Shulou(Shulou.com)11/24 Report--

Highlight:

The deficiency of in-car voice intelligence lies in "semantic understanding". The hot ChatGPT in AI circle has obvious addition to car-borne voice intelligence.

ChatGPT boarding is mainly a matter of cost, which includes user costs, cloud service costs, and targeted training costs.

CTO_ Liang Jiaen, chairman of Yunzhisheng, told TechWeb that ChatGPT technology will certainly make achievements in intelligent interactive applications such as vehicles and homes, but it needs to be optimized in combination with application scenarios.

GE Fujiang, product director of Spice Automotive Division, told TechWeb that there must be commercial landing challenges in the development of new technologies, and AI technology innovation should be combined with scenario applications, and the application of similar ChatGPT in vehicles will pose challenges in computing power optimization, cloud and end intelligent fusion technology.

The fire of ChatGPT suddenly spread to the field of cars.

As we all know, voice interaction is the most concise, humanized and safest way of interaction in the car, and it is also the most important way of interaction in the car in the future. With the enhancement of AI and hardware performance, voice interaction is the absolute mainstream of cars in the future. Voice interaction is mainly vehicle natural speech recognition and speech assistant, it can also be simply said that NLP and NLU technology. Since it is NLP, it should be an opportunity for ChatGPT, who has recently become popular in the AI circle, to show his talents. Is this really the case?

Vehicle voice intelligence, short board in "intelligence" from a technical point of view, intelligent voice interaction mainly has three key points, namely, recognition, understanding and execution. At present, among the manufacturers who provide solutions, the recognition part has become mature, and the recognition rate can reach more than 90%, and some of them have reached about 95%. The pain points of the industry mainly focus on the "understanding" part, most of the vehicle voice interaction systems are not intelligent in "understanding", resulting in a single function of the whole system and a single command word.

So the question is, how to make the in-car voice interaction system understand our words like people?

This involves NLP (Natural language processing) technology, their understanding of user input voice has a close relationship with their own scene strategy and multi-round dialogue, and directly determines the intelligent degree of the vehicle voice interaction system. The mention of NLP is exactly the wish of the recent "hot" ChatGPT, and it is an opportunity for ChatGPT to show his talents.

Historically, there have been several key points in the development of NLP, of which the two most important are 2012 and 2018.

In 2012, deep learning began to be applied to the NLP field; since 2018, the semantic representation pre-training represented by Google BERT has made a great breakthrough, sweeping the major NLP task benchmarks; in May 2020, OpenAI spent a lot of money to build GPT-3, which caused an industry sensation as soon as it was published. This version of the model has 1750 billion parameters and is known as the strongest AI model in the NLP field.

Recently, the popular ChatGPT is based on the large-scale pre-training language model (GPT-3.5). With its strong ability of language understanding and generation, the pre-training language model can better understand human problems and give better responses by learning from large-scale data tagged and fed back manually.

GE Fujiang, product director of Spice Automotive Division, told TechWeb that ChatGPT is currently presented in the form of a text interactive robot, which is suitable for a variety of text processing tasks, and is often used in intelligent question answering, dialogue, text creation and other fields. Turn on the music "has a clear command of the action," the "voice assistant" with a highly anthropomorphic voice output to respond to the demands of car owners. In-vehicle voice interaction is used to liberate drivers' hands and focus their attention to bring a safer and more convenient driving experience. With the application of ChatGPT technology in cars in the future, not only task-based conversations with fixed instructions can be completed, but also cars and people can communicate more efficiently, directly and flexibly in travel, knowledge and chat.

The popularity of "ChatGPT" makes the market see the potential of cognitive intelligence applications. ChatGPT has obvious advantages in reasoning and learning ability. It can not only be used for understanding and dialogue, but also through contextual communication and self-learning to achieve auxiliary creation and knowledge evolution. These capabilities are also applicable to the field of vehicle voice interaction, integrating dialogue intelligence technology, deep learning large model technology, engineering capabilities, and big data's potential to bring smoother and more effective responses. In the limited space in the car, combined with sound field location and multi-speaker judgment, the logical consistency of multi-role and long-context dialogue can be improved, and it can be expanded to meet the needs of unified identification and dialogue of dialects and foreign languages. quickly achieve more flexible, free and personalized interaction. "said GE Fujiang.

From the current use of ChatGPT (including our own), we believe that only for the car intelligent voice NLP, it should be the best and the most intelligent. Does this mean that it will certainly be used in the vehicle intelligent voice system in the short term?

It is well known that the market space is limited, the industrial chain and market challenges still exist. Whether a new technology or product can finally be applied on a large scale, in addition to technical factors, it will also be closely related to many factors, such as the industrial chain of the industry or market, market competition, market space and so on.

Specific to the vehicle intelligent voice system, although ChatGPT performs well in "intelligence", it is relatively late in the whole industrial chain and needs to rely on a long front-end chain, such as signal processing, speech recognition and text output. the factors on the front-end chain will affect the back-end process, for example, signal processing will affect speech recognition, and speech recognition will affect the judgment of NLP if something goes wrong. Each module in the chain needs to improve its reliability in order to ensure the reliability of the overall result. This means that the output of ChatGPT's "intelligent" capability does not depend entirely on its own ability, and any link in its industrial chain will have a positive or negative impact on it.

CTO_ Liang Jiaen, chairman of Yunzhisheng, told TechWeb that ChatGPT technology will certainly make achievements in intelligent interactive applications such as vehicles and homes, but it needs to be targeted and optimized in combination with application scenarios to improve experience and reduce service costs.

"there is a lot of room for experience upgrading in intelligent interactive application scenarios such as vehicles, but at present, ChatGPT is a super-large model, and how to significantly reduce service costs under the condition of maintaining the experience is a key issue."

From the perspective of market competition, according to relevant statistics, the current vehicle voice system market has been monopolized by HKUST Xunfei and Cerence, and they have many years of product and cooperation experience in this field, there are many enterprises of different sizes to participate in it, more importantly, the vehicle voice market has encountered a growing ceiling, which makes the fierce competition at the same time Even HKUST Xunfei and Cerence have begun to take the route of in-car multimodal interaction, cloud service integration and other services in addition to voice, so as to enhance their competitiveness with comprehensive strength. As a latecomer, ChatGPT is bound to face the challenge of strong competitors once it decides to enter the car intelligent voice market.

GE Fujiang added that in terms of cost, ChatGPT research requires huge capital and talent investment, and they need the support of core forces such as supercomputing platforms, algorithms and data, all of which are costs. At present, giant platform companies have the advantages in this respect, and for technology enterprises, they can start with scene integration and seek opportunities for innovation.

From the perspective of commercial scenarios, chatGPT is more suitable for creative industries based on certain background knowledge, as well as scenarios with rigid AIGC requirements and SOP (standard operating procedures) industries, such as intelligent writing, intelligent customer service, document management, code generation, and even game NPC.

Sun Yongjie, the master of tricks, pointed out that from the perspective of the simple car voice market, its market space is not large, which can be seen in the financial reports of the University of Science and Technology and Cerence, which has monopolized the market. In this case, whether it can attract costly ChatGPT entry is also unknown. After all, ChatGPT training is expensive, and the Open AI it belongs to is still losing money.

There are still prospects in the future, cooperation and opening up API or better options as mentioned above, ChatGPT only has advantages in the NLP segment of in-car intelligent voice, although ChatGPT is also said to be conducting AI training in speech recognition and synthesis, hoping to enter the in-car intelligent voice market in the future. But given that ChatGPT is only a way of text interaction, what is the final effect of AI training in speech recognition and synthesis? Whether it can surpass the existing and applied vehicle intelligent voice system on the market is still unknown.

Of course, given the strong capabilities of ChatGPT, TechWeb believes that with the expansion of smart car application scenarios in the future, it is not impossible for ChatGPT to find a real opportunity to use its talents. What is more worth looking forward to is that, in addition to the intelligent car itself, standing at the height of the entire automobile industry, its future applications in automotive design, manufacturing and other fields are full of imagination.

GE Fujiang said: "it is not clear how the application of ChatGPT boarding will develop." Predictably, in the vehicle scene, the large model technology learning ability has obvious advantages. By strengthening the ability of context understanding, thinking chain reasoning and enhanced instruction learning, continuous learning can be achieved to achieve the effect of "answering similar questions". In addition to instruction requirements, daily knowledge and small talk can be more fluent and useful. Generally speaking, the technology will develop towards unified multimodal interaction, strengthening the deep fusion of voice, text, image and other multimodal interaction technology, forming a "car brain" to cope with the interactive needs of complex scenes such as in-car and public space. "

Xiaopeng's technical team told TechWeb that ChatGPT has a strong language organization ability, a large knowledge base and a wide range of areas, so it may give users a better and more intelligent experience. As for whether we should introduce this technical interface or do the integrated development of similar technologies in the future, we are also exploring further.

Based on this, TechWeb believes that cooperation should be the most economical and effective way for ChatGPT to enter the vehicle intelligent voice market. This is the truth of the so-called learning from each other's strengths. The actual situation is that recently, domestic Jidu Automobile announced that it will integrate the comprehensive capabilities of Baidu Wenxin to create the world's first large model artificial intelligence interactive experience for intelligent car scenes, proving the feasibility of this model.

In addition, it is wise to open up your best abilities to third parties through API, and only output your best abilities.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.