Language Model Discriminative Training via Confusion Matrix Subword Probabilities
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition systems face challenges in accurately distinguishing between correct and incorrect transcriptions, especially when no associated correct transcription is available, due to limitations in discriminative training of acoustic and language models.
Innovation Solution
The generation of confusion matrices from acoustic model training data to calculate probabilities of subword unit confusion, insertion, deletion, and substitution errors, which are then used to discriminatively train the language model, enhancing its ability to differentiate between correct and incorrect transcriptions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If discriminative training is applied to acoustic and language models using known correct transcriptions, then the models can distinguish correct results from incorrect results more easily, but the system fails to accurately distinguish between correct and incorrect transcriptions when no associated correct transcription is available
Solution Approach 1:
The patent applies preliminary action by pre-computing confusion matrices from available training data that contains known correct transcriptions. These confusion matrices capture substitution, insertion, and deletion probabilities between subword units. The pre-computed confusion matrices are then stored and applied during inference even when no correct transcription is available, allowing the system to maintain high transcription accuracy in scenarios where traditional discriminative training would fail.
2Ease of manufacture
If traditional language model training is used without confusion matrices, then the training process is simpler, but the language model cannot effectively differentiate between correct and erroneous outputs in real-world scenarios
Solution Approach 1:
The patent introduces confusion matrices as an intermediary component between the training data and the language model. The confusion matrices serve as a mediator that encodes confusion patterns between subword units (substitutions, insertions, deletions) derived from training data with known correct transcriptions. This intermediary structure allows the language model to inherit differentiation capabilities without directly processing complex transcription pairs during inference, thus maintaining training simplicity while improving reliability.
Solution Approach 2:
The patent applies parameter changes by modifying the language model training process to incorporate confusion matrix parameters. Instead of traditional training objectives, the model uses confusion matrix-derived parameters (substitution probabilities, insertion probabilities, deletion probabilities) to guide the differentiation between correct and erroneous outputs. This parameter transformation enables the model to achieve better transcription differentiation while maintaining a relatively straightforward training framework.
Data Source
AI summary
Features are disclosed for discriminative training of speech recognition language models. A confusion matrix can be generated from acoustic model training data for use in discriminative training. The confusion matrix can include probabilities for the substitution, insertion, and/or deletion of some or all subword units of a language. Probabilities can be calculated based on the presence or absence of subword units in a processed acoustic model training data audio recording when compared to a correct transcription of the recording. The probabilities can be used to generate erroneous transcriptions in language model training corpora, and the language model can be trained to distinguish the erroneous transcriptions from the correct transcriptions.