On June 27, iFLYTEK released the iFLYTEK Spark V4.0 at the China National Convention Center in Beijing. After this upgrade, iFLYTEK Xinghuo fully benchmarked ChatGPT 4 Turbo, ranked first in 8 international mainstream test sets, and took the lead in domestic large models.
The Xinghuo model helps the "Chinese-style" intelligent cockpit to the world
Since the implementation of the national strategy for new energy vehicles, automotive intelligence has always been the dominant project of Chinese automobiles, among which the intelligent voice technology represented by iFLYTEK is the most significant. Since 2011, when iFLYTEK took the lead in completing the localization of car voice, in-car voice has become the standard configuration in the Chinese automobile market. However, in the vast overseas market, due to the constraints of the single language market size, the gap in the penetration rate of intelligent voice services is still very large.
Facing the era of the Internet of Everything, iFLYTEK has made another breakthrough in the Spark voice model, released 72 languages/dialects without switching dialogues, solved the problem of speech recognition in strong interference scenarios, released the world's leading voice transcription in extremely complex scenarios, and opened up a broader world for the automotive intelligent cockpit through cloud-edge-end and software-hardware integration solutions.
At the press conference, iFLYTEK used the new Chery Star Epoch ET to demonstrate the switch-free interaction of voice assistants including Northeast dialect, Tianjin dialect, English, and Russian. The Chinese automobile navy represented by Chery has achieved outstanding results in many countries with its leading new energy and intelligent technology. Last year, China's overseas sales exceeded 5.2 million units, ranking first in the world for the first time. Now, relying on the iFLYTEK Xinghuo model, iFLYTEK will have the ability to deliver China's local automotive intelligent experience to more car owners in more countries around the world. Helping Chinese auto brands including Chery, SAIC, GAC, Great Wall, Changan, BYD, etc., to further achieve brand improvement.
The Spark cockpit has been upgraded to create new scenes in multiple modalities
With the release of ChatGPT-4o, multimodal capabilities have become a hot concept for large models. The multimodal model combines the processing capabilities of multiple data types such as text, images, and sounds.
Compared with traditional voice input and small talk, multimodal capabilities greatly expand the application boundaries of large models. The large model that "can hear and see", the cognitive ability has risen from "anthropomorphic" to "human-like", and it has been added to the intelligent cockpit scene, which is like adding an invisible "accompanying all-round assistant" to the car.
For example, with the accelerated popularization of new energy intelligent vehicles, more and more car owners feel that the knowledge they have learned in driving schools in the past is "not enough"; At one time, there were controversial remarks from executives of new energy vehicle companies that "we should cooperate with driving schools to let everyone understand how to use new cars". In response to this problem, iFLYTEK has developed a car assistant based on the Xinghuo model: different from the traditional operation manual, the car assistant can monitor and understand the vehicle condition in real time, and accurately answer the user's questions about the car according to the current road conditions. For example, in different road conditions, it helps users turn on autonomous driving assistance; In different weathers, guide users to use the lights correctly; According to the real-time vehicle condition, accurate maintenance suggestions are given.
The multimodal capability also opens up a whole new application space for the hardware on the vehicle. The traditional DMS is mainly used to monitor the driver's fatigue status, and with the empowerment of large models, visual judgment can obtain a number of physical health indicators including heartbeat, respiration, and blood pressure with high precision, and record and track them for a long time in a state where the user is unconscious. On this basis, iFLYTEK Xinghuo health experts link up with iFLYTEK's medical resources to provide monitoring and diagnosis of more than 30 kinds of health problems for car owners and protect their safe travel.
With the help of hard-core technology, the integration of core computing makes domain control more powerful
Behind the continuous creation of new experiences in large models and multiple scenarios is the continuous upgrading of the computing power demand for automotive intelligent cockpits: in order to support more and more intelligent function applications and support the trend of multimodal integration, the AI algorithms in the car are becoming more and more abundant, and the traditional cockpit SOC will face the dilemma of insufficient CPU computing power. Compared with the frequency of hardware changes in the PC/mobile phone industry, the life cycle of automotive products to accompany users is much longer, and it is more necessary for us to maximize the value of SOC performance for users on mature platforms through algorithm research.
As a leading provider of intelligent cockpit solutions in China, iFLYTEK gives full play to its own technical advantages, deeply integrates and deploys intelligent vehicle algorithms and intelligent vehicle chips, supports multi-modal fusion and interactive applications, and realizes efficient inference, efficient transplantation, and efficient debugging.
Taking iFLYTEK's most representative voice algorithm as an example, after transplanting voice noise reduction, wake-up, recognition, and synthesis from the CPU to the NPU through heterogeneity, the CPU computing power requirement can be reduced by 60%. By deploying a larger model on a resource-rich NPU, it is possible to achieve cloud-like speech recognition locally. This is the technical secret behind the fast and accurate voice recognition on a number of new models such as Hongqi EH7 and NIO ES8.
In the more popular field of large models, through the device-cloud collaborative deployment solution, while using large models in the cloud to achieve multi-round penetration, streaming dialogue, full-domain planning, controllable dialogue, and in-depth understanding of multi-domain knowledge, the device-side model with billion-level parameters is deployed locally to enhance local semantic understanding capabilities, which can not only significantly optimize the response time, but also effectively ensure the closed-loop experience of extremely fast, offline, multi-mode, and privacy and security services. Under the device-cloud collaborative deployment solution, the Spark large model can improve the semantic enhancement of the local large model by 40% and reduce the response speed by 500ms with a cloud intent classification accuracy of more than 98%. Chery Xingtu Star Era, GAC Aion, and Great Wall Wei have all chosen the Xinghuo end-cloud deployment solution to improve the performance of on-board large models.
The integration of core computing is also of great significance to the domestic substitution of automotive chips. In the current macroeconomic context of trade frictions, Chinese car companies are facing risks in the supply chain of foreign-funded chips, and China's domestic on-board chips are ushering in a key opportunity for market substitution. Through the deep integration of algorithms and chips, the terminal performance of the system can be greatly optimized, and the gap between the current independent products and the international leading competitors in terms of design performance and process technology can be bridged. Boost confidence in the choice of domestic alternatives for Chinese car companies, and bring more cost-effective end models to consumers!
In iFLYTEK's view, the development and application of artificial intelligence is not only a technical competition for cooking oil with fire, but also an application of long slopes and thick snow. iFLYTEK has always adhered to the corporate values of achieving customers, using the latest research results to cover the needs of real scenarios, always standing with Chinese automobile companies, and providing a better travel experience for China and the world with the vision of inclusive technology.
Source: Zhongguancun Online