=Paper=
{{Paper
|id=Vol-1199/paper1
|storemode=property
|title=Language identification with limited resources
|pdfUrl=https://ceur-ws.org/Vol-1199/paper1.pdf
|volume=Vol-1199
|dblpUrl=https://dblp.org/rec/conf/timm/ArnalGH14
}}
==Language identification with limited resources==
https://ceur-ws.org/Vol-1199/paper1.pdf
Actas V Jornadas TIMM, pp 7-10
Language identification with limited resources
Emilio Sanchis Mayte Giménez Lluı́s-F. Hurtado
Departament de Sistemes Informàtics i Computació
Universitat Politècnica de València, València, Spain
{esanchis, mgimenez, lhurtado}@dsic.upv.es
Abstract
Language identification is an important issue in many speech applica-
tions. We address this problem from the point of view of classification
of sequences of phonemes, given the assumption that each language has
its own phonotactic characteristics. In order to achieve this classifica-
tion, we have to decode the speech utterances in terms of phonemes.
The set of phonemes must be the same for all the languages, because
the goal is to have a comparable representation of the acoustic se-
quences. We followed two different approaches using the same acoustic
model: we decode the audio using trigrams of sequences of phonemes
and equiprobable unigrams of phonemes as language model. Then a
classification process based on perplexity is performed.
1 Introduction
Language identification (LI) is an important application in multilingual speech environments. This is the case of
multilingual dialog systems where the system has to detect the input language in order to choose the correspond-
ing models associated to each language. Given the interest of this field in speech technologies some evaluation
campaigns have been proposed, as the Albayzin evaluation in Spain [Rod13]. Some methodologies are used for
language identification, some of them directly based on acoustic representation of the signal, and others based on
phonetic representations [Pal13]. Our approach consist of a first process of Acoustic-Phonetic Decoding (APD),
considering the set of Spanish phoneme models, and a classification process of the sequences of phonemes based
on the distance to the different languages. An advantage of this approach is that it can be easily developed
when there are not many resources to learn accurate acoustic representation for each language. It is enough to
have a set universal phonemes, and a not labeled corpus of each language. We have applied this approach to
a multilingual version of the DIHANA corpus, that consist of dialogs for obtaining information about trains in
Spain. We present some experiments over English, French and Spanish.
2 Our language identification approach
Our proposal to LI is based on modeling sequences of phonetic units that characterize each language we want to
identify. The language identification process of a spoken utterance is divided into two phases:
• Acoustic-Phonetic Decoding. The first phase of the LI process is a phonetic transcription of the spoken
utterance which language must be identified. In our proposal, this phase is the same for all languages and,
therefore, it should be language independent.
Copyright c by the paper’s authors. Copying permitted only for private and academic purposes.
In: L. Alfonso Ureña López, Jose Antonio Troyano Jiménez, Francisco Javier Ortega Rodrı́guez, Eugenio Martı́nez Cámara (eds.):
Actas de las V Jornadas TIMM, Cazalla de la Sierra, España, 12-JUN-2014, publicadas en http://ceur-ws.org
7
Emilio Sanchis, Mayte Giménez y Lluís-F Hurtado
• Phonetic sequence classification. Once the spoken utterance is phonetically transcribed, this sequence must
be classified in order to determine the language of the utterance. A language model of sequences of phonetic
units is learned for each language. The selection criterion if based on minimize the perplexity.
Let L be the set of languages, li ∈ L one of this languages, and s the phonetic unit sequence to classify. The
selected language ˆl is the one that minimize the expression:
1
ˆl = argmin 10− |s| log p(s|li )
(1)
li ∈L
where, p(s|li ) is the probability of the sequence s assigned by the model representing language li .
3 Resources and Experimentation
This sections describes the resources used, how we learned the language models, and the preliminary experimen-
tation carried out in this work.
3.1 Description of the used corpus
We have used a corpus of 3446 spoken sentences to learn the language models and evaluate our proposal. The
sentences were uttered by several native English, French, and Spanish speakers. The distribution of the languages
in the corpus was a little unbalanced (1338 in English, 708 in French, and 1400 for Spanish). The domain of the
English and French sentences was queries to a information service about timetable and prices of long distance
trains. The Spanish sentences were extracted from a unrestricted phonetically balanced corpus.
3.2 Learning the models
As phonetic unit, we have chosen context-dependent phonemes. Specifically, we have used triphones, ie, phonemes
with information about the phonemes that appear to their left and right. We have learned the acoustic models
for triphones and the models of sequences of triphones using an independent Spanish corpus. Only triphones
for Spanish have been considered in this work. We have used the same set of Spanish triphones for all the
experimentation.
We have phonetically transcribed all sentences in the corpus using two different Acoustic-Phonetic Decoding
modules. In both modules the set of triphones and the acoustics models associated to them were the same;
the difference was the model of sequences of triphones used as language model. The first APD module used a
trigram model of sequences of triphones. To avoid the bias of using for all languages a trigram model of sequence
of phonetic units (triphones) learned with Spanish corpus, a second module was learned using an equiprobable
unigram model of triphones. This way, all sequences of phonetic units have the same a priori probability.
As result, we got six phonetically transcribed utterances sets, two for each considered language using our two
different APD modules.
3.3 Experimentation
In order to conduct the evaluation of our approach, we split the available corpus by language and use 80% for
training the classification models, leaving the remaining 20% to evaluate the performance of the system. Since
we have two possible different APD modules (trigrams and equiprobable unigrams), we were able to learn two
set of language models. For each set, we learned an trigram language model for every language we are trying to
discriminate.
We used SRILM Toolkit [Sto02] to estimated the phonetic language models of the classifiers and HTK Speech
Recognition Toolkit [You06] to perform the phonetic transcriptions.
Two different experiments were conducted. The first experiment consisted of measuring the perplexity of
the test sets. Table 1 shows the perplexity for all training and test combinations. Each column corresponds
to the test set for a different language and using an specific APD module (Trigrams APD for the APD based
on trigrams of phonetic units and Equiprobable APD for the APD based on equiprobable unigrams of phonetic
units). In addition, each row corresponds to a classifier learned using the transcriptions of the training sentences
of an specific language using an specific APD module.
As expected, Table 1 shows a lower perplexity for combinations where the language of the classifier and the
language of test are the same. Regarding the APD module, lower perplexity occur when an APD based on
8
Language identification with limited resources
Table 1: Perplexity of the phonetic language models
Test set
Trigrams APD Equiprobable APD
French English Spanish French English Spanish
French 8.24 11.62 12.16 27.94 33.07 19.86
Trigrams APD English 10.79 6.63 11.29 40.78 18.86 18.76
Spanish 11.27 10.86 7.57 59.43 39.22 13.98
French 12.06 14.89 17.07 15.64 19.17 19.05
Equiprobable APD English 14.41 8.79 15.13 21.19 10.57 17.43
Spanish 11.57 10.97 8.43 28.53 21.42 10.98
trigrams is used to transcribe the sentences, specially those in the test set. It seems that, the use of trigrams of
phonetic units learned using a corpus only Spanish is not as critic as we a priori expected.
A second experimentation was conducted in order to evaluate the performance of the Language Identification
system. The global accuracy of the system was 0.841 when Trigram APD module was used and 0.775 when
Equiprobable APD module was used. As in the case of perplexity, the best accuracy result is obtained using the
Trigram APD module. Table 2 shows the accuracy considering the different languages involved. The best results
are obtained for Spanish, possibly because the triphones used were just those of Spanish. Although the phonetic
similarity between Spanish and French seems bigger than the phonetic similarity between Spanish and English,
results for English are better than those obtained for French. This may be due to the greater amount of English
sentences available for the experimentation.
Table 2: Accuracy of the Language Identification system
French English Spanish
Trigrams APD 0.793 0.850 0.960
Equiprobable APD 0.771 0.857 0.928
4 Conclusions and future work
In this paper we have presented a preliminary approach to the language identification problem. Our proposal
is based on the classification of sequences of phonemes assuming that each language has its own phonotactic
characteristics. The experimentation shows that our approach is able to predict reasonably well the language
of the speaker, especially considering the limited resources used. We have many ideas on how to improve the
performance of our system, including but not limited to using really language-independent phonetic units, using
the recognizer lattices as input to the classification system.
Acknowledgements
This work is partially supported by the Spanish MICINN under contract TIN2011-28169-C05-01, Spain.
References
[Pal13] Palacios, C.S., D’Haro, L.F., de Córdoba, R., Caraballo, M.A.: Incorporación de n-gramas discrim-
inativos para mejorar un reconocedor de idioma fonotáctico basado en i-vectores. Procesamiento del
Lenguaje Natural 51 (2013) 145–152
[Rod13] Rodrı́guez-Fuentes, L.J., Brümmer, N., Peñagarikano, M., Varona, A., Bordel, G., Dı́ez, M.: The
albayzin 2012 language recognition evaluation. In Bimbot, F., Cerisara, C., Fougeron, C., Gravier, G.,
Lamel, L., Pellegrino, F., Perrier, P., eds.: Interspeech, ISCA (2013) 1497–1501
[Sto02] Stolcke, A.: Srilm - an extensible language modeling toolkit. In: Proc. of Intl. Conf. on Spoken
Language. (2002) 901–904
9
Emilio Sanchis, Mayte Giménez y Lluís-F Hurtado
[You06] Young, S.J., Evermann, G., Gales, M.J.F., Hain, T., Kershaw, D., Moore, G., Odell, J., Ollason,
D., Povey, D., Valtchev, V., Woodland, P.C.: The HTK Book, version 3.4. Cambridge University
Engineering Department, Cambridge, UK (2006)
10