DIVINES – Speech Rec. and Intrinsic Variation W.S.May 20, 2006 Richard Rose DIVINES SRIV Workshop The Influence of Word Detection Variability on IR Performance in Automatic Audio Indexing of Course Lectures Saturday May 20, 2006 Richard Rose 1, Renato Rispoli 1, and Jon Arrowood 2 1 McGill University Dept of ECE Montreal, QC Canada 2 Nexidia Inc. Atlanta, GA USA
DIVINES – Speech Rec. and Intrinsic Variation W.S.May 20, 2006 Richard Rose Indexing Audio Lectures Existing multimedia resources have the potential to make recorded University lectures and seminars accessible online to a wider audience It is important that the audio lectures be searchable … … but, human annotation of large corpora is expensive Automatic Speech Recognition (ASR) based tools can be used to facilitate search of the un-transcribed audio material
DIVINES – Speech Rec. and Intrinsic Variation W.S.May 20, 2006 Richard Rose An Audio Search Tool for Course Lectures Text Query Term Retrieved Segments from Lecture Audio Files Click to listen to audio segment Synchronized Presentation Slides User Interface Developed by Nexidia
DIVINES – Speech Rec. and Intrinsic Variation W.S.May 20, 2006 Richard Rose Audio Indexing of Lectures - Motivation Goal – Provide Disabled and Non-Disabled Students and Scholars Access to a Large Collection (thousands of hours) of Audio Lectures and Seminars Multimedia – Permit Synchronization and Interpretation of audio with Lecture Slides and Video Content Challenges – Large variability in dialect, speaking style, recording conditions, and task domain
DIVINES – Speech Rec. and Intrinsic Variation W.S.May 20, 2006 Richard Rose Issues in Audio Indexing Acoustic – Extraction of query terms from audio –Must be extremely fast during search (>>1,000 X real-time) Information Retrieval (IR) – Definition of relevance measure –Score query against hypothesized audio segment Task Domain - Definition of the notion of relevance –When does relevant segment begin and end? Evaluation Metrics –Acoustic: ASR word error rate, Keyword detection performance –IR: Precision / Recall of relevant segments –Task Domain: Increase in Productivity for the target user community
DIVINES – Speech Rec. and Intrinsic Variation W.S.May 20, 2006 Richard Rose Audio Indexing Task Domains Several techniques have been applied to indexing of spoken audio in several task domains: [Rose, 1991]: –Task: Topic Spotting from Conversational Speech –Method: Keyword spotting [Foote et al, 1997]: –Task: Retrieval of multimedia mail messages (Video mail browser) –Method: Phone lattice based open vocabulary indexing [Garofolo, 2000]: –Task: Spoken Document Retrieval (SDR) from Broadcast News –Method: Large vocabulary continuous speech recognition (LVCSR) Course Lectures: –How to define a topic of interest? –How to segment a continuous lecture by topic? –How to define query terms and extract them from audio?
DIVINES – Speech Rec. and Intrinsic Variation W.S.May 20, 2006 Richard Rose Phone Lattice-Based Search Engine –Off-line Lattice Generation (50 x real-time): Obtain phonetic lattice from utterance (50 x real-time) –Search (100,000 x real-time): Submit text based keyword queries, Obtain phonetic expansion, Find best match in phone lattice A Preliminary Study of Audio Indexing
DIVINES – Speech Rec. and Intrinsic Variation W.S.May 20, 2006 Richard Rose Evaluating Information Retrieval Performance Database – Twelve hours of lectures from McGill ECE Photonics Course (Prof. Andrew Kirk) Domain Experts – Course TA’s Target Domain – Example questions taken from course material … –Sample question: “Explain the modal properties of a conducting waveguide from the point of view of destructive and constructive interference” Relevance Labeling –Domain experts identify lecture segments that are relevant to question –A lecture segment is the audio that overlaps a given lecture slide
DIVINES – Speech Rec. and Intrinsic Variation W.S.May 20, 2006 Richard Rose Relevance Measure Given an audio segment of length seconds, For a Query containing query terms Obtain hypothesized occurrences for term with acoustic posterior scores Combine weighted posterior scores to obtain a measure of relevance for segment w.r.t. query Audio Segment k Acoustic Scores for Query Term i Hypothesized Occurrences of Term i
DIVINES – Speech Rec. and Intrinsic Variation W.S.May 20, 2006 Richard Rose Relevance Measure - Normalization There are two normalization components: –Acoustic Confidence Normalization: Function of the average Figure of Merit observed for query term FOM: Average of the detection prob. over a range of false alarm rates –Document Length Normalization: Estimate of the number of words in audio segment k Relies on estimate of speaking rate: words/sec. Relevance Measure:
DIVINES – Speech Rec. and Intrinsic Variation W.S.May 20, 2006 Richard Rose Acoustic Variability Impact of length of phonetic baseform on word detection performance Word duration in phones: Effect of word length in detection performance Prob. of Detection Baseform PhonesFOM (%) 5 or less or more73.03 Figure of Merit vs. Baseform Length: Figure of Merit (FOM): Average over the range form 0 to 10 false alarms per keyword per hour
DIVINES – Speech Rec. and Intrinsic Variation W.S.May 20, 2006 Richard Rose Acoustic Variability Impact of accuracy of phonetic baseforms on word spotting performance Word pronunciation: Comparison of 2 phonetic expansions of the word “ dielectric ” d iy l eh k t r ih k d ay l eh k t r ih k False Alarms per Keyword per Hour (FA/KW/HR ) Prob. of Detection
DIVINES – Speech Rec. and Intrinsic Variation W.S.May 20, 2006 Richard Rose IR Performance Define a relevance metric based on normalized frequency of occurrence of keywords chosen by domain experts Rank segments of messages based relevance metric Plot Results … Rank (R) % queries with at least one relevant document in top R ranks (text) % queries with at least one relevant document in top R ranks (speech) 575%58.33% %66.67% %75% %83.33%