Bone-Conduction Speech Synthesis for Wearable Audio Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing wearable audio devices face challenges in effectively processing bone conduction signals due to the lack of sufficient corpora for training machine-learning models, leading to expensive and time-consuming data collection processes, which vary by device model and user-specific characteristics.
Innovation Solution
A first machine-learning model synthesizes bone conduction signals from acoustic signals, using a transfer function to train a second model, enabling efficient enhancement and suppression of speech in multimodal processing without extensive data collection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional data collection methods are used to train machine-learning models for bone conduction signal processing, then model accuracy can be improved, but the process becomes expensive and time-consuming
Solution Approach 1:
The patent uses acoustic signals as copies or proxies for bone conduction signals during the training phase. The first machine-learning model learns to map acoustic signals to bone conduction signals, creating a synthetic dataset that mimics real bone conduction data without requiring actual collection. This copying approach enables the second model to be trained effectively using alternative, faster data sources while maintaining model accuracy.
2Measurement precision
If device-specific and user-specific data collection is performed, then processing accuracy for specific devices and users is improved, but the complexity and cost of data collection increases
Solution Approach 1:
The first machine-learning model serves multiple functions: it acts as a transfer function that adapts generic acoustic signal processing to device-specific characteristics, and simultaneously personalizes processing for individual users. By using the acoustic signal as a universal input that can be transformed into device-specific and user-specific bone conduction signals, the system achieves personalized processing without requiring separate data collection processes for each device-user combination.
3Reliability
If extensive corpora are collected for training, then machine-learning model performance is improved, but the manufacturing and deployment process becomes more complex
Solution Approach 1:
The patent replaces the mechanical process of physical data collection with a computational approach. Instead of physically collecting bone conduction signals from multiple users and devices, the system uses machine-learning models to synthesize training data computationally. The first model transforms acoustic signals into synthetic bone conduction signals, creating a virtual corpus that eliminates the need for extensive physical data collection infrastructure and simplifies deployment.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
This approach allows for improved speech enhancement and suppression in wearable audio devices, reducing latency and enhancing user experience in voice communication and extended reality applications.
Implementation Method 1
the vibration sensor(s) may capture bone conducted speech
Data Source
AI summary
Techniques, including wearable audio devices and systems implementing the techniques, for synthesizing bone conduction speech. Such techniques may include (i) inputting a first acoustic signal into a first machine-learning model, (ii) generating, with the first machine-learning model, a first bone conduction signal in a time domain or a spectral domain based, at least in part, on the first acoustic signal, (iii) generating, with the first machine-learning model, a transfer function that characterizes a relationship between the first acoustic signal and the first bone conduction signal based, at least in part, on the first acoustic signal, and (iv) training, using at least one of the first bone conduction signal or the transfer function, a second machine-learning model on a first device.