System and method for disambiguation and error resolution in call transcripts

A system for detecting and resolving mis-transcriptions in a transcript generated by an automatic speech recognition (ASR) system when transcribing a plurality of spoken words, the system being configured to:

• receive the generated transcript of a speech signal;
• receive, at a research on artificial intelligence in linguistics and systems (RAILS) model from the ASR system, a corpus comprising the generated transcript and a plurality of other transcripts, the RAILS model comprising a building block comprising (1) a dataset containing all unigrams, bigrams, trigrams and quadgrams present in the corpus and their respective probabilities, and (2) a phonetic encoding component for determining phonetic similarity between lexical items;
• integrate a pre-trained language model, including a bidirectional encoder representation from transformers (BERT) model or a generative pre-trained transformer (GPT-3) model, to further analyze and process the transcript;
• implement at least one of a first machine learning system and a second machine learning system that
• (i) analyzes the generated transcript in tandem or in parallel,
• (ii) finds a region of low confidence indicative of a mis-transcription utilizing the RAILS model, and
• (iii) analyzes the region of low confidence utilizing the BERT model or the GPT-3 model subsequent to the analyzing utilizing the RAILS model to predict an improvement to the region of low confidence, the predicted improvement comprising a homophone of the found region having substantially a same pronunciation and different meaning therefrom, wherein the same pronunciation indicates a higher level of phonetic similarity than one homophone having only a similar pronunciation;
• receive the generated transcript and the predicted improvement to the region of low confidence, wherein the improvement is predicted responsive to the region of low confidence being partially caused by a degraded signal of a channel used to perform at least one of the receptions or an obtainment of the spoken words;
• based on the predicted improvement to the region of low confidence, select a replacement word for the mis-transcription;
• and
• replace the mis-transcription by the replacement word.

A non-transitory computer readable storage medium comprising instructions for detecting and resolving mis-transcriptions in a transcript generated by an automatic speech recognition system (ASR) when transcribing a plurality of spoken words, the instructions, when executed by a processor, cause the processor to:

• receive the generated transcript of a speech signal;
• receive, at a research on artificial intelligence in linguistics and systems (RAILS) model from the ASR system, a corpus comprising the generated transcript and a plurality of other transcripts, the RAILS model comprising a building block comprising (1) a dataset containing all unigrams, bigrams, trigrams and quadgrams present in the corpus and their respective probabilities, and (2) a phonetic encoding component for determining phonetic similarity between lexical items;
• integrate a pre-trained language model, including a bidirectional encoder representation from transformers (BERT) model or a generative pre-trained transformer (GPT-3) model, to further analyze and process the transcript;
• implement at least one of a first machine learning system and a second machine learning system that analyzes the generated transcript to find a region of low confidence indicative of a mis-transcription utilizing the RAILS model, the at least one of the first machine learning system and the second machine learning system comprising the RAILS model, wherein the first machine learning system is configured to analyze the region of low confidence utilizing the BERT model or the GPT-3 model subsequent to the analyzing utilizing the RAILS model to predict an improvement to the region of low confidence, the predicted improvement comprising a homophone of the found region having substantially a same pronunciation and different meaning therefrom, wherein the same pronunciation indicates a higher level of phonetic similarity than one homophone having only a similar pronunciation;
• select, by a word selector, a replacement word for the mis-transcription based on the predicted improvement to the region of low confidence, wherein the improvement is predicted responsive to the region of low confidence being partially caused by a degraded signal of a channel used to perform at least one of the receptions or an obtainment of the spoken words;
• and
• replace, by the word selector, the mis-transcription by the replacement word.