Dynamic ASR System for Multilingual Edge Speech Recognition

The patent segments the speech recognition system into multiple language-specific models, each trained on a particular language. A language identification model first determines which language is being spoken, then routes the audio to the corresponding language-specific ASR model. This segmentation allows the system to support multiple languages while keeping each individual model relatively simple and manageable.

The patent introduces a language identification (LangID) model as an intermediary component between the audio input and the speech recognition models. This mediator first analyzes the audio to determine the language being spoken, then selects the appropriate ASR model for that language. This intermediary approach enables multilingual support without requiring a single complex model to handle all languages simultaneously.

The patent applies local quality by deploying language-specific ASR models that are optimized for particular languages and can be selectively executed based on the detected language. This allows the system to use smaller, more efficient models for specific language tasks rather than running a single large universal model, reducing computational resource consumption while maintaining accuracy for each language.

The patent implements dynamic model selection where the system adapts its computational resources based on the input language. The language identification model dynamically determines which ASR model to load and execute, allowing the system to optimize resource usage by only activating the necessary model for the current task rather than maintaining all models in active memory simultaneously.

The patent segments the training data into language-specific and accent-specific datasets, creating separate training corpora for different languages and accents. This allows each model to be trained on balanced, representative data for its specific language and accent domain, improving reliability for minority accents while maintaining training efficiency by avoiding the need to process all diverse data in a single model.

The patent changes the training parameters by creating language-specific and accent-specific training datasets with balanced representations. Instead of using a single imbalanced dataset for all languages, the system trains separate models on curated datasets that are balanced for their specific language and accent domains, improving performance reliability across diverse accents while maintaining training efficiency through specialized, focused training data.