Dave Toney, Korin Richmond Toney, Korin Richmond Dave David - - PowerPoint PPT Presentation

Dave Dave Toney, Korin Richmond Toney, Korin Richmond Edinburgh University Edinburgh University David Feinberg David Feinberg St Andrews University St Andrews University

travel/traffic info theatre/concerts restaurant weather directions accommodation

gender gender age age accent accent emotion emotion (location) (location) improved recognition improved recognition consumer preferences consumer preferences interface complexity interface complexity disposable income disposable income safety safety

S: S: Welcome to Welcome to the Glasgow automated tourist information line. the Glasgow automated tourist information line. How can I help you? How can I help you? U: U: Are there any museums nearby? Are there any museums nearby? S: S: The Gallery of Modern Art is closest to you. Admission is free. The Gallery of Modern Art is closest to you. Admission is free. U: U: Are there any plays on tonight? Are there any plays on tonight? S S: : Yes, Macbeth is showing at 6pm in Botanic Park. Yes, Macbeth is showing at 6pm in Botanic Park. You should probably get a taxi or use the underground. You should probably get a taxi or use the underground. U: U: OK, what s the weather like this evening? OK, what s the weather like this evening? S S: : Initially fine, with rain expected later. Initially fine, with rain expected later. U: U: Oh, right. Anything else on? Oh, right. Anything else on?

pitch, jitter shimmer harmonics-to-noise ratio

microphone mobile

ACOUSTI C FEATURE EXTRACTOR NEURAL NETWORK

SPEECH DATABASE

ACOUSTI C FEATURE EXTRACTOR NEURAL NETWORK

.wav gender, age

CTIMIT: CTIMIT: 3303 recordings, 621 speakers 3303 recordings, 621 speakers 429 male, 192 female 429 male, 192 female age 21 age 21 -

• 55

55 Acoustic Feature Extractor: Acoustic Feature Extractor: Praat Praat Neural Network: Neural Network: Netlab (Matlab) Netlab (Matlab)

Gender Gender -

• 94.4% accuracy

94.4% accuracy Age Age -

• no useful learning, possible reasons:

no useful learning, possible reasons: little variation little variation background noise background noise not enough data not enough data

Gender Gender -

• 94.4% accuracy

94.4% accuracy Age Age -

• no useful learning, possible reasons:

no useful learning, possible reasons: little variation little variation background noise background noise not enough data not enough data

MACROPHONE (land-line)

Conversational interfaces Conversational interfaces Personalizing interaction very worthwhile Personalizing interaction very worthwhile Acoustic cues to user profile Acoustic cues to user profile Integrated profiling module Integrated profiling module

gender gender age age accent accent emotion emotion height height improved recognition improved recognition consumer preferences consumer preferences interface complexity interface complexity disposable income disposable income safety safety

dave@cstr.ed.ac.uk