Parallel Speech-to-Text Processing via Audio Portion Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech-to-text systems face inefficiencies and inaccuracies due to the processing of non-dialogue portions, which can lead to false positives and increased processing time, especially when using serial processing methods.
Innovation Solution
The implementation of multi-core parallel processing that identifies and separates audio file portions by type, such as dialogue, music, and silence, allowing for efficient distribution of processor job descriptors to generate text files in order, thereby reducing processing time and improving accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If serial processing methods are used to process all audio file portions, then processing is simpler to implement, but processing time increases and accuracy decreases due to false positives from non-dialogue portions
Solution Approach 1:
The audio file is divided into multiple portions based on time intervals, with each portion processed independently by separate processing units. This segmentation allows parallel processing of dialogue and non-dialogue portions, significantly reducing overall processing time while maintaining system manageability through modular processing units.
Solution Approach 2:
Non-dialogue portions are identified and extracted from the audio file using audio analysis. These extracted non-dialogue portions are then excluded from speech-to-text processing, preventing false positives and improving transcription accuracy while reducing unnecessary processing overhead.
2Measurement precision
If all audio file portions including non-dialogue are processed, then complete coverage is achieved, but accuracy decreases due to false positives
Solution Approach 1:
Audio analysis is performed preliminarily to identify and classify portions as dialogue or non-dialogue before speech-to-text processing begins. This preliminary classification ensures that only relevant dialogue portions are processed, improving accuracy by eliminating false positives from non-dialogue segments.
Solution Approach 2:
Non-dialogue portions are extracted and excluded from the processing pipeline. By removing these irrelevant portions beforehand, the system achieves higher transcription accuracy without the time penalty of processing unnecessary audio segments.
3Productivity
If larger audio file portions are processed in parallel, then productivity increases, but storage and transfer requirements increase
Solution Approach 1:
Processing unit descriptors containing metadata about audio portions are copied and distributed to multiple processing units, rather than copying the actual audio data. This approach enables parallel processing of large portions while minimizing storage and transfer requirements, as only lightweight descriptor data needs to be transmitted.
Solution Approach 2:
Processing unit descriptors act as intermediaries between the audio file and the speech-to-text processing units. These descriptors contain all necessary information (time intervals, portion identifiers) to guide processing without requiring the actual audio data to be transferred and stored across multiple units.
Data Source
AI summary
This specification describes technologies relating to multi core processing for parallel speech-to-text processing. In some implementations, a computer-implemented method is provided that includes the actions of receiving an audio file; analyzing the audio file to identify portions of the audio file as corresponding to one or more audio types; generating a time-ordered classification of the identified portions, the time-ordered classification indicating the one or more audio types and position within the audio file of each portion; generating a queue using the time-ordered classification, the queue including a plurality of jobs where each job includes one or more identifiers of a portion of the audio file classified as belonging to the one or more speech types; distributing the jobs in the queue to a plurality of processors; performing speech-to-text processing on each portion to generate a corresponding text file; and merging the corresponding text files to generate a transcription file.