Speech Transcription Filtering for Human-Like Messaging
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition systems either use statistical language models (SLMs) for transcription or finite grammars for post-processing, but not both simultaneously, limiting their ability to accurately transcribe free-form speech into text that mimics human-written messages, especially in mobile device messaging contexts.
Innovation Solution
A method that receives audio data from a mobile device, transcribes it using SLMs or Hierarchical Language Models, and applies filters such as finite grammars or software filters to modify the text, inserting punctuation, replacing words, and formatting it to resemble human-written messages, which can be communicated over HTTP/HTTPS protocols.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If statistical language models are used for speech transcription, then the system can handle free-form speech, but the accuracy and readability of the transcribed text deteriorates compared to human-written messages
Solution Approach 1:
The patent introduces finite grammar rules and software filters as intermediary components between the statistical language model and the final transcribed text. These intermediaries process the raw transcription output, applying linguistic rules to correct errors, insert appropriate punctuation, and format the text to better mimic human-written messages, thereby improving accuracy while maintaining the ability to handle free-form speech.
2Manufacturing precision
If finite grammars are used for post-processing, then the text accuracy improves, but the system cannot simultaneously use statistical language models for transcription
Solution Approach 1:
The patent merges statistical language models with finite grammar rules and software filters into a unified speech-to-text system. The statistical language model performs initial transcription of free-form speech, and the finite grammar rules are applied in sequence as post-processing filters to refine the output. This combination allows the system to leverage the strengths of both approaches: the adaptability of statistical models for free-form speech and the precision of finite grammars for text accuracy.
3Manufacturing precision
If multiple filters are applied to transcribed text, then the readability and human-like quality improves, but the processing time and system complexity increases
Solution Approach 1:
The patent segments the text processing function into multiple independent filter modules, each responsible for a specific task such as punctuation insertion, capitalization, abbreviation expansion, or grammar correction. This segmentation allows the filters to be applied in a standardized sequence, making the system more manageable and maintainable despite the increased number of components. Each filter operates independently on the output of the previous filter, simplifying the overall system architecture.
Data Source
AI summary
A method for facilitating mobile phone messaging, such as text messaging and instant messaging, includes receiving audio data communicated from the mobile communication device, the audio data representing an utterance that is intended to be at least a portion of the text of the message that is to be sent from the mobile phone to a recipient; transcribing the utterance to text based on the received audio data to generate a transcription; and applying a filter to the transcribed text to generate a filtered transcription, the text of which is intended to mimic language patterns of mobile device messaging that is performed manually by users. The method may also be applied to the audio data of a voicemail, with the filtered, transcribed text being communicated to a mobile phone as, for example, an SMS text message.