Real-Time Speech Interpretation with Cross-Version Feature Conversion

The patent extracts the encoder and decoder models from the mobile device and places them on the server. The mobile device only retains lightweight client-side components that can request and use interpretation services without storing large language models locally. This extraction principle directly resolves the contradiction by removing the memory burden from the device while preserving multi-language interpretation capability through server-side resources.

The server is designed to provide universal interpretation services for multiple languages through a single multi-functional platform. Instead of requiring each device to have separate encoders and decoders for each language, the server hosts a universal set of language models that can serve any client device. This allows the system to support multiple languages (improving adaptability) without increasing the memory requirements of individual devices.

The complex model storage and management functionality is extracted from the mobile device and relocated to the server. The device becomes a simple client that communicates interpretation requests and receives results, eliminating the need for local model storage, version management, and updates. This extraction reduces device complexity while the server handles all model-related operations.

The server acts as an intermediary between the user's speech input and the interpretation output. Instead of the device directly processing large language models, the server mediates the interpretation process by receiving feature vectors from the device, performing the actual translation using stored models, and returning results. This intermediary approach simplifies the device architecture while maintaining interpretation functionality.