Summarization of Spontaneous Speech using Automatic Speech

Recognition and a Speech Prosody based Tokenizer

Gy¨orgy Szasz´ak

, M´at´e

Akos T¨undik

and Andr´as Beke

Dept. of Telecommunications and Media Informatics, Budapest University of Technology and Economics,

2 Magyar tud´osok krt., 1117 Budapest, Hungary

Dept. of Phonetics, Research Institute for Linguistics of the Hungarian Academy of Sciences,

33 Bencz´ur utca, 1068 Budapest, Hungary

Keywords:

Audio, Speech, Summarization, Tokenization, Speech Recognition, Latent Semantic Indexing.

Abstract:

This paper addresses speech summarization of highly spontaneous speech. The audio signal is transcribed us-

ing an Automatic Speech Recognizer, which operates at relatively high word error rates due to the complexity

of the recognition task and high spontaneity of speech. An analysis is carried out to assess the propagation

of speech recognition errors into syntactic parsing. We also propose an automatic, speech prosody based au-

dio tokenization approach and compare it to human performance. The so obtained sentence-like tokens are

analysed by the syntactic parser to help ranking based on thematic terms and sentence position. The thematic

term is expressed in two ways: TF-IDF and Latent Semantic Indexing. The sentence scores are calculated as

a linear combination of the thematic term score and a positional score. The summary is generated from the

top 10 candidates. Results show that prosody based tokenization reaches human average performance and that

speech recognition errors propagate moderately into syntactic parsing (POS tagging and dependency parsing).

Nouns prove to be quite error resistant. Audio summarization shows 0.62 recall and 0.79 precision by an

F-measure of 0.68, compared to human reference. A subjective test is also carried out on a Likert-scale. All

results apply to spontaneous Hungarian.

1 INTRODUCTION

Speech is a vocalized form of the language which is

the most natural and effective method of communica-

tion between human beings. Speech can be processed

automatically in several application domains, includ-

ing speech recognition, speech-to-speech translation,

speech synthesis, spoken term detection, speech sum-

marization etc. These application areas use success-

fully automatic methods to extract or transform the

information carried by the speech signal. However,

the most often formal, or at least standard speaking

styles are supported and required by these applica-

tions. The treatment of spontaneous speech consti-

tutes a big challenge in spoken language technology,

because it violates standards and assumptions valid

for formal speaking style or written language and

hence constitutes a much more complex challenge in

terms of modelling and processing algorithms.

Automatic summarization is used to extract the

most relevant information from various sources: text

or speech. Speech is often transcribed and summa-

rization is carried out on text, but the automatically

transcribed text contains several linguistically incor-

rect words or structures resulting both from the spon-

taneity of speech and/or speech recognition errors. To

sum up, spontaneous speech is “ill-formed” and very

different from written text: it is characterized by dis-

ﬂuencies, ﬁlled pauses, repetitions, repairs and frag-

mented words, but behind this variable acoustic prop-

erty, syntax can also deviate from standard.

Another challenge originates in the automatic

speech recognition step. Speech recognition errors

propagate further into the text-based analysis phase.

Whereas word error rates in spoken dictation can be

as low as some percents, the recognition of sponta-

neous speech is a hard task due to the extreme variable

acoustics (including environmental noise, especially

overlapping speech) and poor coverage by the lan-

guage model and resulting high perplexities (Szarvas

et al., 2000). To overcome these difﬁculties, often

lattices or confusion networks are used instead of 1-

best ASR hypotheses (Hakkani-T¨ur et al., 2006). In

the current paper we are also interested in the as-

Szaszák, G., Tündik, M. and Beke, A.

Summarization of Spontaneous Speech using Automatic Speech Recognition and a Speech Prosody based Tokenizer.

DOI: 10.5220/0006044802210227

In Proceedings of the 8th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2016) - Volume 1: KDIR, pages 221-227

ISBN: 978-989-758-203-5

221

sessment of speech recognition error propagation into

text based syntactic parsing – POS-tagging and de-

pendency analysis. Since POS-tagging and especially

nouns play an important role in summarization, the

primary interest is to see how these are affected by

speech recognition errors.

A possible approach of summarizing written text

is to extract important sentences from a document

based on keywords or cue phrases. Automatic sen-

tence segmentation (tokenization) is crucial before

such a sentence based extractive summarization (Liu

and Xie, 2008). The difﬁculty comes not only from

incomplete structure (often identifying a sentence is

already problematic) and recognition errors, but also

from missing punctuation marks, which would be

fundamental for syntactic parsing and POS-tagging.

Speech prosody is known to help in speech segmenta-

tion and speaker or topic segmentation tasks (Shriberg

et al., 2000). In current work we propose and eval-

uate a prosody based automatic tokenizer which re-

covers intonational phrases (IP) and use these as sen-

tence like units in further analysis. Summarization

will also be compared to a baseline version using to-

kens available from human annotation. The baseline

tokenization relies on acoustic (silence) and syntactic-

semantic interpretation by the human annotators.

In the easier approach, speech summarization is

made equivalent to summarizing text transcripts of

speech, i.e., no speech-related features are used after

the speech-to-text conversion took place. However,

the transcribed speech can be used to summarize the

text (Christensen et al., 2004). Textual features ap-

plied for analysis include the position of the sentence

within the whole text, similarity to the overall story,

the number of named entities, semantic similarity be-

tween nouns estimated by using WordNets (Gurevych

and Strube, 2004) etc.

Other research showed that using speech-related

features beside textual-based features can improve

the performance of summarization (Maskey and

Hirschberg, 2005). Prosodic features such as speak-

ing rate; minimuma, maximuma, mean, and slope of

fundamental frequency and those of energy and ut-

terance duration can also be exploited. Some ap-

proaches prepare the summary directly from speech,

relying on speech samples taken from the spoken doc-

ument (Maskey and Hirschberg, 2006).

This paper is organized as follows: ﬁrst the auto-

matic speech recognizer is presented with analysis of

error propagation into subsequent processing phases.

Thereafter the prosody based tokenization approach

is presented. Following sections describe the summa-

rization approach. Finally, results are presented and

discussed, and conclusions are drawn.

2 AUTOMATIC SPEECH

RECOGNITION AND ERROR

PROPAGATION

Traditional Automatic Speech Recognition (ASR)

systems only deal with speech-to-text transformation,

however, it becomes more important to analyse the

speech content and extract the meaning of the given

utterance. The related research ﬁeld is called Spo-

ken Language Understanding. Nowadays, an increas-

ing proportion of the available data are audio record-

ings, so it becomes more emphasized to ﬁnd proper

solutions in this ﬁeld. There are two possibilities to

analyse the speech content; information can be recov-

ered directly from the speech signal or using parsers

after the speech-to-text conversion (automatic speech

recognition). This latter option also entails the poten-

tial to use analytic tools, parsers available for written

language. However, as mentioned in the introduction,

a crucial problem can be error propagation – of word

insertions, deletions and confusions – from the speech

recognition phase.

Our ﬁrst interest is to get a picture about ASR er-

ror propagationinto subsequent analysis steps. We se-

lect 535 sentences randomly from a Hungarian televi-

sion broadcast news speech corpus, and perform ASR

with a real-time close captioning system described

in (Tarj´an et al., 2014). ASR in our case yields close

to 35% Word Error Rate (WER). Please note that the

system presented in (Tarj´an et al., 2014) is much more

accurate (best WER=10.5%) , we deliberately choose

an operating point with higher WER to approximate a

behaviour for spontaneousspeech. Punctuation marks

are restored manually for this experiment.

Part-Of-Speech (POS) tagging and syntactic

parsing is then performed with the magyarl

anc

toolkit (Zsibrita et al., 2013) both on ASR output and

the reference text (containing the correct transcrip-

tions). Both outputs are transformed into a separate

vector space model (bag of words). Error propagation

is estimated based on similarity measures between the

two models, with or without respect to word order. In

the latter model, the dogs hate cats and cats hate dogs

short documents would behave identically. Although

meaning is obviously different, both forms are rele-

vant for queries containing dog or cat. In our case, not

word, but POS and dependency tags are compared, by

forming uni- and bi-grams of them sequentially.

We use cosine similarity to compare the parses

(for N dimensional vectors a and b):

KDIR 2016 - 8th International Conference on Knowledge Discovery and Information Retrieval

222

sim

cos

(a,b) =

∑

i=1

∑

i=1

∑

i=1

(1)

The cosine similarity of POS-unigrams and POS-

bigrams was found 0.90 and 0.77 respectively for un-

igrams and bigrams formed from bag-of-words for

POS tags of the ASR output and true transcripton,

which are quite high despite the ASR errors. Like-

wise, the result for the dependency labels shows high

correspondence between parses on ASR output and

true transcription; cosine similarity is 0.89, the stan-

dard Labeled Attachment Score and Unlabeled At-

tachment Score metrics (Green, 2011) (these are de-

ﬁned only for sentences with same length) gave 80%

and 87% accuracy. Moreover,the root of the sentence,

which is usually the verb, hits 90% accuracy.

Taking the true transcription as reference we also

calculate POS-tag error rates analogously to word er-

ror rate (instead of words we regard POS-tags). POS-

tag error rate is found to be 22% by 35% WER.

Checking this selectively for nouns only, POS-tag er-

ror rate is even lower: 12%.

These results tell us good news about ASR error

propagation into subsequent analysis steps: POS-tags

and especially nouns are less sensitive to ASR er-

rors. Checking further for correlation between WER

and similarity between parses of ASR output and

true transcription, we can conﬁrm that this is close

to linear (see Figure 1). This means that by higher

WER of the ASR, similarity between true transcrip-

tion and ASR output based parses can be expected to

degrade linearly proportional to WER (see (T¨undik

and Szasz´ak, 2016) for a similar experiment in Hun-

garian), i.e. there is no breakdown point in WER

beyond which parsing performance would drop more

drastically than does the WER.

2.1 ASR for Summarization

For the summarization experiments, we use different

ASR acoustic models (tuned for spontaneous speech)

trained on 160 interviews from BEA (Hungarian lan-

guage), accounting for 120 hours of speech (the in-

terviewer discarded) with the Kaldi toolkit. 3 hid-

den layer DNN acoustic models are trained with 2048

neurons per layer and tanh non-linearity. Input data is

9x spliced MFCC13 + CMVN +LDA/MLLT. A tri-

gram language model is also trained on transcripts

of the 160 interviews after text normalization, with

Kneser-Ney smoothing. Dictionaries are obtained us-

ing a rule-based phonetizer (spoken Hungarian is very

Figure 1: Cosine similarity for POS-tags between true and

ASR transcriptions depending on WER. A linear regression

line is also plotted.

close to the written form and hence, a rule based

phonetizer is available).

Word Error Rate (WER) was found around 44%

for the spontaneous ASR task. This relative high

WER is justiﬁed by the high spontaneity of speech.

3 SPEECH PROSODY BASED

TOKENIZATION

A speech segmentation tool which recovers auto-

matically phonological phrases (PP) was presented

in (Szasz´ak and Beke, 2012). A PP is deﬁned as

a prosodic unit of speech, characterized by a sin-

gle stress and corresponds often to a group of words

belonging syntactically together. The speech seg-

mentation system is based on Hidden Markov Mod-

els (HMM), which model each possible type (7 alto-

gether) of PPs based on input features such as fun-

damental frequency of speech (F0) and speech sig-

nal energy. In segmentation mode, a Viterbi align-

ment is performed to recover the most likely under-

lying PP structure. The PP alignment is conceived in

such a manner that it encodes upper level intonational

phrase (IP) constraints (as IP starter and ending PPs,

as well as silence are modelled separately), and hence

is de facto capable of yielding an IP segmentation,

capturing silence, but also silence markers (often not

physically realized as real silence, but giving a per-

ceptually equivalent impression of a short silence in

human listener). The algorithm is described in detail

in (Szasz´ak and Beke, 2012), in this work we use it to

obtain sentence-like tokenswhich are then fed into the

ASR. We use this IP tokenizer in an operating point

with high precision ( 96% on read speech) and lower

recall ( 80% on read speech) as we consider less prob-

lematic missing a token boundary(merge 2 sentences)

Summarization of Spontaneous Speech using Automatic Speech Recognition and a Speech Prosody based Tokenizer

223

than inserting false ones (splitting the sentence into 2

parts).

The performance of prosody based tokenization is

compared to tokenization obtained from human anno-

tation. Results are presented in section 5.

4 THE SUMMARIZATION

APPROACH

After the tokenization for sentence-like intonational

phrase units and automatic speech recognition took

place, text summarization is split into three main

modules. The ﬁrst module preprocesses the out-

put of the ASR, the second module is responsible

for sentence ranking, and the ﬁnal module generates

the summary. This summarization approach is based

on (Sarkar, 2012), but we modify the thematic term

calculation method. The overall scheme of the sys-

tem is depicted in Fig. 2.

Input speech

ASR

Preprocessing

Feature extraction

Stop word

removal

Stemming

processing

POS

tagging

TF-IDF

LSI

Sentence

position

Sentence ranking

Summary

generation

Output summary

Tokenization

Figure 2: Block scheme of the speech summarization sys-

tem.

4.1 Pre-processing

Stop words are removed from the tokens and stem-

ming is performed. Stop-words are collected into a

list, which contains (i) all words tagged as ﬁllers by

the ASR (speaker noise) and (ii) a predeﬁned set of

non-content words such as articles, conjunctions etc.

Hungarian is a highly agglutinating language,

with a very rich morphology,and consequently,gram-

matical relations are expressed less by the word order

but rather by case endings (sufﬁxes). The magyarl

anc

toolkit (Zsibrita et al., 2013) was used for the stem-

ming and POS-tagging of the Hungarian text. Stem-

ming is very important and often ambiguous due to

the mentioned rich morphology. Thereafter, the POS-

tagger module was applied to determine a word as

corresponding to a part-of-speech. The words are ﬁl-

tered to keep only nouns, which are considered to be

the most relevant in summarization.

4.2 Textual Feature Extraction

In order to rank sentences based on their importance,

some textual features are extracted:

4.2.1 TF-IDF

(Term Frequency - Inverse Document Frequency) re-

ﬂects the importance of a sentence and is generally

measured by the number of keywords present in it.

TF-IDF is a useful heuristic for ranking the words ac-

cording to their importance in the text. The impor-

tance value of a sentence is computed as the sum of

TF-IDF values of its constituent words (in this work:

nouns) divided by the sum of all TF-IDF values found

in the text. TF shows how frequently a term occurs in

a document divided by the length of the document,

whereas IDF shows how important a term is. In raw

word frequency each word is equally important, but,

of course, not all equally frequent words are equally

meaningful. For this reason it can be calculated using

the following equation:

IDF(t) = ln

C(all documents)

C(documents containing term t)

(2)

where C(.) is the counting operator. TF-IDF weight-

ing is the combination of the deﬁnitions of term fre-

quency and inverse document frequency, to produce

a composite weight for each term in each document,

calculated as a dot product:

TF-IDF = TF ∗IDF. (3)

4.2.2 Latent Semantic Analysis

Latent Semantic Analysis (LSA) exploits context to

try to identify words which have similar meaning.

LSA is able to reﬂect both word and sentence impor-

tance. Singular Value Decomposition (SVD) is used

to assess semantic similarity.

LSA based summarization needs the calculation

of the following data (Landauer et al., 1998):

• Represent the input data in the form of a matrix,

where columns contain sentences and rows con-

tain words. In each cell, a measure reﬂecting the

importance of the given word in the given sen-

tence is stored. This matrix is often called input

matrix (A).

• Use SVD to capture relationships among words

and sentences. In order to do this, the input matrix

KDIR 2016 - 8th International Conference on Knowledge Discovery and Information Retrieval

224

is decomposed into 3 constituents (sub-matrices):

A = UΣVT, (4)

where A is the input matrix U represents the de-

scription of the original rows of the input matrix

as a vector of extracted concepts, Σ is a diago-

nal matrix containing scaling values sorted in de-

scending order, and V represents the description

of the original columns of input matrix as a vector

of the extracted concepts (Landauer et al., 1998).

• The ﬁnal step is the sentence selection for the

summary.

4.2.3 Positional Value

The a priori assumption that the more meaningful sen-

tences can be found at the beginning of the document

is generally true. This is even more the case for spon-

taneous narratives, which are the target of our summa-

rization experiments, as the interviewer usually asks

the participants to tell something about their life, job,

hobbies. People tend to answer with keywords, then

they go into details with those keywords recalled. For

the calculation of the positional value, the following

equitation was used (Sarkar, 2012):

= 1/

√

k (5)

where the P

is the positional score of k

sentence.

4.2.4 Sentence Length

Sentences are of different length (they contain more

or less words) in documents. Usually a short sen-

tence is less informative than a longer one and hence,

readers or listeners are more prone to select a longer

sentence than a short one when asked to ﬁnd good

summarizing sentences in documents (Sarkar, 2012).

However, a too long sentence may contain redun-

dant information. The idea is then to eliminate or

de-weight sentences which are too short or too long

(compared to an average). If a sentence is too short or

too long, it is assigned a ranking score of 0.

4.2.5 Sentence Ranking

The ranking score RS

is calculated as the linear com-

bination of the so-called thematic term based score S

and positional score P

. The ﬁnal score of a sentence

k is:

(

αS

+ βP

, ifL

≥ L

& L

≤ L

0 otherwise,

(6)

where α is the lower, β is the upper cut-off for the

sentence position (0 ≤ α,β ≤ 1) and L

is the lower

and L

is the upper cut-off on the sentence length

(Sarkar, 2012).

4.3 Summary Generation

The last step is to generate the summary itself. In this

process, the N-top ranked sentences are selected from

the text. In current work N = 10, so the ﬁnal text

summary contains the top 10 sentences.

5 EXPERIMENTS

For the summarization experiments, we use 4 inter-

views from the BEA Hungarian Spontaneous Speech

database (Neuberger et al., 2014), all excluded form

the ASR’s acoustic and language model training. Par-

ticipants talk about their jobs, family, and hobbies.

Three of the speakers are male and one of them is fe-

male. All speakers are native Hungarian, living in Bu-

dapest (aged between 30 and 60). The total material

is 28 minutes long (average duration was 7 minutes

per participant) and is highly spontaneous.

5.1 Metrics

The most commonly used information retrieval eval-

uation metrics are precision (PRC) and recall (RCL),

which are appropriate to measure summarization per-

formance as well (Nenkova, 2006). Beside recall and

precision, we use the F

-measure:

2∗PRC∗RCL

PRC+ RCL

(7)

The challenge of evaluation consists rather in

choosing or obtaining a reference summary. For this

research we decided to obtain a set of human made

summaries, whereby 10 participants were asked to se-

lect up to 10 sentences that they ﬁnd to be the most

informative for a given document (presented also in

spoken and in written form). Participants used 6.8

sentences on average for their summaries. For each

narrative, a set of reference sentences was created:

sentences chosen by at least 1/3 of the participants

were added to the reference summary. Overlap among

human preferred sentences was ranging from 10% to

100%, with an average overlap of 28%. Sentences

are appended to the summaries in the order of their

appearance in the original text. When using this ref-

erence set for the evaluation of the automatic sum-

maries, we ﬁlter stop words, ﬁllers (ASR output) from

the latter and require at least a 2/3 overlap ratio for

the content words. We will refer to this evaluation

approach as soft comparison.

An automatic evaluation tool is also considered to

obtain more objective measures. The most commonly

used automatic evaluation method is ROUGE (Lin,

Summarization of Spontaneous Speech using Automatic Speech Recognition and a Speech Prosody based Tokenizer

225

2004). However, ROUGE performs strict string com-

parison and hence recall and precision are commonly

lower with this approach (Nenkova, 2006). We will

refer to this evaluation approach as hard comparison.

5.2 Results

Text summarization was then run with 3 setups re-

garding pre-processing (how the text was obtained

and tokenized):

• OT-H: Use the original transcribed text as seg-

mented by the human annotators into sentence-

like units.

• ASR-H: Use the human annotated tokens and

ASR to obtain the text.

• ASR-IP: Tokenize the input based on IP bound-

ary detection from speech and use ASR transcrip-

tions.

Summary generation is tested for all the 3 setups with

both TF-IDF and LSA approaches to calculate the

thematic term S

in Equation (6). Results are shown

in Table 1 for soft comparison, and Table 2 for hard

comparison.

Table 1: Recall (RCL), precision (PRC) and F

– soft com-

parison.

Soft comparison

Setup Method RCL [%] PRC [%] F

OT-H

TF-IDF 0.51 0.76 0.61

LSA 0.36 0.71 0.46

ASR-H

TF-IDF 0.51 0.80 0.61

LSA 0.49 0.77 0.56

ASR-IP

TF-IDF 0.62 0.79 0.68

LSA 0.59 0.78 0.65

Table 2: Recall (RCL), precision (PRC) and F

– hard com-

parison.

Hard comparison (ROUGE)

Setup Method RCL [%] PRC [%] F

OT-H

TF-IDF 0.36 0.28 0.32

LSA 0.36 0.30 0.32

ASR-H

TF-IDF 0.34 0.29 0.31

LSA 0.39 0.27 0.32

ASR-IP

TF-IDF 0.33 0.28 0.30

LSA 0.33 0.32 0.32

Overall results are in accordance with published

values for similar tasks (Campr and Jeˇzek, 2015).

When switching to the ASR transcription, there is

no signiﬁcant difference in performance regarding the

soft comparison, but we notice a decrease (rel. 8%)

in the hard one (comparing OT-H and ASR-IP ap-

proaches). This is due to ASR errors, however, keep-

ing in mind the high WER for spontaneous speech,

this decrease is rather modest. Indeed, it seems that

content words and stems are less vulnerable to ASR

errors, which is in accordance with our ﬁndings pre-

sented in Section 2.

An important outcome of the experiments is that

the automatic, IP detection based prosodic tokeniza-

tion gave almost the same performance as the human

annotation based one (in soft comparison it is even

better). We believe that these good results with IP to-

kenization are obtained thanks to the better and more

infrmation drivenstructurization of speech parts when

relying on prosody (acoustics).

5.3 Subjective Assessment of

Summaries

In a separate evaluation step, volunteers were asked to

evaluate the system generated summaries on a Likert

scale. Thereby they got the system generated sum-

mary as is and had to rate it according to the question

“How well does the system summarize the narrative

content in your opinion?”. The Likert scale was rang-

ing from 1 to 5: “Poor, Moderate, Acceptable, Good,

Excellent”. Results of the evaluation are shown in

Fig. 3. Mean Opinion score was 3.2. Regarding re-

dundancy (“How redundant is the summary in your

opinion?”) MOS value was found to be 2.8.

Excellent

Good

Moderate

Acceptable

Poor

Mean

Error bars: +/- 1 SD

Figure 3: Likert scale distribution of human judgements.

6 CONCLUSIONS

This paper addressed speech summarization for

highly spontaneous Hungarian. The ASR transcrip-

tion and tokenization are sensitive and not yet solved

steps in audio summarization. Therefore the authors

consider that the proposed IP detection based tok-

enization is an important contribution, especially as

KDIR 2016 - 8th International Conference on Knowledge Discovery and Information Retrieval

226

it proved to be as successful as the available human

one when embedded into speech summarization. An-

other basic contribution comes from the estimation of

the ASR transcription error propagation into subse-

quent text processing, at least in terms of evaluating

the similarity of POS and dependency tag sequences

between human and ASR made transcriptions. Re-

sults showed that POS tags and selectively nouns are

less sensitive to ASR errors (POS tag error rate was

2/3 of WER, whereas nouns get confused by another

part-of-speech even less frequently). Given the high

degree of spontaneity of the speech and also the heavy

agglutinating property of Hungarian, we believe the

obtained results are promising as they are comparable

to results published for other languages. The over-

all best results were 62% recall and 79% precision

= 0.68). Subjective rating of the summaries gave

3.2 mean opinion score.

ACKNOWLEDGEMENTS

The authors would like to thank the support of the

Hungarian National Innovation Ofﬁce (NKFIH) un-

der contract IDs PD-112598 and PD-108762.

REFERENCES

Campr, M. and Jeˇzek, K. (2015). Comparing semantic mod-

els for evaluating automatic document summarization.

In Text, Speech, and Dialogue, pages 252–260.

Christensen, H., Kolluru, B., Gotoh, Y., and Renals, S.

(2004). From text summarisation to style-speciﬁc

summarisation for broadcast news. In Advances in In-

formation Retrieval, pages 223–237. Springer.

Green, N. (2011). Dependency parsing. In Proceedings

of the 20th Annual Conference of Doctoral Students:

Part I - Mathematics and Computer Sciences, pages

137–142.

Gurevych, I. and Strube, M. (2004). Semantic similarity ap-

plied to spoken dialogue summarization. In Proceed-

ings of the 20th international conference on Compu-

tational Linguistics, page 764.

Hakkani-T¨ur, D., Bechet, F., Riccardi, G., and T¨ur, G.

(2006). Beyond asr 1-best: using word confusion net-

works in spoken language understanding. Computer

Speech and Language, 20(4):495–514.

Landauer, T. K., Foltz, P. W., and Laham, D. (1998). An in-

troduction to latent semantic analysis. Discourse pro-

cesses, 25(2-3):259–284.

Lin, C.-Y. (2004). Rouge: A package for automatic evalu-

ation of summaries. In Text summarization branches

out: Proc. of the ACL-04 workshop, volume 8.

Liu, Y. and Xie, S. (2008). Impact of automatic sentence

segmentation on meeting summarization. In Proc.

Acoustics, Speech and Signal Processing, ICASSP

2008. IEEE International Conference on, pages 5009–

5012.

Maskey, S. and Hirschberg, J. (2005). Comparing lex-

ical, acoustic/prosodic, structural and discourse fea-

tures for speech summarization. In INTERSPEECH,

pages 621–624.

Maskey, S. and Hirschberg, J. (2006). Summarizing speech

without text using hidden markov models. In Pro-

ceedings of the Human Language Technology Confer-

ence of the NAACL, Companion Volume: Short Pa-

pers, pages 89–92.

Nenkova, A. (2006). Summarization evaluation for text and

speech: issues and approaches. In INTERSPEECH,

pages 1527–1530.

Neuberger, T., Gyarmathy, D., Gr´aczi, T. E., Horv´ath, V.,

G´osy, M., and Beke, A. (2014). Development of

a large spontaneous speech database of agglutinative

hungarian language. In Text, Speech and Dialogue,

pages 424–431.

Sarkar, K. (2012). Bengali text summarization by sen-

tence extraction. In Proc. of International Conference

on Business and Information Management ICBIM12,

pages 233–245.

Shriberg, E., Stolcke, A., Hakkani-T¨ur, D., and T¨ur, G.

(2000). Prosody-based automatic segmentation of

speech into sentences and topics. Speech Communi-

cation, 32(1):127–154.

Szarvas, M., Fegy´o, T., Mihajlik, P., and Tatai, P. (2000).

Automatic recognition of Hungarian: Theory and

practice. Int. Journal of Speech Technology, 3(3):237–

251.

Szasz´ak, G. and Beke, A. (2012). Exploiting prosody

for automatic syntactic phrase boundary detection in

speech. Journal of Language Modeling, 0(1):143–

172.

Tarj´an, B., Fegy´o, T., and Mihajlik, P. (2014). A bilingual

study on the prediction of morph-based improvement.

In Proceedings of the 4th International Workshop on

Spoken Languages Technologies for Under-Resourced

Languages, pages 131–138.

T¨undik, M. A. and Szasz´ak, G. (2016). Sz¨ovegalap´u nyelvi

elemz¨o ki´ert´ekel´ese g´epi besz´edfelismer¨o hib´akkal

terhelt kimenet´en. In Proc. 12th Hungarian Confer-

ence on Computational Linguistics (MSZNY), pages

111–120.

Zsibrita, J., Vincze, V., and Farkas, R. (2013). magyarlanc:

A toolkit for morphological and dependency parsing

of hungarian. In Proceedings of RANLP, pages 763–

771.

Summarization of Spontaneous Speech using Automatic Speech Recognition and a Speech Prosody based Tokenizer

227