[Univ of Cambridge]alt[Dept of Engineering]

MIL Speech Seminars 2006-2007

The MIL Speech Seminars occur roughly every two weeks during full term and occasionally during vacations. The purpose of the seminars is to enable researchers in various fields to present recent work that has potential benefits for speech technology.

The MIL Speech Seminar series schedule for Michaelmas Term was:

31st October 2006 Adria de Gispert TALP Research Centre, Univ. Politecnica de Catalunya (UPC), Barcelona, Spain Use of Linguistic Information and Reordering Strategies for Ngram- based Statistical Machine Translation This seminar will be devoted to an overview of the experience in statistical machine translation at UPC during recent years. Firstly, the Ngram-based SMT system will be described, detailing bilingual unit definition and basic feature functions for a monotone language pair. Secondly, the introduction of linguistic information at various stages will be discussed, including word alignment (investigating correlation between Alignment Error Rate and translation scores), bilingual unit segmentation and direct translation modelling. Results on English-to-Spanish verb form classification will be reviewed, as well as the impact of morphology reduction on bilingual N-gram formulation. For language pairs exhibiting less monotone word order, the reordering strategies implemented will be presented. Particularly, reordered search involving tuple unfolding and extended monotone search by linguistically-driven reordering rules will be compared for Arabic, Chinese and Spanish-to-English tasks. Finally, the seminar will conclude outlining general future research directions towards improving performance of current state-of-the-art SMT systems.
20th November 2006 Arantza Del Pozo (MIL) Tracheoesophageal Speech Repair Tracheoesophageal (TE) speech is the most frequently used voice restoration technique after total laryngectomy. Despite being often cited as the alaryngeal speech alternative most comparable to normal, its quality and intelligibility are still significantly lower than laryngeal speech. Excitation and prosodic deviations are thought to be the main limitations responsible for its quality reduction. This seminar will describe excitation and duration repair systems developed for the enhancement of continuous TE speech. The excitation repair system resynthesises TE speech using a synthetic glottal waveform, reduces its jitter and shimmer and applies a novel spectral smoothing and tilt correction algorithm. For duration repair, TE phone durations are modified based on the predictions of regression trees built from non-pathological data. The perceptual enhancement of each system is evaluated using listening tests. Results show that the repaired sentences are preferred to the original overall in terms of breathiness, harshness and rhythmic naturalness.

If you are interested in giving a seminar presentation or if you would like more information about the seminar series please contact Marcus Tomalin