Gathering micro-messages, many AI algorithm companies attacked in the field of speech recognition, image processing, etc., and welcomed the capital's blitz. But now the wind direction has changed, and AI algorithm company finds that the algorithm cannot be sustained during the process of seeking applications. The 'hands' began to look for ways to transform. No, several domestic voice start-up companies have recently 'hardened' landings, launched AI voice-specific chips, including Sibin Chi, go out to ask, Rokid, Yun Zhisheng And so on. Abandoning their 'infighting and fighting' and talking about each other, the era of AI algorithm alone has apparently turned chapters, the battle of the next AI voice dedicated chip has already started.
The end of the algorithm
The choice of these companies 'invariably' may be due to the times.
Yang Yu-hsin, co-founder of Anchuang Space, said in a real statement that during the recent roadshow of the 5th Anchuang Growth Camp, he stated that it is difficult for the startup company to do the algorithm purely because the algorithm has bottlenecks and ceilings and may not be sold at all. Big money, so it must be combined with hardware to achieve commercial value.
Business value is closely related to the application of the scenario. 'In a subdivided scenario, the business value lies in the ability to achieve corresponding functions with high performance-to-price ratio and the best performance-power ratio of the hardware. These scenarios are large companies' chips when the volume has not yet risen. It is impossible to enter because large companies value the market of tens of millions of shipments. 'Yang Yuxin believes that 'in this 'vacuum' period, the company that performs the algorithm must be vertically integrated, ie guaranteed by the chip Its own core values and interests. Because of their AI capabilities, based on their own algorithms and capabilities, they can develop the best chip for themselves to achieve the best match. This is the inevitable trend of the AI algorithm company's commercial value to maximize.
Obviously this is a result of 'evolution'. The algorithm world must be based on computational power. The algorithm is not sufficient to support future growth. Vertical integration of AI algorithm companies becomes an inevitable path choice.
Dedicated chip into a new battlefield?
Therefore, the AI voice algorithm company's 'invariable' test was the first to test the water. Several domestic voice technology startup companies have successively launched the AI voice special chip as the corroboration. The above mentioned head start up companies in the AI voice field have almost always bet on AI voice. chip.
On the one hand, it is the move of the algorithm company to “catch the birds for change”, and it is also 'in tune' with the market's thrust. In the big market smart speakers for voice applications, sales volume has ushered in the period of blowout, Canalys predicts, global in 2018 Smart speaker shipments will exceed 56 million units. In addition, the automotive market, smart home and other scenes in the hardware began to voice intelligence, will further drive the AI dedicated voice chip outbreak.
At present, the chip solutions in smart speakers on the market are either general-purpose chips that take care of voice processing or have dedicated voice chips. For example, Lynx uses the MT8516 voice chip of MediaTek, and the millet speaker uses the Crystal A113 chip. Under this circumstance, a number of domestic voice technology companies rely on their own accumulation in technologies such as speech recognition, natural language processing, and voice interaction design to make AI voice chip integration and provide voice interaction solutions.
With the participation of a group of AI voice companies, it will make the market structure and future trends unclear.
Chip architecture choice?
Judging from the chip architecture, whether it is the first generation of UniOne chip "Swift" for IoT field released by the cloud sound release, or go out to ask the introduction of the voice chip module Mobvoi A1, or Rokid released SoC chip KAMINO18 , Mainly featured in high integration, low power consumption, low cost, and customizable.
It is reported that Yun Zhisheng's 'Swift' chip adopts a CPU+uDSP+DeepNet architecture, while going out to ask and Rokid announced that its chip is based on the depth customization of Hangzhou Guoxin Technology. Rokid has confirmed that KAMINO18 is based on the 40nm process of the Guoxin GX8010. manufacture.
At the end of last year, Guoxin Technology released two NPU chips GX8008 and GX8010, which are mainly used for AI voice interaction. It has Cadence's latest Tensilica HIFI 4 DSP core, which is low-power, low-cost, offline and integrated. Go out and ask, Manufacturers such as Rokid have achieved their IC by integrating microphone array signal processing, noise reduction, wake-up technology, voiceprint recognition and some voice skills.
SBCI announced the completion of the D round of 500 million yuan financing, led by Yuen Biao Holdings, China Minsheng Investment Group, Shenzhen Venture Capital, Foxconn, MediaTek, and Cast. Sibi Chi said that it is currently related to a chip giant in terms of AI chips. The two parties will establish a joint venture company. The AI chip for intelligent voice interaction will be in the process of R&D and will be streamed in the second half of the year. It is reported that this will be an ASIC chip with acoustic signal processing capability and voice capability.
The chip competition is not only simple and cost-effective. Providing an overall voice interaction solution is the focus of the core. Among them, Yun Zhisheng proposed a solution for the integration of cloud cores, and a one-stop solution is needed to go out and ask questions. Soft and hard voice solution. Sibi announced that it would provide more comprehensive support services including DUI platform, voice chip, and vertical solution, as well as deepening solutions and platform capabilities to provide smart services. Business Assistant will soon be announced. Smart service business.
Market wrestling?
And this is just the beginning of the battlefield for voice-specific chips. In the future, there will be even more giants of algorithm companies 'stirring in'. This war situation is flooding into so many players. It is also destined to be a tough battle.
Whether the AI-specific voice chip can continue to burst, on the one hand, depends on whether these chips can be used on a large scale. On the other hand, it depends on whether the voice interaction capabilities of these products can be favored by the market.
Wei Shaojun, director of the Institute of Microelectronics at Tsinghua University, said in an interview with the media that the killer application of AI has not yet appeared. No matter whether it is a smart speaker or other products, it has not yet become a just-needed one. Therefore, only making speech really become human-computer interaction The mainstream, in order to promote the AI voice chip outbreak.
In addition to integrated solutions, the deep integration with the scene is the real test of the future. Yang Yuxin said that if AI companies only make chips, there will be no problems with algorithms and scenarios, and they will encounter ecological problems immediately. Scenarios, algorithms, and chips. Now that there are algorithms and chips, the key question is how to go deep into the scene to create an excellent voice interaction experience?
Shi Bichi, Yun Zhisheng, go out and ask questions, Rokid et al. have many crossovers in the layout of application scenarios, all of which focus on smart homes, cars, robots and other fields. This should be the main battlefield for the future. Of course, they are also actively exploring. New business scenarios, such as medical care, smart services, etc. Future fronts will also be prolonged.