Speech Recognition Clustering to Reduce Manual Transcription Time
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The development of automatic speech recognition (ASR) systems requires large manually transcribed datasets, which is resource-intensive and time-consuming, and existing methods fail to efficiently process and update grammars to include all user utterances, leading to poor speech recognition.
Innovation Solution
The method involves clustering similar training samples to reduce manual processing, where only a few representative samples from each cluster are manually processed, and assigning weighting factors based on cluster size to ensure accurate training, thereby reducing the need for extensive human involvement and improving grammar updates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a large corpus of training samples is manually transcribed to ensure accurate speech recognition, then the accuracy of speech recognition is improved, but the time and resources required for manual transcription increase significantly
Solution Approach 1:
The patent combines multiple identical or similar audio samples into clusters, where samples with the same or similar transcriptions are grouped together. This merging allows the system to process fewer unique samples manually while still capturing the diversity of the full dataset, thereby reducing transcription time while maintaining accuracy.
Solution Approach 2:
The patent creates synthetic training samples by copying and slightly modifying existing transcribed samples. By generating variations of proven samples (e.g., adding background noise, changing speed slightly), the system expands the training corpus without requiring equivalent manual transcription effort, thus improving accuracy efficiently.
2Adaptability or versatility
If all training samples are manually processed to ensure comprehensive grammar coverage, then the grammar completeness is improved, but the productivity and efficiency of the development process deteriorate
Solution Approach 1:
The patent performs preliminary clustering of samples before manual transcription. By grouping similar samples in advance, the system identifies which samples need manual processing and which can be handled automatically or synthesized, thereby improving development efficiency while ensuring comprehensive grammar coverage through targeted manual processing of representative samples.
Solution Approach 2:
The patent changes the parameter of sample representation by using clustering to identify representative samples from each group. Instead of processing all samples individually, the system selects key representatives and processes them manually, then uses parameter variations (synthetic modifications) to generate additional training data, thereby maintaining grammar coverage while boosting productivity.
3Reliability
If manual transcription is performed on every unique sample to avoid redundancy, then the data quality is improved, but the loss of time and computational resources increases
Solution Approach 1:
The patent merges identical or near-identical audio samples into clusters, recognizing that processing each unique sample separately is redundant. By combining these samples and processing only one representative from each cluster, the system maintains data quality while significantly reducing computational resources required for transcription and processing.
Solution Approach 2:
Instead of manually transcribing every unique sample, the patent creates synthetic copies of transcribed samples with varied parameters (noise levels, playback speeds, pitch variations). This copying approach generates diverse training data with high quality while consuming far fewer computational resources than manual transcription of all samples.
Data Source
AI summary
Techniques for grouping a plurality of samples automatically transcribed from a plurality of utterances. The method comprises forming clusters from the plurality of samples, wherein the clusters include two or more of the plurality of samples. One or more samples are selected from a cluster and manually-processed data samples for the one or more samples are obtained. A weighting factor may be assigned to the data samples based, at least in part, on the number of samples in the cluster associated with the selected data sample.