Distributed Acoustic Model Training via Node Resource Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current large corpus-based acoustic model training methods face inefficiencies due to excessive memory occupancy by Hidden Markov Models (HMMs), limiting parallel processing and resulting in prolonged training times.
Innovation Solution
A method and device for training acoustic models by dividing tasks based on resource occupancies and complexities of nodes, allowing for distributed task execution across multiple devices, including voice parameter extraction and HMM-based Speech Synthesis System (HTS) training, to improve training efficiency and reduce memory requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If single-machine partial-task multi-processes approach is used for training, then memory occupancy is reduced, but training time increases significantly
Solution Approach 1:
The patent divides the acoustic model training into multiple independent tasks (e.g., voice parameter extraction, HMM training, Gaussian mixture modeling) that can be executed in parallel across multiple nodes. Each node handles a subset of tasks, enabling distributed computation that reduces memory occupancy per node while maintaining overall training efficiency through parallel processing.
2Adaptability or versatility
If large corpus is used for acoustic model training, then model context coverage improves, but memory occupancy becomes excessive
Solution Approach 1:
The patent segments the large corpus into multiple smaller subsets and distributes them across different nodes. Each node processes a portion of the corpus for specific tasks, enabling the system to handle large corpora for comprehensive model training while keeping individual node memory occupancy manageable through distributed data storage and processing.
3Productivity
If more processes run in parallel for training, then training efficiency improves, but memory occupancy increases
Solution Approach 1:
The patent transitions from single-machine parallel processing to multi-node distributed processing, adding the dimension of spatial distribution. By deploying processes across multiple physical or virtual nodes, the system achieves parallel processing benefits while distributing memory occupancy across multiple devices, thereby maintaining training efficiency without excessive memory consumption on any single node.
Data Source
AI summary
A method and device for training an acoustic model are provided. The method comprises determining a plurality of tasks for training an acoustic model, obtaining resource occupancies of nodes participating in the training of the acoustic model, and distributing the tasks to the nodes according to the resource occupancies of the nodes and complexities of the tasks. By using computational resources distributed at multiple nodes, tasks for training an acoustic model are performed in parallel in a distributed manner, so as to improve training efficiency.