Multilingual Speech Processing

262 CHAPTER 8. AUTOMATIC LANGUAGE IDENTIFICATION

was shown to improve the performance, as did the combination of the

articulatory-feature system and the baseline phone-based system.

In summary, Table 8.3 lists selected LID implementations of

phonotactic-only components and their performances spanning the last

decade, in chronological order. Note that despite the common use of the

OGI database, the exact subset used to obtain the error rates may vary in

each study. Furthermore, baseline systems with a variable level of com-

plexity were used, therefore the listed error rates may not be seen as an

outline of progress in phonotactic modeling in an absolute sense.

8.7 Prosodic LID

Prosodic information—such as tone, intonation, and prominence—is pri-

marily encoded in two signal components: fundamental frequency (F

0

) and

amplitude. Thus, properties of F

0

and amplitude contours can intuitively be

expected to be useful in automatic LID. In Eady (1982), for instance, two

languages with different prosodic properties were studied: English, which

belongs to the category of languages marking different levels of promi-

nence by means of F

0

and amplitude, and Chinese, in which tonal patterns

are lexically distinctive (see Chapter 2). The author compared the time

contours of the fundamental frequency and the energy extracted from sen-

tences of the same type (e.g., all declarative sentences) and found unique

language-specific differences—such as a higher rate of change within

each contour—and a higher fluctuation within individual syllables for

Chinese.

However, the use of prosody for the purpose of identifying languages is

not without problems. The difficulty of obtaining a clear assessment of the

usefulness of a prosodic component for LID derives from the large number

of additional factors influencing F

0

and amplitude. These include:

•

speaker-specific characteristics (voice type, characteristic speaking

rate, emotional state, health, etc.)

•

lexical choice (word characteristics, such as word stress and lexical

tone)

•

syntactic content of the utterance (statement, question)

•

pragmatic content/function in discourse (contrastive emphasis, given

versus new distinction).

8.7. PROSODIC LID 263

Table 8.3 Examples of phonotactic LID systems and their recognition rates.

System Task Test Signal Duration ID Rate Ref.

Interpolated Trigram OGI-11L 10 s/45 s 62.7%/77.5% Hazen and Zue, 1997

Multilingual Tokenizer

Bigrams (1 tokenizer) OGI-10L 10 s/45 s 54%/72% Zissman, 1996

Gender-Dependent

PPRLM (3 tokenizers) OGI-10L 10 s/45 s 63%/79% Zissman, 1996

PPRLM (6 tokenizers) OGI-6L 10 s/45 s 74.0%/84.8% Yan and Barnard, 1995

Extended N-grams OGI-6L 10 s/45 s 86.4%/97.5% Navrátil, 2001

(PPRLM with 6 streams)

Cross-Stream

OGI-10L mixed 65% Parandekar and Kirchhoff, 2003

Modeling

GMM-Tokenizer CallFriend (12L) 30 s 63.7% Torres-Carrasquillo et al., 2002a

GMM-Tok.+PPRLM CallFriend (12L) 30 s 83.0% Torres-Carrasquillo et al., 2002a

264 CHAPTER 8. AUTOMATIC LANGUAGE IDENTIFICATION

In many cases, particularly in languages from the same family (or the same

prosodic type; see Section 8.2), the language-specific prosodic variation is

overridden by other, more dominant factors, rendering its extraction and

recognition extremely difficult. In order to successfully exploit prosodic

information, the question of how to best separate language-dependent char-

acteristics from speaker-dependent or other irrelevant characteristics needs

to be addressed and indeed remains one of the open challenges in cur-

rent LID technology. A potentially useful application of prosodic features

might be the preclassification of language samples into broad, prosodically

distinct categories, followed by more fine-grained subclassification using

other methods.

LID systems based solely on prosody are relatively rare. One such

approach was described in Itahashi et al. (1994), and Itahashi and Liang

(1995). In this system, fundamental frequency and energy contours were

extracted from the speech signal and approximated via a piecewise-

linear function. In order to preserve prosodic language characteristics,

a heuristically selected set of variables derived from the line-approximated

representation was calculated (mean slope of the line segments, mean rela-

tive start frequency of a line segment with positive and negative slope, the

correlation coefficient between the fundamental frequency and the energy

contour). Vectors of these features were then processed using principal

components analysis (PCA) to perform a dimensionality reduction. The

PCA-reduced samples were stored and compared to test samples via the

Mahalanobis distance. The system was evaluated on speech material from

six languages (Chinese, English, French, German, Japanese, and Korean);

LID rates ranged between 70 and 100%. The number of speakers used in

the experiments, however, was rather small (5 per language). An identical

approach was tested with data in six languages taken from the OGI-11L

corpus, producing accuracy rates of about 30% (Dung, 1997). Similar

prosodic results were reported by Hazen and Zue (1994). A recent study

used similar modeling of F

0

-based and rhythmic features from pseudo-

syllabic segments of read speech in ten languages, utilizing a Gaussian

Mixture Model (GMM)–based classifier (Farinas et al., 2002).

Probably the most thorough study of prosodic features for LID was

presented in Thyme-Gobbel and Hutchins (1996). The authors com-

pared 220 prosodically motivated features that were calculated from

syllable segments represented as histograms, and used for classifica-

tion with a likelihood ratio detector. The experiments involved a series

of pair-wise language classification tasks on four languages: English,