Voice Recognition / Voice Interaction

Definition and Technical Concept

In-vehicle voice recognition and voice interaction refer to a technology suite that enables drivers and passengers to interact with vehicle systems via natural speech, driven by voice signal processing, speech recognition, natural language understanding (NLU), and speech synthesis. Its primary objective is to allow users to set navigation, operate vehicle controls, and perform information queries completely hands-free and eyes-on-the-road, thereby enhancing driving safety and operational convenience. Voice interaction is widely regarded as the most central and intuitive contactless interaction medium within the smart cockpit.

Technological Evolution: From Command Recognition to AI Large Models

In-vehicle voice technology has undergone a leapfrog evolution, transitioning from rigid "fixed commands" to fluid "natural dialogue". Early in-car voice systems could only recognise preset phrases such as "open navigation" or "increase temperature", requiring users to memorise exact command syntax—resulting in clunky interactions with little room for error. With the advent of deep learning and Large Language Models (LLMs), voice interaction has matured into an intelligent agent system capable of multi-turn dialogue, contextual understanding, and nuanced intent inference.

Current technological advancements in voice interaction are evident across multiple dimensions. On-device large model deployment is a primary focus—SAIC-GM has successfully deployed a 1.9B-parameter on-device model on the 8775 chip, achieving mass production of open-domain end-to-end voice interaction where users can access large model capabilities without requiring a separate wake-up phrase. Multi-intent understanding enables in-vehicle assistants to process complex multi-layered requests, such as "wind down the windows, then find me a top-rated café with free Wi-Fi and parking". Zoned interaction technology leverages multi-sound-zone acoustic source localisation and voiceprint recognition to determine whether a command originates from the driver or the front passenger, delivering tailored, per-seat responses.

Core Features and Industry Applications

Today's voice interaction features comprehensively cover all-scenario vehicle use cases. At the vehicle control level, the system can interpret fuzzy descriptions such as "it feels a bit stuffy" to automatically activate ventilation or air purification modes. For navigation, it enables dynamic route recalculation based on live traffic conditions. In terms of cross-domain coordination, a single spoken sentence can simultaneously execute multiple cross-functional tasks, such as "wind up the windows, play music, search for a destination, and set the ambient lighting".

In practical industry applications, the AI voice assistant "Xiao Ni" featured in NISSAN OS 2.0 supports seamless multi-dialect recognition without manual toggling—including Mandarin, Cantonese, and Sichuanese—and identifies occupants via the OMS sensing system to automatically adopt a child-friendly tone for younger passengers. Desay SV's fourth-generation smart cockpit platform incorporates a 1.2B voice large model, facilitating seamless 3D HMI interaction and emotion perception. AISpeech's Tianqin Voice Assistant has been adopted across more than 200 mass-produced models from over 60 automotive brands, including Mercedes-Benz, Audi, and BYD, offering full-duplex wake-up-free dialogue, multi-intent processing in a single breath, and conversation memory retention of up to two minutes.

Cloud-Edge Collaboration and Privacy Protection

Modern voice interaction predominantly adopts a cloud-edge collaborative architecture—the vehicle-side edge handles high-frequency, latency-sensitive tasks, while the cloud manages complex logical reasoning. AISpeech's central large model employs a "1+N" architecture, utilising one central model for global comprehension and task decomposition alongside N vertical-domain specialised models, supporting distributed on-device and cloud deployment to ensure high reliability even without network connectivity. This architecture effectively resolves common pain points, such as cloud service dropouts in signal dead zones like tunnels and underground car parks, as well as in-cabin voice data privacy concerns.

Technical Limitations

In-vehicle voice interaction still faces several technical hurdles. Constrained by on-board computing power and bandwidth, the response latency of on-device large models typically sits between 3 and 5 seconds, making it challenging to support scenarios requiring instantaneous feedback. Furthermore, striking an optimal balance between the complex reasoning of large models and the rapid execution of conventional vehicle control commands—while preventing intrusive "pseudo-proactive services" from disrupting motorists—remains an area of ongoing industry refinement.