Speech Recognition System Combination for Machine Translation

Abstract

The majority of state-of-the-art speech recognition systems make use of system combination. The combination approaches adopted have traditionally been tuned to minimising Word Error Rates (WERs). In recent years there has been growing interest in taking the output from speech recognition systems in one language and translating it into another. This paper investigates the use of cross-site combination approaches in terms of both WER and impact on translation performance. In addition the stages involved in modifying the output from a Speech-to-Text (STT) system to be suitable for translation are described. Two source languages, Mandarin and Arabic, are recognised and then translated using a phrase-based statistical machine translation system into English. Performance of individual systems and cross-site combination using cross-adaptation and ROVER are given. Results show that the best STT combination scheme in terms of WER is not necessarily the most appropriate when translating speech.

Advances in Arabic broadcast news transcription at RWTH

“…However, the system combination yields slightly better results. These results differ from those in [21], although different tasks and error metrics were used.…”

Syllable language models for Mandarin speech recognition: Exploiting character language models

“…When component systems use different word segmentation schemes, a direct combination between their outputs is problematic, for example, in Chinese where different character to word segmentations are used. Hence, for the Mandarin speech recognition tasks considered here, the most successful approach is to perform a character level combination, [3][4][5]11 as is also considered in this paper. This requires the mapping of outputs from a standard word based system to sub-word, character level.…”

“…2 Therefore, the character error rate (CER) is the commonly used evaluation metric for state-of-the-art Mandarin speech recognition systems. [3][4][5] All languages have constrained syllable constructions and syllable sequence rules which enhance intelligibility. 6 These phonological and pragmatic constraints can be exploited for Chinese speech recognition.…”

Digital Comics Image Indexing Based on Deep Learning

“…Concerning the evaluation of text recognition, we rely on standard metrics of speech recognition such as Character Error Rate (CER) and Word Error Rate (WER) for determining the recognition accuracy of the segmentation free OCR output [91]. Note that we measure the quality of the generated pseudo ground-truth only by counting the number of validated text lines because, in our experiment, these data are automatically generated (no manual ground-truth available).…”