TIMIT Annotation Overview

• Corpus contains 6300 sentences; 54,387 words
• Our regular expression, unfiltered, produced 3154 tokens for consideration
• With filters, 2059 tokens
• Of these, 1578 were annotated for -t/d deletion (others were cases of N/A)

Overall Deletion Rate
Total number of tokens deleted: 518 (32.8%)
Total number of tokens retained: 1060 (67.2%)

Unique tokens in TIMIT
The TIMIT Corpus consists of 630 speakers reading a list of 10 phonetically-rich sentences (selected from a larger set).  Despite the use of read speech, there are few tokens that occur multiple times relative to the total number of tokens coded.
 

Appearances in corpus	Number/Percentage of tokens
1	35.4% (n=558)
2	3.6% (n=58)
6	4.2% (n=66)
7	50.6% (n=798)
14	1.7% (n=28)
All other numbers	4.4% (n=69)

Morphological Category

	Deletion Rate	Total N
Monomorphemes	36.6%	1081
Irregular verbs (all-inclusive)**	41.1%	107
Irregular verbs (excluding must)**	24.7%	73
Irregular verbs (excluding must, went & all strong verbs)**	15.4%	39
Regular Verbs	22.6%	513

**See coding scheme for details of treatment of irregular verbs

Preceding Environment

	Deletion Rate	Total N
Alveolar Nasal	52.8%	432
Alveolar Fricative	41.7%	391
Other Fricative	24.7%	73
Stop	23.4%	244
Lateral	16%	162
Other Nasal	16%	25
Rhotic	8.7%	252

Following Environment

	Deletion Rate	Total N
Obstruent	52.9%	607
Rhotic	48.2%	56
Clustering Glide	41.9%	105
Lateral	29.4%	17
Other Glide	21.4%	14
Pause	18.3%	252
Vowel	13.7%	257

Identical Preceding/Following Environment

	Deletion Rate	Total N
s_s (processed_soybeans)	88.9%	36
obstruent_obstruent (stopped_passing) *includes s_s environment	87.0%	46
liquid_liquid (guard_rail, old_lady)	0%	10
overall	71.4%	56

Education

Race/Ethnicity

Age

Sex

	Deletion Rate	Total N
Males	31.6%	1077
Females	35.5%	501

Region

	Deletion Rate	Total N
Southern	38.2%	254
New York	37.8%	119
Mixed (Army Brat)	36.8%	76
New England	35.4%	113
South Midland	33.2%	229
North Midland	32.4%	278
Northern	28.2%	262
Western	27.5%	247

 
 

Rate of Deletion before PAUSE by Geographical Region

VARBRUL Results

First Run

Eight Factor Groups; 44 Factors	Selected?
Rule Application	n/a
Morphology	yes
Preceding	yes
Following	yes
Sex	no
Region	no
Age	no
Race	yes
Education	yes

Second Run
-eliminated non-selected factor groups from first run (Gender, Education, Age)
-recodes

1) Factor Group Race: recode Asian/American Indian/Hispanic --> Other (15 tokens)

RESULTS
Summary Statistics and Factor Weights
 

Group	Factors		Retained (%)	Deleted (%)	Total N (%)	Factor Weight
Morphology	monomorpheme	m	630 (62)	394 (38)	1024 (65)	0.535
	reg. past tense	p	397 (77)	116 (23)	513 (33)	0.428
	irregular	i	33 (80)	8 (20)	41 (3)	0.531
Preceding Segment	lateral	l	135 (84)	26 (16)	161 (10)	0.240
	alveolar fricative	f	228 (58)	163 (42)	391 (25)	0.635
	stop	k	187 (77)	57 (23)	244 (15)	0.426
	rhotic	r	230 (91)	22 (9)	252 (16)	0.161
	alveolar nasal	n	204 (47)	228 (53)	432 (27)	0.756
	other fricative	c	55 (75)	18 (25)	73 (5)	0.433
	other nasal	s	21 (84)	4 (16)	25 (2)	0.390
Following Segment	rhotic	h	29 (52)	27 (48)	56 (4)	0.650
	vowel	v	455 (86)	72 (14)	527 (33)	0.245
	obstruent	b	286 (47)	321 (53)	607 (38)	0.767
	pause	q	206 (82)	46 (18)	252 (16)	0.305
	lateral	a	12 (71)	5 (29)	17 (1)	0.380
	cluster glide	g	61 (58)	44 (42)	105 (7)	0.645
	other glide	o	11 (79)	3 (21)	14 (1)	0.330
Group	Factors		Retained (%)	Deleted (%)	Total N (%)	Factor Weight
Race	white	W	991 (68)	464 (32)	1455 (92)	0.489
	black	L	33 (49)	34 (51)	67 (4)	0.753
	unknown	U	25 (61)	16 (39)	41 (3)	0.433
	other	O	11 (73)	4 (27)	15 (1)	0.552
Education	Bachelors	B	587 (67)	289 (33)	876 (56)	0.514
	High School	H	133 (64)	74 (36)	207 (13)	0.524
	Masters	T	250 (71)	100 (29)	350 (22)	0.436
	PhD	P	45 (78)	13 (22)	58 (4)	0.357
	Unknown	K	13 (42)	18 (58)	31 (2)	0.752
	Associates	A	32 (57)	24 (43)	56 (4)	0.616
TOTAL			1060 (67%)	518 (33%)	1578

 

Third Run
-split the irregular verb category into four new categories:

• semi-weak verbs: includes verbs that have a stem change AND t/d affixation in their past tense (keep > kept; lose > lost)
• strong verbs: includes verbs that only have a stem change in their past tense, no t/d affixation (hold > held; wind > wound)
• must: this is a modal verb, and has only one form in the past and present
• went: there is no transparent connection between this verb's present and past forms (go > went)

RESULTS
Summary Statistics and Factor Weights
 

Group	Factors		Retained (%)	Deleted (%)	Total N (%)	Factor Weight
Morphology	monomorpheme	m	611 (64)	347 (36)	958 (61)	0.521
	reg. past tense	p	392 (76)	121 (24)	513 (33)	0.436
	strong verb	s	18 (62)	11 (38)	29 (2)	0.476
	semi-weak verbs	i	26 (67)	13 (33)	39 (2)	0.553
	must	t	11 (32)	23 (68)	34 (2)	0.747
	went	w	2 (40)	3 (60)	5 (0)	0.837
Preceding Segment	lateral	l	135 (84)	26 (16)	161 (10)	0.244
	alveolar fricative	f	228 (58)	163 (42)	391 (25)	0.635
	stop	k	187 (77)	57 (23)	244 (15)	0.395
	rhotic	r	230 (91)	22 (9)	252 (16)	0.161
	alveolar nasal	n	204 (47)	228 (53)	432 (27)	0.768
	other fricative	c	55 (75)	18 (25)	73 (5)	0.436
	other nasal	s	21 (84)	4 (16)	25 (2)	0.383
Following Segment	rhotic	h	29 (52)	27 (48)	56 (4)	0.649
	vowel	v	455 (86)	72 (14)	527 (33)	0.248
	obstruent	b	286 (47)	321 (53)	607 (38)	0.759
	pause	q	206 (82)	46 (18)	252 (16)	0.313
	lateral	a	12 (71)	5 (29)	17 (1)	0.372
	cluster glide	g	61 (58)	44 (42)	105 (7)	0.655
	other glide	o	11 (79)	3 (21)	14 (1)	0.354
Group	Factors		Retained (%)	Deleted (%)	Total N (%)	Factor Weight
Race	white	W	991 (68)	464 (32)	1455 (92)	0.489
	black	L	33 (49)	34 (51)	67 (4)	0.751
	unknown	U	25 (61)	16 (39)	41 (3)	0.430
	other	O	11 (73)	4 (27)	15 (1)	0.556
Education	Bachelors	B	587 (67)	289 (33)	876 (56)	0.512
	High School	H	133 (64)	74 (36)	207 (13)	0.523
	Masters	T	250 (71)	100 (29)	350 (22)	0.439
	PhD	P	45 (78)	13 (22)	58 (4)	0.361
	Unknown	K	13 (42)	18 (58)	31 (2)	0.725
	Associates	A	32 (57)	24 (43)	56 (4)	0.622
TOTAL			1060 (67%)	518 (33%)	1578

Dual Annotation Results
Reannotation of 5% of TIMIT corpus is complete.  Details coming soon...

DASL Homepage