Feature-Space Transform Discriminative Training for ASR
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Automatic speech recognition (ASR) systems face challenges in discriminative training due to transcription errors in automatically transcribed data, which negatively impact the accuracy of feature-space transforms, especially when only a small amount of manually transcribed data is available.
Innovation Solution
A method involving a three-stage discriminative training process: first using manually transcribed data to obtain a pre-stage trained feature-space transform, then using automatically transcribed data to refine it, and finally returning to manually transcribed data for further refinement, thereby improving the robustness and accuracy of the feature-space transform.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If automatically transcribed data is used for discriminative training, then the quantity of training data increases and computational cost decreases, but transcription errors introduce harmful factors that degrade model accuracy
Solution Approach 1:
The training process is segmented into three distinct stages: pre-stage training with manually transcribed data, main-stage training with automatically transcribed data, and post-stage training with manually transcribed data. This segmentation allows the system to leverage the quantity benefits of automatic transcription while mitigating errors through structured integration with high-quality manual data at critical phases.
Solution Approach 2:
The patent implements a feedback mechanism where manually transcribed data is used to initialize and refine the feature-space transform before and after the main training stage. This feedback loop corrects errors introduced by automatic transcription by using ground-truth manual transcriptions to adjust and optimize the transform parameters, thereby reducing the harmful impact of transcription errors.
2Measurement precision
If manually transcribed data is used for discriminative training, then transcription accuracy is high, but the cost and time required for data preparation increases significantly
Solution Approach 1:
Instead of requiring complete manual transcription of all training data (excessive action), the patent applies partial manual transcription only at critical pre-stage and post-stage phases. This partial action approach maintains high accuracy where needed while reducing overall time and cost by using automatic transcription for the bulk of the training data in the main stage.
Solution Approach 2:
The system performs preliminary action by using manually transcribed data to initialize the feature-space transform before the main training stage. This preliminary setup establishes a solid foundation of accurate transformations that guides subsequent training with automatically transcribed data, reducing the need for extensive manual annotation throughout the entire process.
3Ease of manufacture
If transcription errors are present in training data, then the ease of data acquisition improves, but the reliability of discriminative training statistics deteriorates
Solution Approach 1:
The feature-space transform acts as an intermediary that mediates between the easily acquired automatic transcription data and the reliable manual transcription data. By learning to map automatic transcriptions through this transform and then refining with manual data, the system can utilize the ease of acquisition of automatic data while ensuring reliability through correction by manual data, separating these two qualities in the data processing pipeline.
Data Source
AI summary
A method, a system, and a computer program product are provided for discriminatively training a feature-space transform. The method includes performing feature-space discriminative training (f-DT) on an initialized feature-space transform, using manually transcribed data, to obtain a pre-stage trained feature-space transform. The method further includes performing f-DT on the pre-stage trained feature-space transform as a newly initialized feature-space transform, using automatically transcribed data, to obtain a main-stage trained feature-space transform. The method additionally includes performing f-DT on the main-stage trained feature-space transform as a newly initialized feature-space transform, using manually transcribed data, to obtain a post-stage trained feature-space transform.