The Centre for Speech Technology Research, The university of Edinburgh
EUSTACE
HOME DESIGN EXAMPLES SENTENCES PROCEDURE
RESULTS DISSERTATION LICENCE DOWNLOAD GLOSSARY


Procedure: Utterance production labels


Each utterance was listened to in order to determine whether it had been produced correctly. This judgement involved the application of a number of criteria regarding pronunciation, placement of emphasis and the presence of audible boundaries within the utterance. Reasons for excluding utterances from further analysis and the labels used to indicate such utterances are described below. Utterance production labels, which all begin ``note.'', are placed immediately preceding the utterance identification label. (Note: some of the labels contain a numerical suffix. These were used in an early version of a label extraction script and were arbitrarily assigned. Where more than one version of label exists, all versions are listed below.)

Exclusion due to pronunciation

An utterance was labelled as excluded due to pronunciation if the speaker misread the words in the sentence, or pronounced the keyword differently from their other productions of the same keyword. In some cases, this label was applied during test syllable labelling, when it was determined that some acoustic feature of the keyword meant that segmentation criteria could not be reliably applied.

Labels indicating pronunciation errors:
note.discard-disfluency
note.discard-disfluency_20

Exclusion due to accent placement

An utterance was labelled as excluded due to accent placement if the keyword was perceived as unaccented in the accented condition or accented in the unaccented condition. Often this judgement was straightforward; where it was more difficult, the decision was based on the relative prominence of the keyword and other words in the utterance. On a few occasions, reference was made to the fundamental frequency contour of the utterance, but this often proved inconclusive.

If the keyword correctly carried or did not carry an accent, other misplacement of accent did not exclude the utterance from analysis: in the unaccented condition, where an accent was not placed on a word emphasised with block capitals but elsewhere in the utterance; in the accented condition, where there was an additional accent in the utterance to that on the keyword.

Labels indicating that the keyword has not received an accent in the accented condition, thus excluding the utterance from the original analysis:
note.nuclear_after
note.nuclear_after_13
note.nuclear_before
note.nuclear_before_12
note.key_non_nuclear_18
Labels indicating that the keyword has received an accent in the unaccented condition, thus excluding the utterance from the original analysis:
note.key_nuclear_6
note.key_prominence_7

Exclusion due to stress shift

Another source of accent misplacement was the shifting of primary stress in right-headed polysyllables. For example, the right-headed word reproduce should in this experiment have the primary stress on the final syllable, but subjects occasionally placed the primary stress - thus also the accent if present - on the initial syllable of the word. Utterances in which this occurred were excluded.

Labels indicating stress-shift in right-headed keywords:
note.initial-stress
note.initial_stress

Exclusion due to boundary placement

Utterance production labels were also applied regarding the occurrence of boundaries within the utterance. The judgement of the presence of a boundary was perceptual, and was made where a juncture was perceived corresponding to an intonational phrase boundary: the primary criterion was the presence of an audible boundary tone, such as a steep fall or a fall-rise; additional criteria were the presence of an audible pause or perceptible pre-boundary lengthening, or the existence of two complete intonational phrases within the utterance. Visual inspection of fundamental frequency contours, which were often discontinuous, was seldom used. Utterances were excluded if an intonational phrase boundary was perceived adjacent to the keyword, but were not excluded if the boundary was elsewhere in the utterance.

Labels indicating the presence of a boundary adjacent to the keyword:
note.post_boundary
note.post_boundary_2
note.pre_boundary_1

The ``pre_boundary'' label indicates that there is a boundary preceding the keyword; the ``post_boundary'' label indicates there is a boundary following the keyword.

Total excluded utterances

The number of utterances excluded from analysis according to each criterion was:

Utterances for which both repetitions are excluded represent missing data points. There are 72 such utterances, 3.1% of the total, listed in Tables D.1 - D.3 in Appendix D of English speech timing: a domain and locus approach (see DISSERTATION page).


Additional utterance labels

There are a number of other utterance labels, all preceded by ``note.'', which indicate some feature of the prosody of the utterance or the segmental realisation of the keyword. None of these labels were used to exclude an utterance from the original analysis.

Labels indicating the presence of an additional accent in the accented keyword condition:
note.extra_acc_after
note.extra_acc_after_15
note.extra_acc_before
note.extra_acc_before_14
Labels indicating the misplacement of the accent in the unaccented keyword condition, where this does not affect the keyword directly:
note.mis_accent_further
note.mis_accent_further_8
note.mis_accent_nearer_9
Labels indicating the presence of an utterance-internal boundary not adjacent to the keyword:
note.earlier_boundary_10
note.later_boundary
note.later_boundary_11
Labels used in some cases to indicate the type of evidence for an utterance-internal boundary:
note.boundary_tone_5
note.glot_boundary
note.glot_boundary_4
Miscellaneous labels relating to the segmental realisation of various keywords:
note./d/_released
note./d/_released-unvoiced
note.nasal-plosion
note.s/sh_very_undifferentiated
note.s_seg_probs
note.unstressed_syllables_cannot_be_measured


Up to PROCEDURE page Forward to TEST SYLLABLE LABELS page