Multilingual Speech Processing

9.10. SUMMARY 315

been solved by a combination of technical and sociological means. Two

recent examples are the widespread uptake of cellular telephony, and sev-

eral successful AIDS-awareness campaigns. In both cases, the critical

success factor was the importance of the intervention to the lives of the

users by providing dramatically improved communications in the one case,

and limiting the spread of a debilitating disease in the other. If spoken inter-

faces can be shown to have comparable value in the developing world, one

can be confident that the hurdles to their development will be overcome.

9.10 Summary

In this chapter, we have examined characteristics of and approaches to

automatic recognition of non-native, accented, and dialectal speech. This

topic has received comparably little attention in the past. For the real-world

use of speech-driven interfaces, systems need to be able to handle speech

other than purely native speech; the number of languages for which ASR

systems can be built will always be limited, and non-native words will occur

in many domains. To date, most ASR systems show significant increase in

word error rates if faced with non-native speech.

In order to appropriately model non-native speech, we first have to

determine how non-native speech differs from native speech. We have

tried to give some insight into these differences and their origins. We have

shown that influences on non-native speech are complex, far exceeding the

commonly assumed sole influence of the mother tongue.

A particular challenge in non-native and dialectal speech is that a huge

number of permutations exist in terms of which language was spoken

with the accent of which other language. Detailed investigations are often

constrained by the lack of sufficient amounts of non-native or dialectal

data.

We have also described how standard methods in native speech recog-

nition, such as specialized acoustic modeling, speaker adaptation, and

adaptation of the pronunciation dictionary, apply to the non-native case,

and what the particular challenges are for these methods. We have shown

that these methods can indeed help non-native speech if they are tailored

to this special case.

Apart from the modeling issues described above, we have also investi-

gated how the design of different interfaces needs to be modified in order

316 CHAPTER 9. OTHER CHALLENGES: NON-NATIVE SPEECH

to increase usability for non-native speakers, who often have particular

needs and problems when using automated systems. This is of particular

importance for speech-based interfaces in the developing world, where cul-

tural differences need to be taken into account in successfully deploying

speech-based interfaces.

To summarize, we can say that there are certainly some approaches

that are helping increase the performance of speech recognition systems

when faced with non-native and accented speech. On the other hand, many

questions are left unanswered, leaving plenty of room for future research.