Acoustically Informed Language Model Pruning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Language models used in automatic speech recognition (ASR) face challenges in differentiating between similar sounding words, leading to reduced accuracy when pruned to reduce size and complexity, as pruning can result in incorrect transcriptions due to loss of contextual information.
Innovation Solution
A process for pruning language models that evaluates candidate portions based on both information loss and acoustic confusability, using entropy scores and acoustic scores to determine whether to remove or retain sequences, thereby balancing model size reduction with maintaining transcription accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If the language model is pruned to reduce size and complexity, then processing speed and cost are improved, but transcription accuracy deteriorates due to loss of contextual information needed to differentiate similar sounding words
Solution Approach 1:
The patent applies local quality by differentiating pruning strategies based on acoustic properties of specific word sequences. High-acoustic-confusability sequences (those with similar sounding words) are retained with full contextual information, while low-acoustic-confusability sequences are pruned more aggressively. This localized approach to pruning maintains accuracy where needed while achieving compression elsewhere, resolving the contradiction between model size and transcription accuracy.
Solution Approach 2:
The patent changes the pruning parameter from a uniform size-based metric to an acoustic confusability-based metric. By introducing acoustic confusability as a new parameter for evaluating which sequences to retain, the system can selectively preserve sequences that are critical for accurate transcription (those with homophones) while removing less critical sequences, thus maintaining accuracy while reducing overall model size.
2Measurement precision
If sequences with more items are retained to improve differentiation of similar sounding words, then transcription accuracy is improved, but model size and complexity increase
Solution Approach 1:
The patent makes model complexity non-uniform by applying different pruning levels to different sequences based on their acoustic confusability. Sequences containing homophones or similar sounding words retain full contextual depth (higher complexity), while sequences without such challenges are pruned to shorter contexts (lower complexity). This localized complexity allocation achieves accurate transcription where needed without uniformly increasing model complexity.
Solution Approach 2:
The patent applies partial action by selectively retaining full contextual sequences only for high-acoustic-confusability cases rather than uniformly retaining all long sequences. This partial retention strategy achieves the necessary transcription accuracy for difficult cases while avoiding the unnecessary complexity burden of retaining all possible long sequences, thus resolving the contradiction between accuracy and complexity.
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for pruning a language model are disclosed. The methods, systems, and apparatus include actions of selecting a candidate portion of the language model to evaluate for pruning, obtaining an entropy score representing information loss that would result from pruning the candidate portion of the language model, obtaining an acoustic score representing acoustic confusability of one or more words modeled by the candidate portion of the language model, and evaluating whether to prune the candidate portion of the language model using the entropy score and the acoustic score.