Parallel Hidden Markov Model Training via Local Cache Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for training Hidden Markov Models for speech recognition are inefficient, requiring extensive time and resources, with previous parallel computing approaches achieving less than 3× speedup over sequential runs and being unsuitable for continuous observation models or leveraging special model structures.
Innovation Solution
The method optimizes the calculation of matching scores between phone states and acoustic frames by distributing phone states across local caches on a chip, scheduling calculations based on the matrix structure, and utilizing parallel computing to evaluate phone states independently across cores, allowing simultaneous evaluation of multiple phone states and reducing processing time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If parallel computing is used to train Hidden Markov Models, then training speed is improved, but previous approaches achieved less than 3× speedup and are not suitable for continuous observation models
Solution Approach 1:
The patent segments the training process by distributing different phone states to different processing cores. Each core independently processes matching scores for its assigned phone states, enabling parallel computation. This segmentation allows the system to handle continuous observation models effectively while achieving significant speedup, as each core can process continuous probability distributions without interference from other cores.
Solution Approach 2:
The patent transitions from sequential processing to parallel processing by adding a spatial dimension to the computation. Multiple cores operate simultaneously on different phone states, transforming the training process from a one-dimensional sequential operation to a multi-dimensional parallel operation. This dimensional change enables handling of continuous observation models while achieving over 3× speedup.
2Speed
If phone states are stored in local caches on a chip, then access speed is improved, but memory distribution complexity increases
Solution Approach 1:
The patent applies local quality by storing phone states in local caches on the chip, bringing data closer to where it is processed. Each processing core has access to its assigned phone states in local memory, reducing access latency and improving computation speed. This localized data placement optimizes the trade-off between access speed and memory distribution complexity.
Data Source
AI summary
A method of optimizing the calculation of matching scores between phone states and acoustic frames across a matrix of an expected progression of phone states aligned with an observed progression of acoustic frames within an utterance is provided. The matrix has a plurality of cells associated with a characteristic acoustic frame and a characteristic phone state. A first set and second set of cells that meet a threshold probability of matching a first phone state or a second phone state, respectively, are determined. The phone states are stored on a local cache of a first core and a second core, respectively. The first and second sets of cells are also provided to the first core and second core, respectively. Further, matching scores of each characteristic state and characteristic observation of each cell of the first set of cells and of the second set of cells are calculated.