Improving Automatic Forced Alignment for Dysarthric Speech Transcription
Yu Ting Yeung
2
, Ka Ho Wong
1
, Helen Meng
1,2
1
Human-Computer Communications Laboratory,
Department of Systems Engineering and Engineering Management
2
Stanley Ho Big Data Decision Analytics Research Centre
The Chinese University of Hong Kong, Hong Kong SAR, China
{ytyeung,khwong,hmmeng}@se.cuhk.edu.hk
Abstract
Dysarthria is a motor speech disorder due to neurologic deficits.
The impaired movement of muscles for speech production leads
to disordered speech where utterances have prolonged pause in-
tervals, slow speaking rates, poor articulation of phonemes, syl-
lable deletions, etc. These present challenges towards the use
of speech technologies for automatic processing of dysarthric
speech data. In order to address these challenges, this work be-
gins by addressing the performance degradation faced in forced
alignment. We perform initial alignments to locate long pauses
in dysarthric speech and make use of the pause intervals as an-
chor points. We apply speech recognition for word lattice out-
puts for recovering the time-stamps of the words in disordered
or incomplete pronunciations. By verifying the initial align-
ments with word lattices, we obtain the reliably aligned seg-
ments. These segments provide constraints for new alignment
grammars, that can improve alignment and transcription quality.
We have applied the proposed strategy to the TORGO corpus
and obtained improved alignments for most dysarthric speech
data, while maintaining good alignments for non-dysarthric
speech data.
Index Terms: automatic forced alignment, speech recognition,
dysarthric speech, word lattices
1. Introduction
Dysarthria is a motor speech disorder due to problems in the
nervous system [1]. The disorder is usually a consequence of
diseases such as stroke, cerebral palsy and amyotrophic lateral
sclerosis (ALS). These diseases also cause impaired movement
of the limbs, prohibiting the use of sign languages, computer
input by keyboard or pointing devices. The muscular impair-
ment also affects speech production which leads to disordered
speech. The impairment presents great difficulty to the subjects
in daily communication.
The recent advances of speech technologies bring a new
hope to dysarthric subjects on improving their quality of life,
such as using their remaining speech abilities to perform spoken
control of daily tasks [2–4], speech synthesis and voice conver-
sion for more lively and animated expression [5, 6], and com-
puter programs for speech rehabilitation [7]. Today’s speech
technologies are mostly data-driven and rely on speech corpora.
For dysarthric speech, several English dysarthric speech cor-
pora are publicly available, such as Nemours [8], the Universal
Access-Speech (UA-Speech) [9], and the TORGO corpus [10].
This project is partially sponsored by a grant from the HKSAR Govern-
ment General Research Fund (ref. no. 415513).
Forced alignment automatically generates time-stamps of
linguistic units (e.g. words, syllables, or phonemes) according
to the transcriptions of the recordings. The characteristics of
dysarthric speech pose challenges for automatic forced align-
ment. The automatic forced alignment algorithm should be ro-
bust to the prolonged pause intervals and slow speaking rates.
The algorithm should also be robust to the poor articulation
of phonemes, insertion and deletion of syllables and phones.
Moreover, the subjects may delete a word or restart an utter-
ance. The spoken content in the recordings may be different
from transcriptions.
Manual alignment of dysarthric speech data is a difficult
task due to poor speech intelligibility. Automatic forced align-
ment of dysarthric speech is thus preferable. We may ap-
ply a well-trained acoustic model (AM) from normal (i.e.,
non-dysarthric) speech data for automatic forced alignment of
dysarthric speech. Noticing the risk of speech data mismatch,
we must verify the quality of the alignments. Dysarthric speech
usually contain long pauses. We first perform initial alignment
to locate all the long pauses in the dysarthric utterances. Previ-
ous works [11–13] successfully applied large vocabulary con-
tinuous speech recognition (LVCSR) in aligning very long tran-
scriptions. We adopt the speech recognition approach for the
flexibility of inserting and deleting words in the alignments.
The word lattices from LVCSR system also provide us with
larger search spaces to locate the reliably aligned segments. The
reliably aligned segments provide the constraints for new align-
ment grammars that aim at improving the alignments.
This paper is organized as follows. In Section 2, we briefly
describe our speech data. We also discuss the challenges of
aligning dysarthric speech. In Section 3, we present our method
to obtain and update word alignments on dysarthric speech. In
Section 4, we discuss the alignment results on both dysarthric
and non-dysarthric speech. Finally in Section 5, we present our
conclusions and future directions.
2. Dysarthric speech
2.1. The corpus
We use the TORGO (LDC2012S02) corpus [10] for dysarthric
speech data. The corpus includes 8 dysarthric subjects (5 males
and 3 females) and 7 non-dysarthric speakers (4 male and 3 fe-
males). The corpus includes action tasks for articulation move-
ments, speaking tasks of repeating patterns, words, sentences
and picture descriptions. We only use speech data from single-
word and sentence speaking tasks as transcriptions are avail-
able. The data set consist of about 4,900 non-dysarthric and
2,400 dysarthric speech utterances, with at least 100 utterances
Appears in INTERSPEECH 2015 pp. 2991-2995
Updated version with typos corrected
Figure 1: A comparison of different alignments on dysarthric
speech (M01/Session2 3/0094.wav). The waveform and spec-
trogram of the utterance are included. (a) Manual alignment.
(b) Pause-aware initial alignment. (c) Conventional automatic
forced alignment without handling the pauses. A pause is repre-
sented as “sil”. <s>and </s> represent the sentence start and
end of the utterance respectively.
Table 1: The statistics of pause interval length in dysarthric and
non-dysarthric speech.
Sheet1
Non-dysarthric Dysarthric
Mean (ms)
74 383
25th percentile (ms)
30 80
Median (ms)
50 200
75th percentile (ms)
80 540
Page 1
for each speaker. The speech data are recorded with head-
mounted microphones. The TORGO corpus also includes Fren-
chay Dysarthria Assessment (FDA) [14] results of the subjects,
which provide references to the severity of dysarthria.
2.2. The challenges of aligning dysarthric speech
2.2.1. Prolonged pause intervals
The existence of long pauses is characteristic of dysarthric
speech. Table 1 shows the statistics of pause intervals of the
TORGO corpus from the pause-aware initial alignment stage
described in Section 3.1. For non-dysarthric speech, the pause
intervals are usually short. The length is consistent among dif-
ferent speakers. These pauses are commonly modeled as “short
pause” (“sp tee-model” in HTK toolkit [15]) in acoustic mod-
eling. The pause intervals of dysarthric utterances are signif-
icantly longer. The length of a pause is comparable to that
of a syllable. We need to take extra care on these pause in-
tervals during alignment. Figure 1 shows the waveform, spec-
trogram and alignments from different alignment methods of a
dysarthric utterance. The speaker pauses nearly at every word,
and even between two syllables within the word “noticed”. The
longest pause lasts for nearly 300 ms, leading to a slow speak-
ing rate of about 2 syllables per second. The pause-aware ini-
tial alignment (b) generally agrees with the manual alignment
(a). When we apply the conventional forced alignment setting
(“short pause” model and transcriptions without pause mark-
ings) as in (c), alignment errors are observed (between 1.3-2.4 s,
for the words “up” and “and”).
The pause intervals are as important as words in aligning
dysarthric speech. The pause intervals in dysarthric speech can
be modeled well by the pause model derived from the acous-
tic model of normal speech, due to similar acoustic properties.
Once a pause interval is located, it can serve as an anchor point
in the alignment. The start-time of a pause is the same as the
end-time of the previous word. The end-time of a pause is the
same as the start-time of the next word.
Initial alignment of words
and pauses with time-stamps
Utterance-
s
pecific LM
LVCSR General LM
LVCSR output with
time-stamps
Word-lattice
Acoustic model
Audio
Grammar-
b
ased speech
decoder
Grammar:
(
<s> A [sil] quick [sil] brown [sil]
fox [sil] jumps [sil] over [sil] the
[sil] lazy [sil] dog </s>)
Transcription
Figure 2: The flow diagram of the grammar-based speech de-
coder for generating initial alignments. The flow diagram also
includes the LVCSR system for generating word lattices and
speech recognition outputs. An optional pause is represented
by [sil] in the grammar.
2.2.2. Pronunciation deviations
Pronunciation deviations are common in dysarthric subjects.
The physical limitations of the subjects lead to poor articulation
of phonemes, insertion and deletion of syllables and phones.
The observed pronunciation deviation patterns of the TORGO
corpus are described in [4]. We decide to perform alignment
at word-level as the linguistic boundaries are better defined by
the pause intervals. The contextual information from language
models in LVCSR also helps to compensate for the acoustic
mismatch due to pronunciation deviations. We may perform
phone-level alignment at each word after segmenting a sentence
into words according to the word-level alignment.
3. Alignment of dysarthric speech
3.1. Stage 1: Pause-aware initial alignment
The first stage is to locate the long pauses in dysarthric ut-
terances. We recognize each utterance with a grammar-based
speech recognizer based on the HTK toolkit [15]. The grammar
is built upon the original word-level transcription, but optional
pauses are allowed to be inserted between words. We refer this
stage to as initial alignment. Figure 2 shows the corresponding
flow diagram and an example of the decoding grammar. We
prepare a speaker-independent acoustic model (AM) for the ini-
tial alignment task. The AM is trained with training data of
the TIMIT (LDC93S1) corpus [16] and represents the phonetic
characteristics of normal speech from a wide variety of speak-
ers. The monophone AM consists of 128 Gaussian components
at each acoustic state.
3.2. Stage 2: LVCSR output
We adopt the approach of using LVCSR outputs to verify the
initial alignments [11–13]. Figure 2 also shows the flow dia-
gram of the LVCSR system. We use the same 128-state TIMIT
monophone AM. We do not use a triphone AM, as speech
recognition performance is similar on dysarthric speech [4].
General word-based bigram and trigram language models
(LMs) are prepared from the entire original transcriptions of
the TORGO corpus. For each utterance, we further interpolate
the general bigram and trigram LMs with the text of the ini-
Verification
Updated alignment and
transcription
Acoustic model
Audio
Grammar-
based speech
decoder
Grammar for transcription editing:
(<s>Usually minus several sil
[several] buttons </s>)
Initial alignment of words and
pauses with time-stamps
LVCSR output with time-
stamps
Word
-lattice
Figure 3: The flow diagram of the alignment verification stage
and the grammar-based speech decoder for updating the align-
ments. An example of update grammar is given, where an op-
tional word is braced by brackets [].
tial alignment for utterance-specific LMs. We include the pause
markings into the utterance-specific LMs to model the pause in-
tervals in the utterance. The interpolation weight is set to 0.9,
to bias toward the text of the initial alignment.
A two-pass LVCSR is developed with the HTK toolkit [15].
For each utterance, word lattice is generated with the utterance-
specific bigram LM. The word lattice is then re-scored by the
utterance-specific trigram LM for 1-Best LVCSR output. The
1-Best output accompanies with time-stamps of the recognized
text. When the text of the 1-Best output (pauses ignored) is the
same as the original transcription, the original transcription is
considered as reliable and the 1-Best output is the final align-
ment of the utterance.
3.3. Stage 3: Alignment verification and update
The 1-Best LVCSR output may be different from the original
transcription. We have to identify whether the difference is due
to speech recognition errors or imperfect original transcription.
Speech recognition errors of dysarthric speech are commonly
caused by pronunciation deviations of the subjects. Subjects
may also pause between syllables of a word. A longer word may
be recognized as several shorter words. We thus process the 1-
Best output by grouping the recognized text into segments. The
time-stamps of the segments are aligned with the time-stamps
of the words and the pauses in the initial alignment. We com-
pute the Levenshtein distance between the segments and the text
(including the pauses) of the initial alignment for the mismatch
patterns, i.e., insertion, deletion and substitution. We weight
equally on all types of mismatch patterns. We also add an ad-
ditional constraint for pattern matching. The matched patterns
should also match with either start-time or end-time.
A word lattice contains alternative paths in addition to the
1-Best output. We hope that a substituted word can be recov-
ered from the word lattice. When we discover an arc of the sub-
stituted word with the same end-time as the initial alignment,
we consider the word as reliable. We match only the end-time
to allow potential word insertion in the updated alignment. A
grammar for updating the alignment is generated according to
the rules described in the following. The rules make use of the
mismatched patterns, word reliability, and the time-stamps of
the substituted and recognized words. We allow a maximum of
200 ms tolerance for time-stamp matching between the initial
alignments and the 1-Best outputs. The rules are:
1. If a substituted word is considered as reliable, the word
is always recovered in the updated alignment.
2. If the substituted word is reliable, we replace the rec-
ognized word with the substituted word in the updated
alignment under the following conditions: (1) the start-
time or the end-time between the substituted and the rec-
ognized words are matched, or (2) the start-time and the
end-time are matched with larger time tolerance.
3. Following Rule 2, when the substituted word is unreli-
able, the word may be replaced by the recognized word.
4. If the recognized word ends after the substituted word,
or vice versa, the updated alignment may include both
words. The word order follows the end-time of the
words. The recognized word is always optional in the
updated alignment.
5. An inserted word in the 1-Best output is optional in the
updated alignment.
6. If a word in the initial alignment is deleted in the 1-Best
output, the word is optional in the updated alignment.
An extension is especially useful for handling the pronunciation
deviations of dysarthric speech. We may abolish Rule 3 and ex-
tend Rule 2 to replace the recognized word with the substituted
word as long as their time-stamps are matched. Finally, we re-
align the utterances with a grammar-based speech recognizer
according to the derived grammar.
4. Result analysis
4.1. Agreement of LVCSR outputs
We align the dysarthric speech data with the settings described
in Section 3. We define the agreement rate which is equivalent
to sentence correctness (1- sentence error rate) of the 1-Best
LVCSR outputs (pauses ignored) with the original transcrip-
tions as the reference. We expect that the agreement rate should
be reasonably high when aligning speech data in a carefully de-
signed corpus. We do not expect a 100% agreement rate. An
occasional mismatch between the spoken content and the tran-
scription can occur in a well-designed corpus.
A hyper-parameter of LVCSR speech decoder is the lan-
guage model (LM) weight. The LM weight controls the con-
tribution of acoustic similarity according to acoustic model and
contextual information from the utterance-specific LMs. We
have considered two LM weights. A lighter LM weight p = 15
is within the common range of LVCSR systems, and a heav-
ier LM weight p = 30. Table 2 shows the agreement rates of
individual speakers.
For non-dysarthric speech data, the 1-Best LVCSR outputs
generally agree with the original transcriptions regardless the
LM weights. The medians of the agreement rates are 91.89%
for p = 15 and 95.09% for p = 30. The over 90% agree-
ment rate suggests that the original transcriptions are generally
in good quality. The high agreement rates also suggest that the
acoustic mismatch between the TORGO and the TIMIT corpora
is negligible for non-dysarthric speech. The agreement rates are
also consistent among different non-dysarthric speakers.
The LVCSR agreement rates drop substantially in
dysarthric speech data. The medians of the agreement rates are
61.7% and 85.1% for p = 15 and 30 respectively. The situa-
tion is expected due to the distorted pronunciations. For lighter
LM weight p = 15, the agreement rates are coarsely related to
Table 2: 1-Best LVCSR output agreement rates of individual
speakers. The severity of the subjects are described in [4].
p=15 p=30 p=15 p=30
F01 Severe 63.16 87.72 FC01 91.89 95.27
F03 Moderate 80.91 88.18 FC02 96.27 96.69
F04 Mild 84.68 89.52 FC03 92.17 95.09
M01 Severe 53.53 79.35 MC01 93.05 95.96
M02 Severe 60.26 82.56 MC02 81.31 90.99
M03 Mild 90.12 91.60 MC03 88.54 89.17
M04 Severe 54.74 74.82 MC04 91.36 92.48
M05
Moderate-
to-severe
26.55 79.65
Dysarthric subjects
Non-dysarthric subjects
Agreement rate
(%)
Agreement rate
(%)
Speakers
Severity
Speakers
Table 3: Agreement rates between the updated alignments and
the original transcriptions of individual speakers.
(a) (b) (a) (b)
F01 92.11 94.74 FC01 98.65 98.65
F03 94.90 94.90 FC02 99.27 99.27
F04 94.35 95.16 FC03 97.29 97.49
M01 86.96 90.49 MC01 98.40 98.50
M02 92.01 93.56 MC02 94.14 94.14
M03 97.28 97.52 MC03 92.43 92.43
M04 85.71 87.91 MC04 97.28 97.28
M05 85.84 90.27
Agreement rate (%)
Non-dysarthric speakers
Speakers
Agreement rate (%)
Speakers
Dysarthric subjects
the severity of the subjects. The mild subjects achieve higher
agreement rates, although the rates are still lower than those of
non-dysarthric speakers. For the severe subjects, their agree-
ment rates are generally lower. M05 gets the lowest agreement
rate although the severity level is not the worst. This is probably
due to serious pronunciation deviations of the subject. A note in
the TORGO corpus states that the speech of M05 is intelligible
when the subject speaks slowly.
A heavier LM weight p = 30 increases the agreement
rates of the dysarthric subjects. The agreement rates become
more consistent among different dysarthric subjects and no
longer reflect their severity. The severe subjects tend to achieve
larger improvement on the agreement rates with the heavier LM
weight. The agreement rate of M05 is no longer the lowest. The
pronunciation deviations of the subjects are compensated by the
strong contextual information from the utterance-specific LMs.
4.2. Alignment agreement after verification and update
We proceed to the verification stage with the LVCSR outputs
based on LM weight p = 30 due to the higher and more con-
sistent LVCSR agreement rates. We compare two different set-
tings for the alignment verification process: (a) the basic set-
ting described in Section 3.3, and (b) the extension in which the
recognized words are replaced by the substituted words when
their time-stamps are matched. Table 3 shows the agreement
rates between the updated alignments and the original transcrip-
tions. The agreement rates of non-dysarthric speakers are close
to 100%. The results from different settings are similar. For
dysarthric subjects, the medians of the agreement rates increase
to 92.4% and 94.2% for setting (a) and (b) respectively.
Figure 4 shows the example outputs of three utterances.
The first utterance is from a non-dysarthric speaker. The
speaker paused shortly and restarted the sentence. The 1-Best
usually minus several buttonssil
usually minus several sil several buttons
usually minus several several minus
Initial
alignment:
1-Best LVCSR:
Edited
alignment:
Grammar: usually( minus several sil [several] buttons )
usually minus several several buttonsGround truth:
(a) FC01/Session 1/0084.wav
grandfather likes to hissil
(c) F01/Session 1/0130.wav
languageinbe modern
grandfather and in languageinmoderna
grandfather to languagebe modern
sil
)
likes sil hisin sil
grandfather to languagebe modernlikes sil hisin sil
grandfather to languagebe modernlikes sil hisin sil(
Initial
alignment:
1-Best LVCSR:
Edited
alignment:
Grammar:
Ground truth:
i sil up twoand
( )
(b) M01/Session2_3/0094.wav
oldsil sil
looked
Initial
alignment:
1-Best LVCSR:
Edited
alignment:
Grammar:
Ground truth:
noticed sil men
i sil
two
twoand old silandtwo men
i sil(looked | two)
old
sil
i sil up twoand oldsil sillooked noticed sil men
up twoand oldsil sil[noticed] sil men
i sil up twoand oldsil siltwo noticed sil men
Figure 4: Examples of ground truths, initial alignments, LVCSR
outputs and updated alignments. The length and the location
of the boxes illustrate the time-alignment of the words in the
utterances. In the initial alignments, reliable words found in
word lattices are shaded. In the grammars, an optional word is
braced by brackets [], and a choice is represented by ( | ).
LVCSR output detects the restart at the word “several” although
the word “buttons” is wrongly recognized. The verification
step recovers the word “buttons”. The grammar also hypoth-
esizes the inserted word “several”. The updated alignment now
matches with the ground truth. In the second utterance, we re-
visit the example of Figure 1. The utterance is from a subject
with severe dysarthria. Pronunciation derivation is observed
at the word “looked”. The word is recognized as “two”. The
first half of the 1-Best LVCSR output is erroneous, but the re-
liable segments of the initial alignment are still identified from
the word lattice. After the alignment verification, most part of
the new alignment matches with the ground truth, but the word
“looked” is not recovered under setting (a). The word can be
recovered under setting (b), due to the matched time-stamps be-
tween “looked” and “two”. The third utterance is spoken by an-
other severe subject. Although there are some recognition errors
in the 1-Best LVCSR output, the word lattice reveals that the
initial alignment is reliable. The initial alignment is accepted as
the final alignment output. As shown in the examples, a pause
is a reliable feature in dysarthric speech. The pauses are always
included in the grammar to update the alignments.
5. Conclusions
We have developed a strategy for automatic forced alignment
on dysarthric speech data with an acoustic model trained with
normal speech data. We have applied the strategy to align the
TORGO corpus and verified a set of reliable alignments of
dysarthric speech data. These alignments
1
should be of good
quality to support research in dysarthric speech. We continue
to improve the alignment strategy to process the remaining
TORGO speech data and other dysarthric speech corpora.
1
Alignments available: https://github.com/ytyeung/IS2015alignment
6. References
[1] J. Duffy, Motor Speech Disorders: Substrates, Differential Diag-
nosis, and Management, 3rd ed. Elsevier Health Sciences, 2013.
[2] M. Parker, S. Cunningham, P. Enderby, M. Hawley, and
P. Green, “Automatic speech recognition and training for severely
dysarthric users of assistive technology: The stardust project,”
Clinical Linguistics & Phonetics, vol. 20, no. 2-3, pp. 149–156,
2006.
[3] M. Hasegawa-Johnson, J. Gunderson, A. Penman, and T. Huang,
“HMM-based and SVM-based recognition of the speech of talkers
with spastic dysarthria,” in Acoustics, Speech and Signal Process-
ing (ICASSP), 2006. IEEE International Conference on, 2006, pp.
1060–1063.
[4] K. Mengistu and F. Rudzicz, “Adapting acoustic and lexical mod-
els to dysarthric speech,” in Acoustics, Speech and Signal Process-
ing (ICASSP), 2011. IEEE International Conference on, 2011, pp.
4924–4927.
[5] A. Kain, X. Niu, J.-P. Hosom, Q. Miao, and J. P. v. Santen, “For-
mant re-synthesis of dysarthric speech,” in Fifth ISCA Workshop
on Speech Synthesis, 2004, pp. 25–30.
[6] J.-P. Hosom, A. Kain, T. Mishra, J. van Santen, M. Fried-Oken,
and J. Staehely, “Intelligibility of modifications to dysarthric
speech,” in Acoustics, Speech and Signal Processing (ICASSP),
2003. IEEE International Conference on, 2003, pp. 924–927.
[7] A. Hatzis, P. Green, J. Carmichael, S. Cunningham, R. Palmer,
M. Parker, , and P. O’Neill, “An integrated toolkit deploying
speech technology for computer based speech training with ap-
plication to dysarthric speakers,” in EUROSPEECH-2003, 2003,
pp. 2213–2216.
[8] X. Menendez-Pidal, J. Polikoff, S. Peters, J. Leonzio, and H. Bun-
nell, “The Nemours database of dysarthric speech,” in The Fourth
International Conference on Spoken Language Processing (IC-
SLP), vol. 3, Oct 1996, pp. 1962–1965 vol.3.
[9] H. Kim, M. Hasegawa-Johnson, A. Perlman, J. Gunderson, T. S.
Huang, K. Watkin, and S. Frame, “Dysarthric speech database
for universal access research,” in INTERSPEECH 2008, 2008, pp.
1741–1744.
[10] F. Rudzicz, A. Namasivayam, and T. Wolff, “The TORGO
database of acoustic and articulatory speech from speakers with
dysarthria,” Language Resources and Evaluation, vol. 46, no. 4,
pp. 523–541, 2012.
[11] P. J. Moreno, C. F. Joerg, J.-M. Van Thong, and O. Glickman,
“A recursive algorithm for the forced alignment of very long au-
dio segments,” in The Sixth International Conference on Spoken
Language Processing (ICSLP), 1998, pp. 2711–2714.
[12] T. J. Hazen, “Automatic alignment and error correction of hu-
man generated transcripts for long speech recordings.” in INTER-
SPEECH 2006, 2006, pp. 1606–1609.
[13] P. Moreno and C. Alberti, “A factor automaton approach for the
forced alignment of long speech recordings,” in Acoustics, Speech
and Signal Processing (ICASSP), 2009. IEEE International Con-
ference on, 2009, pp. 4869–4872.
[14] P. M. Enderby, Frenchay dysarthria assessment. San Diego, Cal-
ifornia: College-Hill Press, 1983.
[15] S. Young, J. Odell, D. Ollason, V. Valtchev, and P. Woodland, The
HTK Book. Cambridge University, 1995.
[16] Garofolo, John, et al., “TIMIT acoustic-phonetic continuous
speech corpus LDC93S1,” Web download. Philadelphia: Linguis-
tic Data Consortium, 1993.
![Figure 3: The flow diagram of the alignment verification stage and the grammar-based speech decoder for updating the alignments. An example of update grammar is given, where an optional word is braced by brackets [].](/figures/figure-3-the-flow-diagram-of-the-alignment-verification-2627faes.webp)
![Figure 2: The flow diagram of the grammar-based speech decoder for generating initial alignments. The flow diagram also includes the LVCSR system for generating word lattices and speech recognition outputs. An optional pause is represented by [sil] in the grammar.](/figures/figure-2-the-flow-diagram-of-the-grammar-based-speech-2ib9iq2x.webp)
![Table 2: 1-Best LVCSR output agreement rates of individual speakers. The severity of the subjects are described in [4].](/figures/table-2-1-best-lvcsr-output-agreement-rates-of-individual-3rpzbbmw.webp)
