the digital signal in order to facilitate the subsequent data processing.

After the sound file is parsed into a sonogram, the machine can simply identify and judge the sound, such as the sex and age of the speaker.

We know that there is a very obvious difference between male and female voices. Generally speaking, the male voice is loud, the amplitude is higher and the frequency is lower; the female voice is sharp and thin, the frequency is higher and the amplitude is lower. At the same time, there are some abnormal conditions, such as shouting for help and moaning during illness, which can also be described by frequency and amplitude.

The intelligent speech recognition system in the female psychologist can also translate the voice content into text in real time and make an early warning according to the voice content.

How do computers understand human language?

Therefore, it is also necessary to teach the computer to “understand” human speech, and make it have a certain level of professional knowledge, and be able to carry out emotional analysis and independent reasoning on the text content, so as to achieve intelligent early warning and better assist research and judgment.

How can it be realized? Let’s take a look at step four.

We know that sentences are made up of words, including pause words (of, and, ground, de, between.), positive words (cheap, clean, beautiful, good and cheap.), negative words (dead, dirty, bad, bad.), degree words (OK, very good, make do, general, special.) Interrogative words (whether, not, unexpectedly, unexpectedly, after all, simply, simply, on the contrary, why, why.) and negative words (no, Mo, no, Fu, no, no.) and so on. In order to understand the emotion and attitude of a sentence, it is necessary to analyze the part of speech of each word in the sentence. * so we need to segment the translated text in the previous stage, and then synthesize the emotional tendency of each word to get the overall emotional attitude of the sentence. *

When the computer carries on the emotional analysis of the text, it still has to transform it into a mathematical problem before it can be solved.

Computers generally use the following types of mathematical expressions to calculate the emotional attitude of a sentence:

Generally speaking, whether the emotion of vocabulary is positive or negative, there are different standards and opinions in different fields. For example, “loud” is actually a very positive word in the audio industry, but in the home appliance industry, saying that a washing machine is “loud” actually expresses a negative attitude. * therefore, algorithm engineers in different industries need to build a characteristic thesaurus for their industry or field. *

At this point, if you look at the intelligent speech recognition system in the female psychologist, it is easy to understand its “cleverness”. In the field of psychological counseling or rescue, “jumping off a building” and “suicide” are negative words, and there is also a negative word “no” in the sentence “don’t come to me”. Therefore, when the system judges that the voice content is very negative and negative, it will automatically pop up the corresponding alarm.

▲ photo source: Youku online drama “female psychologist”

Seeing this step, congratulations on putting your left foot into the door of intelligent speech recognition science!

As you can see, in the alarm pop-up box in the above picture, there is also a reminder that “professional support is needed”, with a high level of intelligence. In fact, in real life scenarios, the intelligent customer service systems built by many enterprises in e-commerce, Internet medical and other industries have also evolved to a very high level of intelligence, and they can not only understand the content of the text, but also do reasoning and association on their own, put forward relevant professional suggestions, and assist decision-making.

Knowledge graph, which is essentially a semantic network that reveals the relationship between entities, is widely used in search engines, text mining and other fields. For example, when users use a search engine to search for “fruit”, there will be related terms such as “classification of fruit”, “nutritional value of fruit” and “where is the nearest fruit store”. Behind it is the knowledge graph in the field of “fruit”.

The “intelligence” of the intelligent customer service system we use in real life comes from the continuous learning of the knowledge graph of a specific industry. For example, the intelligent customer service of e-commerce platform relies on the knowledge graph of goods, orders, logistics and so on. When users inquire about a commodity, intelligent customer service will call the relevant map to provide users with product details, order status, logistics status, historical price trend, commodity use and other relevant information and suggestions.

After reading this article, I believe everyone has a deep understanding of the technical principles of intelligent speech recognition, knowledge graph and so on. Technology is not only productivity, but also temperature. In the correct application posture, big data and artificial intelligence can play a huge positive value in a variety of industries.