Videolectures are currently being digitised all over the world for its enormous value as reference resource. Many of these lectures are accompanied with slides. The slides offer a great opportunity for improving ASR systems performance. We propose a simple yet powerful extension to the linear interpolation of language models for adapting language models with slide information. Two types of slides are considered, correct slides, and slides automatic extracted from the videos with OCR. Furthermore, we compare both time aligned and unaligned slides. Results report an improvement of up to 3.8 % absolute WER points when using correct slides. Surprisingly, when using automatic slides obtained with poor OCR quality, the ASR system still improves up to 2.2 absolute WER points.

No abstract

Abstract. An effective approach to transcribe handwritten text documents is to follow a sequential interactive approach. During the supervision phase, user corrections are incorporated into the system through an ongoing retraining process. In the case of multilingual documents with a high percentage of out-of-vocabulary (OOV) words, two principal issues arise. On the one hand, a minor yet important matter for this interactive approach is to identify the language of the current text line image to be transcribed, as a language dependent recognisers typically performs better than a monolingual recogniser. On the other hand, word-based language models suffer from data scarcity in the presence of a large number of OOV words, degrading their estimation and affecting the performance of the transcription system. In this paper, we successfully tackle both issues deploying character-based language models combined with language identification techniques on an entire 764-page multilingual document. The results obtained significantly reduce previously reported results in terms of transcription error on the same task, but showed that a language dependent approach is not effective on top of character-based recognition of similar languages.

Abstract. With the spread of MOOCs and video lecture repositories it is more important than ever to have accurate methods for automatically transcribing video lectures. In this work, we propose a simple yet effective language model adaptation technique based on document retrieval from the web. This technique is combined with slide adaptation, and compared against a strong baseline language model and a stronger slideadapted baseline. These adaptation techniques are compared within two different acoustic models: a standard HMM model and the CD-DNN-HMM model. The proposed method obtains improvements on WER of up to 14% relative with respect to a competitive baseline as well as outperforming slide adaptation.