Voice Topic Spotting via Fast Keyword Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing voice topic spotting systems face challenges in processing speed and efficiency, particularly on computing platforms with limited processing power, and require significant user input for annotation, which hinders practical implementation.
Innovation Solution
A voice topic spotting system that employs machine learning to identify keywords from ASR output and applies a fast keyword filtering system to reduce processing load, using a learning component to generate models that speed up topic identification by rejecting non-relevant audio segments and focusing processing on potentially topically relevant segments.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If ASR/LVCSR is used to extract words from audio recordings for topic spotting, then topic identification accuracy is improved, but processing time and computation cost increase significantly
Solution Approach 1:
The system segments the audio processing task into two distinct stages: a fast keyword spotting stage that identifies potential relevant segments using simple acoustic matching, and a secondary ASR stage that processes only those segmented portions for accurate topic classification. This segmentation allows the computationally intensive ASR to operate on a small fraction of the total audio data while maintaining high topic identification accuracy.
Solution Approach 2:
The system performs preliminary keyword spotting using fast acoustic matching before applying ASR. This preliminary action filters out irrelevant audio segments, so that ASR is only applied to segments that contain potential topic keywords. This preliminary filtering significantly reduces the overall processing time while preserving accurate topic identification for relevant segments.
2Measurement precision
If ASR/LVCSR is used to extract words from audio recordings for topic spotting, then topic identification accuracy is improved, but computation cost increases significantly
Solution Approach 1:
The processing pipeline is segmented into a low-cost keyword spotting phase and a high-cost ASR phase. By segmenting the task this way, the system ensures that expensive ASR computation is only applied to audio segments that have already been pre-identified as potentially relevant through fast keyword matching, thereby dramatically reducing total computation cost while maintaining accuracy.
Solution Approach 2:
The system applies ASR only partially - specifically, only to audio segments that contain keyword matches from the preliminary filtering stage. This partial application of the computationally intensive ASR process avoids the excessive computation cost of running ASR on entire audio streams, while still achieving accurate topic identification for the relevant portions.
3Quantity of substance
If existing voice topic spotting systems process large volumes of audio recordings, then comprehensive topic coverage is improved, but processing speed decreases
Solution Approach 1:
The system performs preliminary keyword indexing on large volumes of audio recordings during an offline preparation phase. This preliminary action creates a searchable keyword database that enables fast online querying and topic spotting, allowing the system to maintain high processing speeds even when dealing with large volumes of audio data.
Solution Approach 2:
The system extracts and indexes only the essential keyword information from large volumes of audio recordings during offline processing. This extraction creates a compressed representation that can be quickly searched and matched during online operation, enabling the system to handle large audio volumes at high speeds without processing the entire audio data in real-time.
4Measurement precision
If user-specified topics are implemented with detailed annotation, then topic identification accuracy is improved, but user input requirement increases
Solution Approach 1:
The system automatically generates topic models and keyword associations by processing user-provided example audio segments without requiring manual annotation. The ASR and keyword spotting components self-service by automatically extracting relevant terms and building topic representations, eliminating the need for users to perform time-consuming manual annotation while maintaining high topic identification accuracy.
Data Source
AI summary
A voice topic spotting system includes a learning module and a voice topic classifier module. The learning module receives training audio segments with topic labels and generates a fast keyword filter model based on a set of topic-indicative words and generates a topic identification model based on a training set of topic keyword-containing lattices. The voice topic classifier module includes an automatic speech recognition engine arranged to identify one or more keywords included in a received audio segment and output the one or more keywords. A fast keyword filter, implements the fast keyword model to output the received audio segment if a topic-indicative word is detected in the audio segment. A decoder generates a topic keyword-containing lattice associated with the audio segment. A voice topic classifier implements the voice topic identification model to determine a topic associated with received audio segment.