Acoustic Model Training Using Segmented Imperfect Transcription Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing automatic speech recognition technologies face challenges in utilizing large amounts of imperfectly transcribed speech data, leading to suboptimal acoustic models due to transcription errors, which are often costly and difficult to correct.
Innovation Solution
The technique involves dividing a corpus of training data into parts, decoding utterances with incremental acoustic and language models, aligning transcriptions, and selecting contiguous matching segments to train incremental acoustic models, allowing for the use of more data and improving model accuracy over rounds.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If large amounts of imperfectly transcribed speech data are used for training, then the quantity of training data increases, but the accuracy of the acoustic model deteriorates due to transcription errors
Solution Approach 1:
The training corpus is divided into multiple parts, and the training process is segmented into multiple rounds. In each round, only segments with sufficient contiguous matching words (at least Q contiguous matches) are selected for training. This segmentation allows the system to process large amounts of imperfect data while maintaining model accuracy by selectively using only the reliable portions.
Solution Approach 2:
The patent converts the harmful effect of transcription errors into a beneficial filtering mechanism. By using the incremental language model to decode utterances and compare with original transcriptions, errors are automatically identified and excluded. The mismatches caused by errors become the basis for selecting only high-quality training segments, thus converting data imperfections into a quality control mechanism.
2Manufacturing precision
If transcription errors are corrected to improve model accuracy, then the accuracy of the acoustic model improves, but the cost and time required for data preparation increases
Solution Approach 1:
The system performs automatic error detection and filtering without human intervention. The incremental language model decodes utterances and compares them with original transcriptions, automatically identifying segments with errors. This self-service mechanism eliminates the need for manual transcription verification, achieving high model accuracy without the time cost of human data preparation.
Solution Approach 2:
The patent replaces the mechanical process of manual transcription correction with an automated computational system. The incremental language model and decoding process substitute for human editors, automatically identifying and excluding erroneous segments through algorithmic comparison and matching, thus eliminating time-consuming manual data preparation.
3Manufacturing precision
If manual correction of transcription errors is performed, then the quality of training data improves, but the ease of manufacture deteriorates due to increased complexity
Solution Approach 1:
The system automatically performs quality control through the incremental decoding and comparison process. The language model and decoding mechanism self-identify erroneous segments without requiring external intervention or complex manual procedures, maintaining ease of manufacture while ensuring high data quality.
4Reliability
If incremental acoustic models are trained with selected segments, then the reliability of the model improves, but the productivity of training decreases due to multiple processing rounds
Solution Approach 1:
The training process is segmented into multiple rounds, each processing a subset of the corpus. While this improves reliability through iterative refinement, the segmentation into manageable parts also allows for efficient parallel processing and targeted training, mitigating the productivity loss.
Solution Approach 2:
The system performs preliminary filtering and segment selection before actual model training. By pre-identifying and selecting only high-quality segments with sufficient contiguous matches, the system prepares optimized training data in advance, making the subsequent training rounds more efficient and reducing overall training time.
Data Source
AI summary
Techniques and systems for training an acoustic model are described. In an embodiment, a technique for training an acoustic model includes dividing a corpus of training data that includes transcription errors into N parts, and on each part, decoding an utterance with an incremental acoustic model and an incremental language model to produce a decoded transcription. The technique may further include inserting silence between a pair of words into the decoded transcription and aligning an original transcription corresponding to the utterance with the decoded transcription according to time for each part. The technique may further include selecting a segment from the utterance having at least Q contiguous matching aligned words, and training the incremental acoustic model with the selected segment. The trained incremental acoustic model may then be used on a subsequent part of the training data. Other embodiments are described and claimed.