Keyword Detection Using IPA Acoustic Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing keyword detection methods face challenges in accurately recognizing keywords across multiple languages due to variations in pronunciation caused by external factors and the limitations of training neural networks on a single language, leading to decreased accuracy in non-trained languages and environmental noise.
Innovation Solution
The method employs an acoustic model trained with International Phonetic Alphabet (IPA) phoneme mapping and multiple languages, using a foreground model to match keywords and a background model to filter non-keywords, eliminating the need for Dynamic Time Warping (DTW) and enhancing robustness across languages.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If DTW algorithm is used for keyword detection, then processing speed is high and implementation is simple, but keyword detection accuracy decreases significantly in environmental noise
Solution Approach 1:
The patent replaces the mechanical DTW algorithm with a neural network-based acoustic model that learns phoneme patterns. The neural network processes audio features through multiple layers to identify keywords, substituting the direct characteristic comparison approach with a learned representation approach that is more robust to environmental noise while maintaining efficiency.
2Measurement precision
If neural network is trained on a single language, then keyword detection performance is good for that language, but performance deteriorates when extended to other languages
Solution Approach 1:
The patent creates a universal acoustic model using International Phonetic Alphabet (IPA) representations that can handle multiple languages. By mapping different language phonemes to IPA symbols, the neural network learns language-agnostic phoneme patterns, enabling the same model to detect keywords across various languages without retraining, thus achieving both accuracy and multi-language adaptability.
3Adaptability or versatility
If audio samples from multiple languages are used for training, then multi-language detection capability is improved, but model complexity and training difficulty increase
Solution Approach 1:
The patent introduces IPA phoneme mapping as an intermediary layer between raw audio input and the neural network. This intermediary representation standardizes different language phonemes into a common framework, allowing the model to learn universal phoneme patterns without being confused by language-specific variations, thereby reducing training complexity while maintaining multi-language capability.
Data Source
AI summary
An electronic device with one or more processors and memory trains an acoustic model with an international phonetic alphabet (IPA) phoneme mapping collection and audio samples in different languages, where the acoustic model includes: a foreground model; and a background model. The device generates a phone decoder based on the trained acoustic model. The device collects keyword audio samples, decodes the keyword audio samples with the phone decoder to generate phoneme sequence candidates, and selects a keyword phoneme sequence from the phoneme sequence candidates. After obtaining the keyword phoneme sequence, the device detects one or more keywords in an input audio signal with the trained acoustic model, including: matching phonemic keyword portions of the input audio signal with phonemes in the keyword phoneme sequence with the foreground model; and filtering out phonemic non-keyword portions of the input audio signal with the background model.