Evalita 2011: Automatic Speech Recognition Large Vocabulary Transcription

In this paper we describe design, setup and results of the speech recognition task in the framework of the Evalita campaign for the Italian language, giving details on the released corpora and tools used for the challenge. A general discussion about approaches to large vocabulary speech recognition introduces the recognition tasks. Systems are compared for recognition accuracy on audio sequences of Italian par- liament. Although only a few systems have participated to the tasks, the contest provides an overview of the state-of-the-art of speech-to-text transcription technologies; the document reports systems performance, computed as Word Error Rate (WER), showing that the current approaches provide effective results. The best system achieves a WER as low as 5.4% on the released testset