•
Dialect identification using Gaussian Mixture Models
Pedro A. Torres-Carrasquillo, Terry P. Gleason and Douglas A. Reynolds
Lincoln Laboratory, Massachusetts Institute of Technology
ptorres@ll.mit.edu, tgleason@ll.mit.edu, dar@ll.mit.edu

ABSTRACT
three of the languages, recorded over domestic telephone
lines. The corpus consists of a training partition used to train
Recent results in the area of language identification have
the language models of the system, a development partition
shown a significant improvement over previous systems. In
for parameter tuning and an evaluation partition used to test
this paper, we evaluate the related problem of dialect
performance. The 12 languages are: Arabic, English, Farsi,
identification using one of the techniques recently
French, German, Hindi, Japanese, Korean, Mandarin,
developed for language identification, the Gaussian mixture
Spanish, Tamil and Vietnamese. Of particular interest for the
models with shifted-delta-cepstral features. The system
work shown in this paper are the dialects included for three
shown is developed using the same methodology followed
(English, Mandarin, and Spanish) of the 12 languages. Each
for the language identification case. Results show that the
of these languages includes two dialects: North and South
use of the GMM techniques yields an average of 30% equal
for English, Mandarin and Taiwanese for Chinese, and
error rate for the dialects in the Miami corpus and about
Caribbean and Non-Caribbean for Spanish.
13% equal error rate for the dialects in the CallFriend

corpus.
The training partition of the CallFriend corpus includes 20
conversations for each dialect in the three languages,
1. INTRODUCTION
English, Mandarin and Spanish, resulting in 40
conversations per language. Each training conversation in
Over the last few years, the area of language identification
the CallFriend corpus is about 30 minutes long. The
has seen the development of new techniques including
development partition and the evaluation partition for the
novel approaches using Gaussian mixture models, support
CallFriend corpus include 80 testing utterances per dialect,
vector machines, and improvements on the more classical
except for the case of English where an additional group of
approaches based on phone recognition and language
about 320 utterances are included in the evaluation partition.
modeling, such as the PPRLM system[1-3]. The problem of
This set of 320 utterances include utterances from the King
dialect identification, which is closely related to the
corpus, the OGI corpus and Switchboard. Each of the speech
language identification problem, has not received the same
utterances available in the development and evaluation
level of research interest. The task of dialect identification
partitions and used for this work is about 30 seconds.
is that of identifying the spoken dialect from within a set of

utterances in a known language, (e.g., North versus South,
A second corpus, the Miami corpus, consists of two dialects
in the case of American English). Due to the similar nature
within the Spanish language. This corpus, which is described
of dialects within a language, dialect identification poses a
in [5], includes utterances for Cuban Spanish and Peruvian
more difficult problem than language identification.
Spanish. The corpus is partitioned in three disjoint sets

called E, F, and G as shown in Table 1.
The potential applications for dialect identification are

similar to those in the area of language identification,
Set Dialect
Number
of
including among others pre-processing of the incoming
utterances
speech for other downstream processing such as automatic
speech recognition systems.
E
Cuban 37

Peruvian 23
The work presented in this paper is focused on applying
F
Cuban 39
some of the techniques developed for language identification
Peruvian 20
community to the area of dialect identification. The
G
Cuban 16
organization of this paper is as follows: Section 2 describes
Peruvian 8
the two corpora used for the dialect ID experiments. Section
TABLE 1. Partition of the Miami corpus into three sets E, F,
3 describes briefly the system based on GMM acoustic
and G.
scores. Section 4 discusses the dialect ID results obtained for

both corpora and Section 5 presents conclusions and
The speech utterances in the Miami corpus are about 3
proposals for future work.
minutes long captured from an interview of native speakers.
The effective size of the utterances is around 1.5 minutes as
2. CORPORA
the interviewer’s voice is eliminated from the utterances.
Each of these utterances was recorded using a boom
The CallFriend corpus [4] is a collection of unscripted
microphone and recorded with a Sony digital audio tape with
conversations for 12 languages, including two dialects for
a sampling rate of 48 kHz and 16-bit quantization.

• This work is sponsored by the Air Force Research Laboratory under Air Force Contract F19628-00-C-0002. Opinions,
interpretations, conclusions and recommendations are those of the authors and are not necessarily endorsed by the United States
Government.

---

3. GMM-SDC system
F) is used for training the GMMs with the remaining set (E
or F) along with set G used for evaluation. The GMM
The Gaussian mixture model (GMM) using shifted delta
order used for this set of experiments is 256 as higher
cepstral features (SDC) or GMM-SDC, system employed
order models did not trained reliably. Results for the
in this work has been previously described in more detail
Miami corpus are summarized in Table 2, with the 95%
in [2, 3] and is shown in Figure 1. The GMM-SDC system
confidence intervals for these values at approximately
consist of a set of GMMs trained for each dialect of
±10.0 for testing on sets E and F, and ±20.0 for testing on
interest. For example, for the case of discriminating
set G.
between the English dialects, two models are trained, one

for each dialect. Similar to the work performed for LID,
Training
Evaluation
Classification
EER (%)
the features used for the system are based on shifted-delta
set
set
Error (%)
cepstra, which are a set of features derived from the
E F 28.51 35.59

classical set of delta cepstra.
E G 45.83 41.67

F E 23.73 32.20
FRONT-END
FEATURE VECTORS
F G 41.67 41.67
PROCESSING
TABLE 2. Summary of results for the Miami corpus.
INPUT SPEECH

The performance obtained for the Miami corpus is worse
than that obtained in [5]. The system evaluated in [5]
resulted in a better classification error by about 15%, when
HYPOTHESIZED
DIALECT 1 MODEL
the E partition was used for training, and about 5% better
LANGUAGE
CHOOSE
when the F partition was used for training. Two subtleties
MAX
DIALECT 2 MODEL
in the Miami corpus need to be studied further. First is the
amount of training data available and the availability of
FIGURE 1. GMM-SDC system as used for the dialect ID
almost twice as much data for Cuban compared to
problem.
Peruvian. Additional experiments conducted showed that

the classification is highly biased toward Cuban as the
The training and testing procedures are as follows. During
GMM order is increased. This result might be an
training, each of the training conversations is processed by
indication of a need for balancing the training data set to
the front-end and a feature set using a 7-1-3-7 SDC
include similar amount for each dialect. This limitation in
parameterization is obtained for each frame. The full set of
the GMM orders that could be reliably trained is also
feature vectors is used to build a Universal Background
considered an important factor in the performance obtained
Model (UBM) using all the training conversations in the
Second, the choice of SDC parameterization values needs
CallFriend corpus. The dialect-dependent models are then
to be evaluated further to understand its effect for this data
trained by adapting the UBM similar to the technique
set.
described in [6].

4.2. CallFriend corpus
During testing, an incoming speech utterance is processed
For the speech utterances in the CallFriend corpus, three
by the front-end and scored against each dialect-dependent
experiments were conducted dealing with the classification
model using a fast scoring scheme as described in [7]. The
of the dialects on each of the languages available, English,
model scoring highest is hypothesized as the model of the
Mandarin and Spanish. The results for the experiments are
incoming utterance.
shown in Table 3, with the 95% confidence intervals for

these values at approximately ±2.5.
For the case of the CallFriend corpus, where a significant

amount of data is available, we will evaluate the use of a
From the results in Table 3, the results for the CallFriend
backend classifier instead of a maximum likelihood
corpus are better than those obtained for the Miami corpus.
decision. Employing this backend classifier enhances the
Although such a comparison is not necessarily fair as the
classification performance by correcting some of the errors
conditions for each corpus are different, the increased
produced in the maximum likelihood decision criteria.
performance obtained can be due to either the higher
amount of training data, which allows for a GMM order of
4. EXPERIMENTS
2048 to be trained, or the use of a parameterization for the
This section presents the results obtained for both the
SDC features that was chosen because of its performance
CallFriend corpus and the Miami corpus. The set of results
for the language identification case on previous
presented for the CallFriend corpus is more extensive as
experiments for this corpus. Additionally, the collection
the amount of data available allows more flexibility for
for each corpus is different which needs to be addressed in
further analysis and improvements.
future experiments.

4.1. Miami corpus
Additional experiments were conducted to assess the use
of a backend classifier (BE) in the dialect identification
The first set of results shown is based on the experiments
case. The first experiment conducted was performed by
conducted for the Miami corpus. The results obtained are
scoring the development set against the two dialect models
presented using two scenarios. Either one of two sets (E or

---

available. For each case, the experiments resulted in a
decrease in performance for two of the three set of dialects
evaluated.

Dialects Classification
EER
Error (%)
(%)
English 28.45
15.06
Mandarin 28.21
11.54
Spanish 33.33
13.73
TABLE 3. Summary of results for the CallFriend
corpus.


FIGURE 3. Comparison of results using a backend
classifier in the case of Mandarin dialects.


FIGURE 2. Comparison of results using a backend
classifier in the case of English dialects.

The results in Figures 2, 3 and 4 show that no
improvements are obtained when a Gaussian backend
classifier is employed.

A second experiment was conducted to determine whether
additional improvements can be obtained when auxiliary
information is added to the backend classifier. In this

experiment, rather than limiting the data seen by the
FIGURE 4. Comparison of results using a backend
classifier to only that of the different dialects of the same
classifier in the case of Spanish dialects.
language (two scores per test utterance), the classifier is

designed to utilize all the data available in the CallFriend
corpus (15 scores per test utterance). All the development
utterances are scored and used for training the classifier.
Results for this case are shown in Fig. 5, 6 and 7.

The results shown clearly demonstrate the advantage of
using the auxiliary information for the backend classifier.
The impact of the backend classifier for the case of the
English dialects is not quite as large as for the other two
cases. This result originally was considered to be due to
the presence of the 320 out of corpus utterances in the
evaluation set for English, but additional experiments
conducted showed that eliminating the out of corpus
utterances did not improved performance.

---

FIGURE 5. Comparison of the two backends
FIGURE 7. Comparison of the two backends
evaluated for the English dialects.
evaluated for the Spanish dialects.


Future work in this area includes development of newer
techniques and better understanding of the technique as it
applies to the particular case of dialect identification, along
with the evaluation of other language identification systems
like PPRLM and SVM Additional work is expected to
include evaluation of other algorithm developed for
language identification along with better tuning of the
algorithms for the problem at hand.
6.REFERENCES
[1]P. A. Torres-Carrasquillo, D. A. Reynolds, and J. R.
Deller Jr., "Language Identification Using Gaussian
Mixture Model Tokenization," at ICASSP, Orlando, Fl. ,
USA, 2002.
[2]P. A. Torres-Carrasquillo, E. Singer, M. A. Kohler, R.
Greene, D. A. Reynolds, and J. R. Deller Jr.,
"Approaches to Language Identification using Gaussian
Mixture Models and Shifted Delta Cepstral Features," at

ICSLP, Denver, Co, 2002.
FIGURE 6. Comparison of the two backends
[3]E. Singer, P. A. Torres-Carrasquillo, T. Gleason, W. M.
evaluated for the Mandarin dialects.
Campbell, and D. A. Reynolds, "Acoustic, Phonetic and
Discriminative Approaches to Automatic Language
5.CONCLUSIONS
Recognition," at EuroSpeech, 2003.
[4]"CallFriend Corpus," Linguistic Data Consortium,
This paper has presented results in the area of dialect
1996.
identification by evaluating one of the current state of the art
[5]M. A. Zissman, T. P. Gleason, D. M. Rekart, and B. L.
systems developed in the area of language identification.
Losiewicz, "Automatic Dialect Identification of
The results shown demonstrate the potential of the GMM
Extemporaneous, Conversational, Latin American
technique as a candidate technique in the area of dialect
Spanish Speech," at ICASSP, Atlanta, Georgia, 1996.
identification.
[6]E. Wong and S. Sridharan, "Methods to Improve

Gaussian Mixture Model Based Language Identification
The performance obtained for the GMM system evaluated in
System," at ICSLP, Denver,CO, 2002.
this paper is lower than that obtained in previous work with
[7]J. McLaughlin, D. A. Reynolds, and T. Gleason, "A
the Miami corpus[5] but the system provides very good
Study of Computation Speed-UPS of the GMM-UBM
performance for two of the dialects in the CallFriend
Speaker Recognition System," at EuroSpeech, Seattle,
corpus. Of particular importance is the fact that the
Washington, 1999.
technique has been ported without any specialization for the

case of dialect identification and the results obtained are
promising.

---

Document Outline

• ÿ
• ÿ
  • ÿ
• ÿ

---