ASR Model Combining WFST and Attention Decoder for IVR
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional automatic speech recognition (ASR) systems struggle with high accuracy when processing short utterances, such as names or commands, in interactive voice response (IVR) systems, leading to user frustration.
Innovation Solution
An ASR model that combines a trained weighted finite state transducer (WFST) with an attention decoder to efficiently identify candidate names or commands, using a WFST to configure the system to listen for a wide range of commands and an attention decoder to select the best result with high accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional ASR systems are used to process short utterances in IVR systems, then the system structure remains simple, but the recognition accuracy deteriorates
Solution Approach 1:
The ASR system is segmented into distinct functional modules: acoustic model for phoneme recognition, language model for contextual understanding, and decision model for final transcription. This segmentation allows each module to specialize in specific aspects of speech recognition, improving overall accuracy for short utterances while maintaining manageable system complexity through modular design
Solution Approach 2:
The system transitions from traditional single-dimension speech-to-text processing to a multi-dimensional approach by integrating acoustic, linguistic, and decision-making dimensions. This dimensional expansion enables the system to capture both phonetic information and contextual meaning, significantly improving recognition accuracy for brief speech inputs
2Reliability
If generic ASR software is used for IVR commands, then the system is easy to implement, but the understanding of short commands deteriorates
Solution Approach 1:
The ASR system implements local quality by optimizing specific components for IVR command recognition rather than using uniform generic processing. The language model is tailored with domain-specific vocabulary and grammar patterns relevant to IVR systems, while the acoustic model is fine-tuned for the particular speech characteristics of command utterances, thereby improving reliability without requiring complete system redesign
Solution Approach 2:
The system performs preliminary action by pre-training and fine-tuning models with IVR-specific data before deployment. The acoustic and language models are预先 configured with knowledge of typical IVR command structures, enabling them to reliably understand short commands from the outset rather than requiring extensive adaptation during operation
Data Source
AI summary
One example method includes receiving an audio input from a user; determining, using a first trained model, a plurality of candidate commands; determining, using a second trained model, a recognized command from the plurality of candidate commands; and identifying a corresponding valid command in a set of valid commands based on the recognized command.