Neural Network Language Identification via Score Normalization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional multi-lingual speech processing devices face delays in language identification, consuming significant computational resources and taking time to initiate speech processing, especially when dealing with unknown languages, which hinders quick communication between speakers of different languages.
Innovation Solution
A language identifying device that uses a neural network to output scores for each language, normalizes these scores, and selectively processes the language with the highest normalized score, allowing for rapid language identification and switching to ensure efficient speech processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If speech recognition is performed in multiple languages to identify the language of the speech signal, then language identification accuracy is improved, but processing time increases and computational resources are consumed
Solution Approach 1:
The patent segments the language identification task by first performing acoustic feature extraction, then feeding these features into a neural network that outputs language scores independently of full speech recognition. This segmentation allows language identification to proceed without completing entire speech recognition pipelines for multiple languages, thereby reducing processing time while maintaining accuracy.
Solution Approach 2:
The patent performs preliminary acoustic feature extraction from the speech signal before language identification. These pre-computed acoustic features are then used by the neural network to determine language probabilities. This preliminary action avoids redundant feature extraction that would occur if full speech recognition were performed for each candidate language, thus reducing overall processing time.
2Measurement precision
If speech recognition is performed in multiple languages simultaneously, then language identification accuracy is improved, but computational resource consumption increases
Solution Approach 1:
The patent extracts only the language identification function from the full speech recognition system. Instead of running complete speech recognition pipelines for multiple languages, the system extracts acoustic features and feeds them into a dedicated neural network for language scoring. This extraction eliminates the computational overhead of full speech recognition while preserving language identification accuracy.
Solution Approach 2:
The patent uses a neural network that is trained to copy the language discrimination capability of multiple speech recognition systems without requiring those systems to actually run. The network learns language identification patterns from training data and reproduces this capability independently, avoiding the need to instantiate and execute multiple speech recognition engines simultaneously.
3Ease of operation
If the device is set to a default language (e.g., Japanese), then the owner can operate the device easily, but speakers of other languages cannot understand the device interface
Solution Approach 1:
The patent makes the device interface dynamic by automatically detecting the language of the speech signal and adapting the display language accordingly. When a speech input is received, the system identifies the language using the neural network and switches the interface language to match the speaker's language. This dynamic adaptation allows the device to serve both the owner (who can set a default) and foreign speakers (whose language is automatically detected and displayed).
Data Source
AI summary
A language identifying device includes: a neural network responsive to an input of a speech signal, for outputting a score for each of a plurality of languages indicating that the speech represented by the speech signal is the speech of the corresponding language; a selecting device for selecting, among the scores output by the neural network, scores of a smaller number of languages; a normalizing device for normalizing the scores selected by the selecting device; a determining device for determining whether the maximum of the normalized scores normalized by the normalizing device is equal to or higher than a threshold; and a language deciding device, responsive to a positive determination by the determining device, for selectively performing a process of deciding the language corresponding to the maximum of the scores as the language of the speech represented by the speech signal, or a process of discarding outputs of the score selecting device.