8.1. Introduction

When listening to their native language, speech and hearing-enabled humans can identify the language being spoken immediately. For non-native but familiar languages, we are able to do almost as well. This is the capacity we would like automatic devices to share. Human capabilities are certainly impressive on short speech fragments, however the number of native and familiar languages generally remains very limited compared to the number of distinct spoken languages in the world. Furthermore the risk of error, even if low, is not negligible.

Technically speaking, language identification (LId) involves identifying the language being spoken by an unknown speaker using a given speech sample. Automatic LId appears as one of the expected capacities of "smart" speech-to-text or speech-to-speech systems in a multilingual context. Automatic LId raises many challenging research issues which contribute to progress towards multilingual speech technologies [SCH 06, NAV 06, GEO 04, ZIS 99]. Major questions remain in the domain of automatic language identification. What levels of information contribute most effectively to high accuracy language identification? What are the best acoustic parameters to catch language-specific information? Which algorithms, methods and approaches are the most promising for LId? Which language-specific resources are required for spoken language modeling? What is the correlation between signal duration and the confidence limits of language identification? Related questions of interest also concern human performance: how do humans proceed in identifying languages? Can insight into human cognition be helpful for automatic devices and vice versa? Another important...