With the advent of the Internet of Things era, voice interaction has become an important part of "communication" between humans and machines. The voice-oriented layout has become a battleground for many technology companies.
However, the voice circuit also has its distinctive industry characteristics: first, the interaction process and logic of voice are more complex than that of image, which determines the complexity and difficulty of the commercial landing of its product application, and the user's expectation of product performance is higher. Secondly, from the perspective of the domestic market, compared with the voice track, the image has a more deterministic scene application. As a voice company, to survive, it also means that the team's "market hunting" ability is more tested.
At the same time, due to the fragmentation of the voice scene, industry players not only need to layout in advance in terms of technology, repeated polishing, but also need to have a sensitive business sense, identify and layout market opportunities in advance.
The more "complex" an industry, the more can test the concentration and confidence of enterprises in the industry. In the voice track, there is such a company focusing on voice AI technology -- Yunzhisheng. After eight years of grinding, it has explored its own survival rules and market sense on the voice track.
Over the past eight years, Yunzhisheng has gradually grown from an unknown enterprise to a leading enterprise in the industry, taking the lead not only in technology but also in commercialization. And in the recent days, Cloud know Voice officially submitted IPO prospectus materials to the Shanghai Stock Exchange, the company is "Science and technology innovation board AI voice first share" launched an impact.
Dare to be the first
How can you tell if an AI company is on the right track? Yunzhisheng provides a good example of how a company's path is judged mainly by the few steps it takes at key moments, rather than by how much it does the same thing as everyone else most of the time.
"Computing power, algorithm, data" is the basic element of building the core technology of artificial intelligence. One of the keys to lead the AI industry out of the deep water and realize the real industrial landing is to improve the supply of the algorithm side. In this direction, Yunzhichun took the lead in applying deep neural network (DNN) to commercial speech recognition system in 2012, and continued to carry out business practice of cutting-edge algorithm in the subsequent wave of artificial intelligence.
Based on the company's pioneering introduction of DNN into the field of speech, it continues to carry out the business practice of cutting-edge algorithm in the subsequent wave of artificial intelligence. For example, Later emerging mainstream technologies such as convolutional neural network (CNN), regression neural network (RNN), end-to-end sequence modeling, generative adversative network (GAN), attention transformation model (Transformer), bidirectional coding representation transformation model (BERT), knowledge distillation (KD), self-supervised learning (SSL), etc. Yunzhisheng is one of the earliest industrial practitioners in the industry.
From the choice of the vast majority of enterprises, DNN has become the mainstream choice. In other words, Yunzhisheng's advanced technical prediction turned out to be correct.
In addition to taking the lead in embracing DNN, another key step for Yunzhisheng is to start its own core-making in 2015.
It is understood that in order to further reduce the power consumption of intelligent voice interaction scheme, improve the sensitivity of its wake, and then meet the needs of more Internet of Things equipment loading voice interaction function, cloud know voice since 2015, forward-looking layout of artificial intelligence voice chip, set about the development of uDSP processor and DeepNet IP technology.
On this basis, Yunzhishen took the lead in the delivery of artificial intelligence voice chip -- "Swift" in 2018. Later, in 2019, it successively launched the vehicle scale chip "Snow Leopard" and the second upgraded version of the chip "Hummingbird" series for the home field, and started the research and development of the chip "Dolphin" with the multi-mode interaction function of "image + voice", continuing to lead the industry.
It should be noted that at present, more and more AI voice companies are simultaneously breaking into the new racetrack of core making. It has become a necessary way to "apply both soft and hard" from the algorithm, but it also appears that the voice AI market has become extremely competitive. In contrast, does the chip launched by Cloud know Voice have the quality to "stand out" in this competition?
Superior chip performance
Take "Hummingbird", one of the main chip products of Yunzhisheng, as an example. According to the research results of Jishi Consulting, "Hummingbird" has obvious advantages in speech processing efficiency and commercial cost compared with the main products of other market participants in the same period.
Specifically, the "Hummingbird" chip adopts advanced DSP+ neural network +MCU heterogeneous design. Compared with the traditional general CPU architecture, the heterogeneous design can greatly improve the operating efficiency of the chip on the one hand, and reduce the chip area and power consumption on the other hand. Its floating-point computing capability and fixed-point computing capability can reach 16MAC/ clock cycle and 32MAC/ clock cycle respectively, which are ahead of the performance of similar chips in the same area.
In addition, "Hummingbird" chip is integrated with 1.5MBSRAM and equipped with Yunji Voice's own speech recognition algorithm, which can realize far-field speech recognition at a distance of 5m with an accuracy of >98%. Compared with the same competitors, RAW capacity is reduced by nearly half, which is crucial in the cost-sensitive and competitive voice chip market.
At the same time, in order to maximize the integration of the chip, "Hummingbird" chip integrated FLASH inside the chip, to achieve a real sense of single chip voice solution.
In the first half of this year, thanks to the large-scale promotion of "Hummingbird" and the giving up of products with low gross profit rate, the direct hardware procurement cost of intelligent voice interaction products of Yunzhisheng dropped to 26.17%, and the gross profit rate recovered significantly.
In addition, in combination with the artificial intelligence accelerator IP with its own property rights, Yunzhisong also designed and developed the vehicle gauge level voice AI chip -- "Snow Leopard". This chip is vertically integrated with Yunzhisong's artificial intelligence software engine and artificial intelligence solution of the Internet of vehicles, forming a complete cloud core integrated solution.
At present, the vehicle gauge level chip "Snow Leopard" has entered the product stability test and vehicle gauge level test stage of Geely Auto, after passing, it can be launched on the mainstream platform of Geely Auto.
Through the core development idea of "only fast not broken", cloud know voice soon in the voice recognition tuyere comparable to the industry leader.
In the IPO of the Science and Technology innovation Board, Yunzhisheng plans to invest 317 million yuan for the project of "Construction of artificial intelligence chip R&D platform for the edge computing of the Internet of Things". The project aims to enhance the edge computing capability of iot terminals through AI chip research and development, and improve the chip design level in the application field.
Unique integration of cloud core
Dr. Liang Jia-en, chairman /CTO of Yunzhichao, has pointed out that Yunzhichao has built the Atlas machine learning computing platform known as "TensorFlow GKE (Google Kubernetes Engine)" of Yunzhichao very early. Based on this platform, the horizontal expansion and vertical iteration of signal (AEC/SSP/ISP), speech (ASR/TTS/VPR/CALL), language and cognition (NLU/NMT/SDS/ICI), image (FID/OD/OCR) and other technical fields are supported upward.
By encapsulating the capability on the self-developed AI chip, Yunzhichun took the lead in polishing the unique "cloud core" product middle platform, and conducted modular carding and integration of the core technology, so as to have the ability of rapid productization to meet the market demand, and support the company to provide services for many well-known enterprises with a rich product portfolio.
It is understood that the two core scenes of "cloud-end-core" products are "AI life (home, vehicle, robot, etc.)" and "AI service (medical, education, government affairs, hotel, etc.)" in the middle platform of Yunzhisheng, and then through the ecological loop of Yunzhisheng from AI technology to industrial application.
In the field of home appliances, the company's intelligent voice interaction program has been adopted by Gree and other leading home appliances; In the hotel, commercial real estate and residential areas, Yunzhisheng has reached a strategic cooperation with Shimao Group and established a joint venture company to deliver intelligent iot solutions in bulk to many hotels under Shimao Group, which has brought rapid growth of sales revenue.
In the medical field, through cooperation with Peking Union Medical College Hospital, Yunzhisheng has taken the lead in providing medical record translation solutions in China, and has been rapidly promoted in the industry. At present, it has been put into use in more than 100 hospitals and tested in more than 500 hospitals. In addition, the company established a joint venture with Ping An Good Doctor to test the development of intelligent hardware products in the medical field.
In the automotive field, Yunzhichun has also established a joint venture with Egatone to provide in-vehicle voice solutions for Daimler Benz and other auto manufacturers, and jointly explore the application of in-vehicle voice solutions based on self-developed vehicle gauge artificial intelligence voice chip "Snow Leopard".
Relying on the "cloud-end-core" product center, Yunzhisheng has formed three series, dozens of products complement each other, highly integrated product matrix. In addition, benefiting from the unique product middle platform system, Yunzhisong's technical capability to the product landing period has been significantly shortened, and the frontier algorithm with effective verification can usually be applied to the terminal products within 3-6 months, which ensures the development efficiency and competitiveness of the company's products and enhances the market competitiveness of Yunzhisong.
conclusion
At present, artificial intelligence companies are ushering in a golden age of enterprise development. On the one hand, since 2015, artificial intelligence has become a hot market, and related industries have been formally put on the national strategy. For example, the draft of the 14th Five-Year Plan points out that "we should target the frontier fields of artificial intelligence, quantum information, integrated circuits, life health, brain science, etc. We will implement a number of forward-looking and strategic major national science and technology projects.
According to the central plan, the future AI core industry and the integration of AI and traditional industries are strategic development priorities. Ping An Securities estimates that by 2020, the market scale of China's AI core industry will exceed 160 billion yuan, and the market scale of related industries will exceed one trillion yuan.
On the other hand, the more flexible listing mechanism adopted by the Science and Technology Innovation Board will help a number of AI enterprises to accelerate their entry into the capital market. At present, the review status of the Science Innovation Board of Yunzhisheng is changed to "asked", and the IPO process is progressing steadily. The company is expected to become the first share of AI voice on the Science innovation Board.
