AI chip applications continue to evolve in multi-dimensional directions, such as CV (computer vision), autonomous driving, smartphones, and voice interaction. In the field of voice interaction, China's intelligent voice market is maintaining rapid growth. According to the Deloitte report, it is expected that the consumer-level application scenarios will exceed 70 billion yuan in 2030, and the development space of enterprise-level scenarios is also expected to reach the scale of 100 billion.
In order to make the intelligent terminal have a better interactive experience, to create a matching AI chip for the voice algorithm, the use of soft and hard integration solutions is almost the most common choice in the industry, which is also the inevitable path of technology iteration. Nowadays, the intelligent voice track has gathered many players, such as Baidu, iFLYTEK, Yunzhisheng, Sibichi, Mobvoi, Qi Ying Tailun and so on have laid out the chip industry.
To solve the problem of dedicated chip technology, enterprises still need to break down the commercialization barriers
Intelligent speech is inherently more complex in terms of technical correlation and data level, and the ability to understand and process natural language alone has cost scientists decades of hard work in exchange for today's interactive experience. The ability to integrate speech recognition, semantic understanding, natural language processing, speech synthesis, voice noise reduction and other technologies into a "cloud core" and extend the business to chips and even hardware is a reasonable path to commercialize the technology.
The general-purpose chip architecture is not specifically designed for AI, and the limitations of performance and power consumption are commonplace. In recent years, with the efforts of many enterprises in the industry, the adaptation problem of traditional general-purpose chips has been solved, and companies have also invested in the manufacture of special chips.
After the technical problems are solved, AI voice chips still face many challenges on the road to commercialization:
First, how to achieve performance optimization under cost constraints. Intelligent voice technology is tightly coupled, and the technology of patchwork cannot get the ideal interactive effect. What needs to be taken into account is the full-stack solution on the chip, and each additional function means an increase in cost. The low-cost, easy-to-implement, low-power product features need to be closely integrated with the solution.
Second, throughout the layout of AI voice chip companies, their selected application terminals are concentrated in home, electrical appliances, robots, vehicles and other scenarios. However, the large number of product categories in these types of scenes is a major feature, especially home appliances, from a large air conditioner to a socket have a demand for voice chips. How to adapt the chip on these devices, the necessity of judging each of the functions of the chip, requires a deep grasp of the terminal product function Know-How.
Third, due to the natural dispersion of customer manufacturers, standard product plus tool customization is the most efficient cooperation model. Having an efficient tool chain and reducing the time and marginal cost of customization will greatly enhance the commercialization process of voice chips.
Tens of millions of shipments, how do enterprises do it "core"?
As one of the first enterprises in the industry to design and provide software and hardware integration products for edge-side chips, Yunzhisheng has released a total of 6 full-stack voice AI chip software and hardware integrated products based on self-developed or third-party chips, and the shipment volume of chips and modules has reached the level of tens of millions, accumulating nearly 800 cooperative customers.
As early as 2015, Yunzhisheng began to build a chip team, not only from its industry sense of smell, but also from the environment. Yunzhisheng's chip solutions are built-in deep neural network acceleration schemes optimized for speech recognition to achieve off-line recognition of the device's speech. At the same time, the core link of voice interaction has also made a major breakthrough. The speech recognition link breaks through the single point capability, from far-field recognition, to speech analysis and semantic understanding, and has also been greatly improved, presenting an overall interaction scheme.
In 2018, Yunzhisheng's first-generation UniOne "Swift" chip was successfully tape-out for the first time and mass-produced. The "Swift" chip is a high-performance, low-cost and integrated chip solution launched by Yunzhisheng for smart home voice interaction scenarios. The "Swift" architecture built-in digital signal processor uDSP, as well as the AI accelerator DeepNet (Yunzhisheng's completely self-developed deep neural network processor, NPU), supports DNN/LSTM/CNN and other deep neural network models, which can achieve deep learning computing acceleration required for speech recognition, understanding, and synthesis. Compared with the general-purpose CPU, the processing speed and efficiency of this ASIC chip have been significantly improved.
The accumulation of industry Know-How
For the choice of business scenarios, Yunzhisheng's end-side intelligent voice chip accurately targets the small household appliance market. The opportunity to choose this track began in 2014 with gree, midea and other electrical appliance manufacturers. In this field, Yunzhisheng has accumulated more than 7 years of experience and is familiar with the characteristics of scenes and electrical products.
Taking microphone array technology as an example, there is no doubt that the more microphones there are, the easier it is to achieve better noise reduction and speech enhancement. However, compared with the multi-microphone array scheme that is in full swing in the industry, single microphone has become the largest shipment solution in the home appliance industry. This is because the difference in effect only has an impact on products with sound source positioning needs, and for the home appliance market, some equipment that originally needs to be placed against the wall, such as air conditioners, televisions, etc., the application of eight microphone arrays is obviously redundant. Therefore, considering the implementation cost, structural design and production installation, single and double microphone is the most suitable solution for this application scenario - these experiences come from the mastery of the industry function Know-How.
As a result, Yunzhisheng's second-generation chip "Hummingbird" for the small household appliance market was successful in 2019. The "Hummingbird" chip is a high-performance, highly integrated, cost-effective voice intelligent IoT chip specially designed by Yunzhisheng for offline far-field voice interaction scenarios. Compared with the general chip "Swift" in the voice industry, the positioning of the "Hummingbird" chip is lighter and more flexible, which can continuously reduce the threshold of adapted equipment and provide customers with more cost-effective solutions.
Hummingbird series chips apply far-field pickup, high-performance recognition, and low-power wake-up functions. Equipped with single and double microphones, it can achieve 10 meters of long-field pickup, and the recognition rate of 5 meters in quiet environments reaches more than 95%. Mainly facing the kitchen, living room, bedroom, bathroom, etc. in the home environment, including white electricity and small household appliances (lamps, kitchen appliances, smart sockets, etc.) product field. As a leading large-scale mass production voice solution in the field of white electricity and small household appliances, Hummingbird and related series of chips have shipped tens of millions. It covers domestic first-line home appliance manufacturers including Gree, Midea, Haier, Oaks, Vantage and so on.
At the same time, with the public's attention to user privacy issues, compared with cloud services that require the system to connect and upload data, the demand for offline voice interaction in specific scenarios is gradually increasing. Low power consumption, low cost, rapid response and the integration of off-line interaction mode, the system can intelligently decide the off-line processing mode has become a necessary condition for the development of voice AI chips. Therefore, Yunzhisheng's chip product matrix also lays out high-end chips for the automotive market.
Yunzhisheng and Geely Group's Yijiatong established a joint venture company, Xinzhi Technology, and launched a high-performance car-grade off-line voice chip "Leopard", which is expected to be launched this year. This chip also integrates the perception and cognitive technologies of Yunzhi sound. Facing the unstable characteristics of the in-vehicle network environment, in the pure offline scenario, "Snow Leopard" provides natural language interaction and nationwide addressing comparable to the online experience, as well as multi-microphone noise reduction functions. At the same time, Snow Leopard has obtained the vehicle specification grade AEC-Q100 certification, with completely independent intellectual property rights, providing pure local voice solutions for automobiles. This localized service not only allows users to experience smooth voice interaction functions, but also solves the problem of user privacy leakage.
As autonomous driving and smart cockpits also become hot industries, the construction of a vehicle-centric ecosystem based on the Internet of Vehicles is also a strategic direction to accelerate the development of intelligent voice enterprises. The full integration of Internet ecology, user personality, environmental interaction, etc. is undoubtedly the consideration of Yunzhisheng's choice to specialize in vehicle chips.
Empower efficient product landing with convenient tools
In different application scenarios of the Internet of Things, massive terminal devices must cooperate with the cloud to achieve functional intelligence, that is, to form a dynamic balance between edge computing power and cloud computing power. The proposition of cloud interaction requires strong support from AI chips, which further profoundly affects the design and ultimate delivery of chips.
As google's Switch Transformer model launched in 2021 refreshes public perception with the learning ability of small samples or even zero samples, pre-training large models also herald the outbreak of new trends in AI technology. People can more smoothly inject the knowledge they need to learn into the model, which means that complex and large number of custom tasks can be adapted and reasoned for calculations. Similarly, Yunzhisheng also uses pre-training models to efficiently solve the pain points of IoT product customization, and establishes a developer platform for AIoT product customization.
Voice control, docking IoT control and device control, through the platform self-service generation can achieve the effect of the algorithm original factory offline support and R & D personnel coding to achieve the effect, the original work cycle that took several weeks to 30 minutes. For example, the offline standard scheme can be configured with the number and spacing of the microphone, custom wake words, timbre configuration, command words and answers, and burned with one-click download versions. Customers only need to focus on the electronic control part they are familiar with, and can customize the exclusive intelligent voice solution without additional hardware capabilities.
The standardized delivery model greatly reduces the threshold of the client, and through the one-stop development of the product side and the cloud, intelligent voice control is quickly realized, so that zero-based enterprises can easily achieve intelligence. Coupled with the characteristics of strong operation, the platform supports hundreds of configurable cloud skills, and can also customize skills or access third-party skills. As of now, active customers have built more than 25,000 product versions on the platform.
epilogue
In the past decade, AI technology has continuously made breakthrough competition, and the AI industry will also usher in its second half. In the face of the current booming intelligent voice market and the addition of Internet giants, although enterprises with similar technical levels, the trade-offs in vertical industries and subdivision scenarios will also cause completely different orientations, and the landing and scenarios of AI chips also need to be closely combined.
After solving the problems of performance optimization under functional constraints, free configuration in different scenarios, and convenient toolchains, intelligent voice technology enterprises represented by Yunzhisheng need to continue to accelerate the penetration and layout of vertical industries, so that intelligent voice can empower multi-form terminals and build an industry-wide ecosystem with the power of "China core".
Leifeng Network