Portfolio item number 1
Published:
Short description of portfolio item number 1
Posts by Collection
portfolio
Portfolio item number 2
Published:
Short description of portfolio item number 2
publications
Europarl-ST: A Multilingual Corpus for Speech Translation of Parliamentary Debates
Published in ICASSP, 2020
Speech Translation datasets are a scarce resource, and this greatly hampers research in the area. Europarl-ST, first released in 2019, was a game-changer for Speech Translation research, thanks to the wide range of languages covered and a careful filtering pipeline. Currently (early 2023), close to 100 publications have cited Europarl-ST.
Direct Segmentation Models for Streaming Speech Translation
Published in EMNLP, 2020
Machine Translation systems are trained with full sentences, but in the Cascaded Speech Translation scenario, the output of the ASR system does not necessarily form sentences, which hampers performance. This publication introduces a streaming-ready segmenter applied to the output of the ASR system, in order to maximize downstream translation quality.
Streaming cascade-based speech translation leveraged by a direct segmentation model
Published in Neural Networks, 2021
This paper extends the previous one (EMNLP2020) with additional experiments and by moving from a simulated Streaming scenario, which used an offline MT system, to a real streaming scenario with a simultaneous MT system.
Stream-level Latency Evaluation for Simultaneous Machine Translation
Published in Findings of EMNLP, 2021
A reliable evaluation metric is critical for any technical and scientific task. However, the standard simultaneous MT latency metrics (AP, AL and DAL) are not robust when applied to the Streaming scenario. This paper studies this phenomenon and proposes a re-segmentation solution that provides reliable and interpretable results for the Streaming scenario.
From Simultaneous to Streaming Machine Translation by Leveraging Streaming History
Published in ACL, 2022
The Streaming ST scenario presents many challenges, but there are also opportunities that can be used to improve translation quality. This work introduces the concept of Streaming history, which holds the information of the previously translated segments. The proposed MT system is able to leverage this contextual information in order to improve translation quality.
talks
Towards Streaming Speech Translation
Published:
On this talk I gave an overview of how Speech Translation is carried out at MLLP research group, with a focus on the MT side.
Unfortunately, the talk was not recorded due to a technical issue, but you can access the slides here.
Towards Streaming Speech Translation - SIGSLT
Published:
This talk provides an overview of the Speech Translation task and its particularities. I then go into detail explaining two contributions of my thesis: 1) a procedure for Streaming ST evaluation, and 2) streaming-specific models with streaming history.
This talk goes into greater detail than my previous one at UPC, so you should definitely watch this one if you interested in Streaming Speech Translation.
The talk is available on Youtube, and you can access the slides here.
My Thesis Defense: Streaming Neural Speech Translation
Published:
This is the presentation of my PhD Thesis. I speak more about this in one of my blog posts. The actual presentation was not recorded, but I sat down a couple of days later and tried to recreate it as best as possible. You can find the video here.