Open Domain Speech Recognition & Translation:Lectures and Speeches

Abstract

For years speech translation has focused on the recognition and translation of discourses in limited domains, such as hotel reservations or scheduling tasks. Only recently research projects have been started to tackle the problem of open domain speech recognition and translation of complex tasks such as lectures and speeches. In this paper we present the on-going work at our laboratory in open domain speech translation of lectures and parliamentary speeches. Starting from a translation system for European parliamentary plenary sessions and a lecture speech recognition system we show how both components perform in unison on speech translation of lectures.

The ISL RT-06S Speech-to-Text System

“…During ML-SAT training, these models were applied to the CHIL training data with a weight of 4.0. Comparing the resulting system to the system used in [1], we improved our second pass result by 1.4% absolute (see Table 2, second row).…”

“…The training setup was based on experiments performed during the development of the lecture translation system [1]. We selected the training data that performs best on close talking audio, using only the ICSI, NIST and TED data and skipping the CMU and the Hub4-BN training material.…”

“…Both the MVDR and the FFT system were trained in the same way, resulting in a size of 16k distributions over 4k models, with a maximum of 64 Gaussians per model. The training was similar to that used in [1], with one modification. A second pass for incremental growing of Gaussians was performed after the STC training, which leads to an additional gain of 0.3% resulting in a WER of 32.0% on the IHM data of lectDEV.…”

“…In this paper, we present the ISL's most recent speech-to-text systems for lectures and conference meetings, which have evolved significantly over previous versions [1][2][3] and which were evaluated in the NIST RT-06S Rich Transcription Meeting Evaluation.…”

“…The systems described in [1] and [3] shared many common elements, e.g. front-end, phoneme set, and training strategy.…”