Notes
Slide Show
Outline
1
Introduction to
Speech Corpora@Stanford
  • Florian Jaeger,
  • tiflo@stanford.edu


  • For LIN110,
  • May 10th, 2004
2
Before we get to the real stuff…
  • This presentation will be available online at:
    • http://www.stanford.edu/dept/linguistics/corpora/material/X_speech_corpora/
    • (that is: X_speech_corpora)


  • Local support
  • Where are our corpora?
  • Setting up your account on AFS
3
Local support
  • Where can you get help with your project?
    • Your TA
    • The Corpora@Stanford website (http://www.stanford.edu/dept/linguistics/corpora/)
    • The corpora@csli.stanford.edu email list (you have to subscribe first), cf.:
      • http://www.stanford.edu/dept/linguistics/corpora/cas-support.html
    • The corpus TA (tiflo@stanford.edu)
4
Where are our corpora? (1)
  • AFS:
    • AFS is Stanford’s file sharing system
    • The linguistic corpora are stored at:


    • /afs/ir/data/linguistic-data/


    • You need to register for AFS access:
      • http://www.stanford.edu/dept/linguistics/corpora/cas-rules.html#reg-standard
    • You need to set up your account:
      • http://www.stanford.edu/dept/linguistics/corpora/cas-setup.html
5
Where are our corpora? (2)
  • Corpus Computer
    • The computer is the one closest to the printer in the linguistics department’s computer cluster (MJH, 1st floor)
    • Login: ‘user’
    • Pwd:
  • The corpora are stored on partition D:\
  • Mapping the drive via a network:
    • http://www.stanford.edu/dept/linguistics/corpora/material/corpora-list/021204_FYI%20NEW%20local%20resource.txt
    • http://www.stanford.edu/dept/linguistics/corpora/material/corpora-list/021204_addendum%20to%20last%20message.txt
6
The real part
  • Example project


  • Overview of available corpora
  • Where to find them
  • How does the annotation look like?


  • How to search speech corpora
7
Example projects (1)
  • Differences in the realization of phonemes depending on their context
    • ‘Context’ can be segmental [1]
      • How does the realization of syllabic /m/ differ depending on the preceding onset?
      • Word final vowel aspiration
    • ‘Context’ can be supra-segmental: [3]
      • How does the realization of syllabic /m/ differ at the beginning/end of conversations/utterances/sentences?
      • Reduction of complex clusters
8
Example projects (2)
    • ‘Context’ could also include the register, style (formal vs. informal), genre (reading a fairy tale vs. reading an article), different dialects, etc. [2]
  • Pitch contours related to specific meanings [1]
    • Steady-state pitch contours
9
Available corpora
  • Cf. Colleen’s handout
  • See also:
    • http://www.stanford.edu/dept/linguistics/corpora/cas-corpora.html
10
 
11
Switchboard – spontaneous AE speech
    • Transcripts uploaded to AFS:
      • /afs/ir/data/linguistic-data/Switchboard/
    • Sound files available on CD


    • available in several formats:
      • All in one file
      • Separate files for
        • Syllables
        • Words
        • Orthographic transcription
12
Example annotations (Switchboard)
  • Some files in Switchboard
13
Switchboard – all in one file
Annotation key (1)
  • Key:


  • SENTENCE: word1 word2 ...  (2005_A_0041)
  • WORD: word canonical? [lm-probs] [rates] [positions] [morebigrams] part-of-speech phone1 phone2 ...
  • SYL: baseform transcribed syl_structure stress length [lm-probs] [rates] [positions]
  • PHONE: baseform stress syl_part [lm-probs] [rates] [positions] tran1 tran2 ...
14
Switchboard – all in one file
Annotation key (2)
  • [lm-probs]= trigram unigram trigram-unigram
  • [rates]= seg_tr_syl seg_tr_phn lex_syl lex_phn enrate vrate nvrate mrate mfrate enmmfrate mmfrate
  • [positions] = word_num_in_utterance word_num_in_turn
  • [morebigrams] = bigram reverse-bigram reverse-trigram center-trigram
  • part-of-speech = syntactic part of speech (currently only done for the word "to")
  • wordX= word number X in acoustically segmented `sentence'
  • canonical?= can if canonical (pronlex) pronunciation, alt otherwise
  • trigram= p(word | previous two words)
  • unigram= p(word)
  • trigram-unigram = difference between two probabilities
  • seg_tr_syl= transcribed syllable rate between closest two pauses
  • seg_tr_phn= transcribed phone rate between closest two pauses
  • lex_syl= lexical syllabic rate (i.e. as determined from wd transcription)
  • lex_phn= lexical phone rate (i.e. as determined from wd transcription)
15
Switchboard – all in one file
Annotation key (3)
  • enrate= old enrate measure
  • vrate= voicing rate
  • nvrate= another voicing rate
  • mrate= sub-part of mrate measure
  • mfrate= sub-part of mrate measure
  • enmmfrate= *this is what we call mrate* average of enrate, mrate, mfrate
  • mmffrate= average of mrate, mfrate
  • baseform= pronunciation as written in dictionary
  • transcribed= transcribed syllable
  • syl_structure= onset/nucleus/coda markings from dictionary
  • stress= syllable stress marking from dictionary P=primary S=secondary N=none
  • length= syllable length
  • tranX= transcribed phone X corresponding to baseform phone
16
Example annotations
(Switchboard – all in one file)
  • SENTENCE: like finding a proper nursing home (2005_A_0041)
  • WORD: like 1 can -2.408 -2.152 -0.256 4.64 10.43 3.87 9.89 3.80 2.32 5.79 2.32 4.64 3.59 3.48 0 26 l ay k
  • SYL: l_ay_k l_ay_k O_N_C P 0.258 -2.408 -2.152 -0.256 4.64 10.43 3.87 9.89 3.80 2.32 5.79 2.32 4.64 3.59 3.48 0 26
  • PHONE: l P O -2.408 -2.152 -0.256 4.64 10.43 3.87 9.89 3.80 2.32 5.79 2.32 4.64 3.59 3.48 0 26 l
  • PHONE: ay P N -2.408 -2.152 -0.256 4.64 10.43 3.87 9.89 3.80 2.32 5.79 2.32 4.64 3.59 3.48 0 26 ay
  • PHONE: k P C -2.408 -2.152 -0.256 4.64 10.43 3.87 9.89 3.80 2.32 5.79 2.32 4.64 3.59 3.48 0 26 k
  • WORD: finding 2 alt -3.604 -4.256 0.652 4.64 10.43 3.87 9.89 3.80 2.32 5.79 2.32 4.64 3.59 3.48 1 27 f ay n ih ng
  • SYL: f_ay_n f_ay_n O_N_C P 0.358 -3.604 -4.256 0.652 4.64 10.43 3.87 9.89 3.80 2.32 5.79 2.32 4.64 3.59 3.48 1 27
  • PHONE: f P O -3.604 -4.256 0.652 4.64 10.43 3.87 9.89 3.80 2.32 5.79 2.32 4.64 3.59 3.48 1 27 f
  • PHONE: ay P N -3.604 -4.256 0.652 4.64 10.43 3.87 9.89 3.80 2.32 5.79 2.32 4.64 3.59 3.48 1 27 ay
  • PHONE: n P C -3.604 -4.256 0.652 4.64 10.43 3.87 9.89 3.80 2.32 5.79 2.32 4.64 3.59 3.48 1 27 n
  • SYL: d_ih_ng NULL_ih_ng O_N_C N 0.117 -3.604 -4.256 0.652 4.64 10.43 3.87 9.89 3.80 2.32 5.79 2.32 4.64 3.59 3.48 1 27
  • PHONE: d N O -3.604 -4.256 0.652 4.64 10.43 3.87 9.89 3.80 2.32 5.79 2.32 4.64 3.59 3.48 NULL 1 27
  • PHONE: ih N N -3.604 -4.256 0.652 4.64 10.43 3.87 9.89 3.80 2.32 5.79 2.32 4.64 3.59 3.48 1 27 ih
  • PHONE: ng N C -3.604 -4.256 0.652 4.64 10.43 3.87 9.89 3.80 2.32 5.79 2.32 4.64 3.59 3.48 1 27 ng
17
Boston Radio Transcripts
  • Includes read news etc. (i.e. non-spontaneous read speech)


  • Transcripts uploaded to AFS at:
    • /afs/ir/data/linguistic-data/Boston-University-Radio
  • Sound files available on CD
18
Example annotations (Boston Radio)
  • Boston News Corpus
    • H#   0 4
    • >endsil
    • DH   4 5
    • IH+1   9 10
    • S 19 9
    • >This
    • HH 28 5
    • AA+1 33 9
    • L 42 12
    • AX 54 4
    • DCL 58 3
    • D 61 1
    • EY 62 16
    • >holiday
    • S 78 11
    • IY+1 89 14
    • Z 103 7
    • EN 110 20
    • …
19
Example annotations (Boston Radio)
  • XWAVES/PRAAT readable:
    • signal st43/f3ast43p1
    • type 1
    • color 76
    • font -*-times-medium-r-*-*-17-*-*-*-*-*-*-*
    • separator ;
    • nfields 1
    • #
    •     0.035000   76 H#
    •     0.085000   76 DH
    •     0.185000   76 IH+1
    •     0.275000   76 S
    •     0.325000   76 HH
    •     0.415000   76 AA+1
    •     0.535000   76 L
    •     0.575000   76 AX
    •     0.605000   76 DCL
    •     0.615000   76 D
    •     0.775000   76 EY
    •     0.885000   76 S
    • …
20
CALLHOME Mandarin - Transcripts
  • CALLHOME – Mandarin
    • Transcripts uploaded to AFS:
      • /afs/ir/data/linguistic-data/CALLHOME/CALLHOME-Mandarin-Transcripts/
    • Lexicon with pronunciation information available at:
      • /afs/ir/data/linguistic-data/CALLHOME/CALLHOME-Mandarin-Lexicon/
    • Sound files only available on CD/DVD, but I could put them on the corpus computer
21
Example annotation (CALLHOME)
  • 150.41 150.82 B: ßÀ ºß ,
  • 154.57 161.32 B: °¥ , ÄǸö ßÀ Äã ˵ Äã ˵ Ëû ÓÐûÓÐ &¿­Àñ& µÄ µØÖ· °¡ ? ÎÒ Ôõô [channel_noise] ÕÒ ÁË °ëÌì ÄØ Ã» ÕÒ ×Å Ëû µÄ µØÖ· °¡ ?
  • 161.45 162.11 A: ÊÇÂð ?
  • 162.24 162.70 B: ßÀ ,
  • 162.93 163.88 A: %Ŷ% ,
  • 163.96 164.41 A: [background_speech_((°¥_Ó´))]
22
TIMIT – dialect variation
  • Telephone recording of 8 major dialects of American English
  • (orthographic) transcripts on AFS, sound files available on CD


  • Comparable dialect corpora exist for the British Isles (IViE; stored on the corpus computer)
23
Example annotations (TIMIT)
  • TIMIT
    • Word label (.wrd):
      •         7470 11362 she
      •         11362 16000 had
      •         15420 17503 your
      •         17503 23360 dark
      •         23360 28360 suit
      •         28360 30960 in
      •         30960 36971 greasy
    • Phonetic label (.phn):
      • (Note: beginning and ending silence regions are marked with h#)
      •         0 7470 h#
      •         7470 9840 sh
      •         9840 11362 iy
      •         11362 12908 hv
      •         12908 14760 ae
      •         14760 15420 dcl
      •         15420 16000 jh
      •         16000 17503 axr
24
How to search transcribed corpora?
  • Either load the files into your favorite text editor
  • Or use a command from the ‘grep’ family (run on a UNIX shell)
    • This allows you to search many files as once for patterns that are described by regular expressions
    • For help, see our tutorial page at:
      • http://www.stanford.edu/dept/linguistics/corpora/cas-tut-grep.html