Searching and Summarizing Speech Julia Hirschberg CS 6998 7/15/2016 1 Today Speech browsing and search Speech summarization: 2 views Hori et al Barzilay et al Speech data mining 7/15/2016 2 Searching Audio Data Today, large amounts of audio data available: on the web, in company archives, in our homes But what can we do with it? We have tools supporting random access to text – but for audio we’re limited to serial search Goal: tools to search audio as easily as text 7/15/2016 3 Why? Searching online news and archives Searching a/v archives, movies Searching trial recordings and legislative sessions Browsing meetings, customer care exchanges, focus groups Telephone calls and voicemail 7/15/2016 4 Audio Browsing/Retrieval for Voicemail Motivated by interviews, surveys and usage logs of heavy users: Hard to scan new msgs to find those you need to deal with quickly Hard to find msg you want in archive Hard to locate information you want in any msg How could we help? 7/15/2016 5 Caller SCANMail Architecture SCANMail Subscriber Corpus Collection Recordings collected from 138 AT&T Labs employees’ mailboxes 100 hours; 10K msgs; 2500 speakers Gender balanced: 12% non-native speakers Mean message duration 36.4 secs, median 30.0 secs Hand-transcribed and annotated with caller id, gender, age, entity demarcation (names, dates, telnos) 7/15/2016 7 Transcription and Bracketing [ Greeting: hi R ] [ CallerID: it's me ] give me a call [ um ] right away cos there's [ .hn ] I guess there's some [ .hn ] change [ Date: tomorrow ] with the nursery school and they [ um ] [ .hn ] anyway they had this idea [ cos ] since I think J's the only one staying [ Date: tomorrow ] for play club so they wanted to they suggested that [ .hn ] well J2 actually offered to take J home with her and then would she 7/15/2016 8 would meet you back at the synagogue at [ Time: five thirty ] to pick her up [ .hn ] [ uh ] so I don't know how you feel about that otherwise M_ and one other teacher would stay and take care of her till [ Date: five thirty tomorrow ] but if you [ .hn ] I wanted to know how you feel before I tell her one way or the other so call me [ .hn ] right away cos I have to get back to her in about an hour so [ .hn ] okay [ Closing: bye [ .nhn ] [ .onhk ] 7/15/2016 9 SCANMail Demo http://www.fancentral.org/~isen hour/scanmail/demo.html Audix extension: 8380 Audix password: (null) 7/15/2016 10 Information Extraction from Speech Jansche & Abney ‘02 7/15/2016 11 Speech Summarization: Extraction Techniques Hori et al ‘02 Inoue et al ‘04 7/15/2016 12 Domain Specific Summarization (Barzilay et al ‘00) Motivation: lab experiments show little facilitation of speech summarization by techniques that do improve search Domain: Broadcast News Idea: knowing what type of speaker (anchor, reporter, interviewee) is speaking provides structural clues that can “outline” the newscast since programs are predictable 7/15/2016 13 SCAN: Spoken Content-based Audio Navigator TREC SDR corpus of Broadcast News Segment speech `documents’ into audio `paratones’ acoustically Segmentation module trained on handlabeled discourse structure annotation in another domain Classify recording conditions, e.g. Music, telephone bandwidth, wide-band Run ASR with appropriate acoustic models (~70% wac) Index (errorful) transcripts using SMART IR 7/15/2016 14 Results in WYSIAWY (“What you see is almost what you hear”) GUI Transcript prosodically formatted Overview provides abstract structure 7/15/2016 15 Acoustic Condition Classification Paratone Detector Broadcast News corpus Recognition SCAN db Information Retrieval GUI 7/15/2016 16 Search Overview Transcript 7/15/2016 17 Patterns in Newscasts Anchors present headlines and introduce stories Most frequent speakers Anchor/reporter turn alternation Reporter/guest turntaking during stories 7/15/2016 18 Data 35 broadcasts of “All Things Considered” Human and ASR transcripts (without commercials but with turn boundaries) Features to predict speaker role Lexical: ngrams 1-5, explicit introductions (current and prior segment) Contextual: labels and features of prior turns Durational: turn length (absolute and relative to previous) 7/15/2016 19 Methods and Results Boosting and maximum entropy --> simple weighted rules to predict speaker role Baseline: guess anchor (35.4%) Result on human transcripts: BoostTexter 79% MaxEnt 80.5% Result on ASR transcripts: BoostTexter 72.8% MaxEnt 77% 7/15/2016 20 Speech Data Mining How does it differ from text data mining? Maskey et al ‘04 7/15/2016 21