multiple level articulatory segmental HMMs

1
ABI speech corpus
Shona DArcy, Martin Russell Electronic,
Electrical and Computer Engineering University of
Birmingham
2
Motivation

• Wide variation in pronunciation of English in the
  British Isles
• Perceived as problem for speech technologies
• No systematic study published to date
• Need for corpus of speech representing different
  accents

3
Variety of Accents

• British isles
• Northern England
• Southern England
• Wales
• Scotland
• Northern Ireland
• Southern Ireland

4
Corpus design

• What is an accent?
• What accents are required for corpus?
• List of 13 accents to be recorded
• Standard Southern english

5
London
6
London
West County
7
London
West County
8
Ireland
London
West County
9
East Anglia
Ireland
London
West County
10
East Anglia
Ireland
Birmingham
London
West County
11
Liverpool
East Anglia
Ireland
Birmingham
London
West County
12
N Ireland
Liverpool
East Anglia
Ireland
Birmingham
London
West County
13
N Ireland
Lancashire
Liverpool
East Anglia
Ireland
Birmingham
London
West County
14
N Ireland
Yorkshire
Lancashire
Liverpool
East Anglia
Ireland
Birmingham
London
West County
15
N Ireland
Newcastle
Yorkshire
Lancashire
Liverpool
East Anglia
Ireland
Birmingham
London
West County
16
Glasgow
N Ireland
Newcastle
Yorkshire
Lancashire
Liverpool
East Anglia
Ireland
Birmingham
London
West County
17
Scottish Highlands
Glasgow
N Ireland
Newcastle
Yorkshire
Lancashire
Liverpool
East Anglia
Ireland
Birmingham
London
West County
18
Corpus design

• Who are good subjects?
• People who had lived in the area most of their
  lives
• People whos parents had lived there most of
  their lives
• Between the ages of 18 and 50
• 10 male and 10 female

19
Text to be recorded

• Accent specific texts
• when a sailor in a small craft faces the might
  of the vast Atlantic ocean today
• Scribe sentences
• Where were you while we were away
• Vowels in control contexts,
• Had, whod, Hudd (to rhyme with bud), hid, heard

20
Texts to recorded

• Standard prompts for speech technologies
• Digits triples
• Letters, phonetic alphabet
• GYP, Golf, Yankee, Polo
• Equipment specific commands
• Climate control 71 degrees

21
(No Transcript)
22
Preparation

• Recording software
• List of prompts
• Venue selection
• Chose town
• Library/community centre

23
Hardware

• Laptop with external soundcard, (Edirol)
• Emkay head mounted microphone
• Generic desk mounted microphone

24
Strategy for subjects

• Press releases
• Local media, radio (25)
• TV(7) newspapers (40)
• Interviews
• Free phone
• Subject selection
• Subject criteria
• Based on audio

25
Statistics

• 14 locations
• 20 people per location, between 16 and 70
• 95 hours of recordings
• Annotated at phrase level

26
Lessons learnt

• Identifying accents
• Choosing a town likely to have appropriate accent
  and produce enough candidates
• Choice of recording location

27
Rural settings

• Association of accent with place not reliable
• Expectation of situation
• Less migration, hence less diluting of accents
• Smaller variations over large area
• Findings
• Young people more likely to adopt other accents
• Mainly older people had good accent

28
Urban settings

• Most cities homogeneous accent
• Liverpool, Glasgow
• Huge variety of accents in London

29
Choice of location

• Library
• Relatively consistent noise level
• Library users available to recruit
• Other (e.g. community centre, universities)
• More variation in noise
• Older people

30
Press release free-phone

• Mostly successful (would do it again)
• Coverage approximately equal in each location
• Variable response
• When successful
• Many volunteers with good accents within our age
  range
• When unsuccessful
• Not enough volunteers
• Had to recruit from the street

31
Future recordings

• Fill in holes existing in current database
• Incorporating lessons learnt
• Record accents not covered in this corpus