Multilingual Speech Processing

268 CHAPTER 8. AUTOMATIC LANGUAGE IDENTIFICATION

8.9 Trends and Open Problems in LID

Over the past 10–15 years, language identification has become a well-

established branch of speech research, with regular conference sessions

dedicated to its problems, developments, and applications. Recent LID

evaluations organized by the National Institute of Standards and Tech-

nology (NIST) in 2003 and 2005, as a continuation of previous such

evaluations from 1995–96, contributed to continued attention to this

problem and spurred new activity in this area.

We have discussed how the main approaches can be described in

terms of the signal features they seek to exploit as well as in terms of

their modeling techniques. However, the various degrees of difficulty

in the LID experiments, such as test duration, amount of training data

and data annotation, channel quality, and test training mismatch, render

a consistent comparison of the individual algorithms difficult. Based on a

somewhat rough comparison, a performance-oriented ranking will place

the LVCSR-based systems among the most powerful methods, followed

by the phonotactic, the acoustic, and finally the prosodic approaches.

A major trend in current LID methodology is system and component

fusion. While a system fusion aims at combining entire LID systems, the

component fusion exploits the fact that different sources of discriminative

information within one LID system provide partially decorrelated output

and therefore may lead to improved performance when integrated into one

decision process. Most common fusion implementations combine outputs

of the individual components, such as phonotactic, acoustic, and prosodic

modeling, into a final hypothesis by means of a higher-level classifier. This

classifier can be a simple linear function (Hazen and Zue, 1997) or a non-

linear classifier, such as a multilayer perceptron (Navrátil, 2001; Yan et al.,

1996).

An interesting and equally important aspect of comparing LID

approaches is the intended application. Differentapplications pose different

demands due to their choice of languages, typical test duration, availability

of training material, and so on. Table 8.6 summarizes some of the advan-

tages distinguishing individual methods from several application-oriented

aspects.

While LVCSR-based LID systems can be considered the most accu-

rate solution to LID today (Hieronymus and Kadambe, 1997), their

high computational cost as well as the requirement for sufficient

8.9. TRENDS AND OPEN PROBLEMS IN LID 269

Table 8.6 Comparison of basic LID approaches from an application

development aspect.

Basic Strength Constraint Example

Approach Application

Prosodic Robust in channel Mostly suitable Preclassifier of

mismatch for distinguishing tonal versus stress

language groups languages

Acoustic Low cost in training Usable in Language and accent

and testing (data combination with ID in a multiapproach

and computation) other components system

Phonotactic Good performance- Useful for tests LID system with large

-to-cost ratio, with duration language population,

No linguistic >5 seconds including rare languages

knowledge without linguistically

required to train labeled training data

LVCSR/ High accuracy, Significant Multilingual

Keyword Short tests training effort, dialog systems with

Spotting

Linguistic input LVCSR components,

required Audio mining systems

word-transcribed training data, limits their area of use. Potential appli-

cation scenarios are multilingual dialog or translation systems, in which

a multilingual LVCSR infrastructure is already in place, independently

of the LID task. In order to reduce computational demands, speech sig-

nals can first be decoded by all LVCSR systems in parallel, followed by

a process of hypothesis pruning and eventual convergence on a single

language hypothesis. In the extreme case of systems with limited vocab-

ularies and finite-state grammars, the LID task can be implemented as an

implicit by-product of word recognition by unifying all words in a common

(multilingual) dictionary.

In terms of performance-to-cost ratio, the phonotactic approaches

seem to fare best. Their main advantage is that they are capable of

modeling arbitrary languages in terms of sound units of other lan-

guages (Zissman, 1996) and hence do not require manual labeling and

segmentation. They therfore represent an acceptable solution for many

practical applications—especially those involving large numbers of lan-

guages or rare and resource-poor languages. The many improvements in

the area of N-gram model smoothing (Section 8.6.2), context clustering