Marchex, Inc.
Automatic speech recognition (ASR) model training

Abstract:

The disclosed system continuously refines a model used by an Automatic Speech Recognition (ASR) system to enable fast and accurate transcriptions of detected speech activity. The ASR system analyzes speech activity to generate text transcriptions and associated metrics (such as minimum Bayes risk and/or perplexity) that correspond to the quality of or confidence in each generated transcription. The system employs a filtering process to select certain text transcriptions based in part on one or more associated quality metrics. In addition, the system corrects for known systemic errors within the ASR system and provides a mechanism for manual review and correction of transcriptions. The system selects a subset of transcriptions based on factors including confidence score, and uses the selected subset of transcriptions to re-train the ASR model. By continuously retraining the ASR model, the system is able to provide ever faster and more accurate text transcriptions of detected speech activity.