Text and speech machine translation
José B. Mariño
After the latest advances in speech recognition and synthesis, these technologies have become sufficiently evolved to make it possible to undertake speech translation. Despite the fact that in the field of natural language processing there is already a firm grounding in text machine translation, the unique features of spoken language mean that the techniques used for texts are not particularly successful when applied to speech. Current state-of-the-art speech translation systems take a statistical approach that is closely related to the methods that have emerged from speech recognition. In the framework of national and European projects (ALIADO, FAME, LC-STAR I TC-STAR), the TALP is developing its own system of statistical speech translation to attain improved performance by incorporating linguistic (morphological, syntactic and semantic) knowledge. Spanish, Catalan and English are the priority languages, but activities are also being carried out in Mandarin and standard Arabic. The TALP’s translation system regularly undergoes assessments by international organizations and has obtained similar results to existing systems. The TALP’s speech translation system can also be applied to texts, for which it has been shown to perform as well (if not better) than conventional text machine systems in the fields of application for which it has been trained.