Automotive Voice Industry Report, 2025

Automotive voice research: high-level voice function installation rate significantly increases, automotive voice moves towards 'cognitive interaction'

1. Automotive voice installation rate exceeds 83%, high-level voice function installation rate significantly increases

From January to November 2024, installations of automotive voice systems reached 16.76 million units, with an installation rate of 83.3%. Compared to the full year of 2023, installations increased by 5 percentage points. By energy type, EREV (Extended-Range Electric Vehicle) had the highest installation rate for automotive voice systems, reaching 100% from January to November 2024. Typical models under this energy type include the Li Auto L series, AITO M series, and Deepal S series.

In terms of voice function, installations and installation rate for continuous dialogue, see-and-speak, and wake-up-free functions greatly increased in 2024.

For the see-and-speak function, from January to November 2024, its installations reached 4.66 million units, with an installation rate of 23%, an increase of 18 percentage points compared to the full year of 2023. By energy type, EREV had the highest installation rate at 92.1%, while fuel vehicles had the lowest at only 7.1%. By price range, the 'see-and-speak' function had the highest installation rate in the over 500,000 RMB range, with representative models such as Zeekr 009, Yangwang U8, and NIO ES8. This range also saw the largest increase in installation rate, up by 48 percentage points. This also indicates a significant improvement in the intelligence level of automotive voice systems in 2024.

2. The cockpit accesses more ecological resources, voice assistants gain deep service capabilities

In the era of foundation models, a voice assistant that 'knows a lot and can serve' relies more on the access to diverse ecological applications. For example, when users issue vague commands such as 'the car is almost out of power,' 'I'm hungry,' or 'what should I wear for the Chinese New Year,' the voice assistant's response requires support from applications like maps, local life services, and online information.

In addition to common applications like AMAP, iQiyi, Tencent Video, NetEase Cloud Music, and QQ Music, Li Auto has implemented voice calls to Xiaohongshu (Little Red Book) platform content and launched a deeply customized voice skill for Meituan. For example, users can wake up Lixiang Tongxue to ask ' Chinese New Year outfits recommended by Xiaohongshu,' 'find a Beijing travel guide on Xiaohongshu,' or 'help me find a Cantonese restaurant on Meituan with an average price of 200 RMB and a rating above 4.5.'

3. Foundation model applications accelerate the development of automotive voice from 'command interaction' to 'cognitive interaction'

Different from the previous command-based interaction, automotive voice systems empowered by foundation models have better capabilities in spoken language understanding, logical reasoning, knowledge Q&A, painting creation, and perceiving the vehicle's surrounding environment. For example:

XPeng's XGPT-powered Xiao P assistant has capabilities in spoken language understanding, logical reasoning, knowledge encyclopedia, painting & story & fairy tale creation, and recognizing objects around the vehicle.

Li Auto's Mind GPT-powered Lixiang Tongxue has fuzzy search capabilities, such as asking Lixiang Tongxue 'I forgot the name of a movie, there's a black pianist, do you know what it is?'; search by image description, where Lixiang Tongxue can read movie poster content and express it freely, allowing children who cannot read to choose movies by describing the poster.

Xiaoai Tongxue's application of foundation models also gives it the ability to understand and respond to vague commands. For example, it can recognize and respond to commands like 'Where is my phone?', 'Turn off the lights at home', 'What mountain is that ahead?', and 'What car is that ahead?'.

Taking XPeng Motors as an example, XPeng Motors has developed its own XGPT (Lingxi) foundation model and integrated it into the voice system. Additionally, it has integrated the ZhiPu AI base foundation model and multimodal models, giving the voice assistant Xiao P stronger language understanding, image recognition, and generation capabilities, which can be linked with in-vehicle perception system and external environment.

4. AI foundation models become a must-have for OEMs to build intelligent automotive voice systems

By 2024, the number of brands equipping their intelligent cockpits with foundation models has significantly increased, with Chinese independent brands being the primary drivers of this trend. Some brands have already completed the development path from cooperative supply to joint R&D, and finally to independent research. For example, in January 2024, Geely applied Baidu's ERNIE Bot foundation model in its Galaxy L6. In the same month, Geely released its self-developed full-scenario AI foundation model - Geely Xingrui AI Foundation Model.

Based on the Xingrui AI Foundation Model architecture, Geely has also developed derivative models such as the Xingrui NLP Language Foundation Model and the Xingrui Multimodal Foundation Model. Among these, the Xingrui NLP Language Foundation Model is entirely self-developed by the Xingrui Intelligent Computing Center, with a total training data volume exceeding 3 trillion tokens. It includes an emotional module, enabling excellent logical reasoning and contextual memory capabilities, allowing for human-like emotional interactions.

In January 2025, Geely showcased its development path for an in-cabin intelligent assistant based on the Xingrui AI Foundation Model at CES 2025 - moving from 'Assisted Intelligence' to 'Agent Intelligence' and finally to 'Autonomous Intelligence.' With the support of the foundation model, in-car assistant will evolve from 'accurately responding to commands' to 'understanding the environment and autonomously completing tasks,' and ultimately to 'possessing self-awareness and autonomous emotional capabilities.'

Chinese independent brands such as BYD, SAIC, Dongfeng, GAC, Changan, Chery, and emerging OEMs like NIO, Li Auto, XPeng, AITO, and Xiaomi have also implemented AI foundation models in automotive voice systems. As automotive intelligence enters its second phase, AI foundation models are gradually becoming a necessary option for building intelligent voice interaction systems.