The application of large language models (LLMs) to speech translation (ST) or, in general, to machine translation (MT) has recently provided excellent results, superseding conventional encoder-decoder MT systems in the general domain. However, this is not clearly the case when LLMs as MT systems are translating medical-related materials. In this respect, the provision of multilingual training materials for oncology professionals is a goal of the EU project Interact-Europe in which this work was framed. To this end, cross-language technology adapted to the oncology domain was developed, evaluated and deployed for multilingual interspecialty medical education. More precisely, automatic speech recognition (ASR) and MT models were adapted to the oncology domain to translate English pre-recorded training videos, kindly provided by the European School of Oncology (ESO), into French, Spanish, German and Slovene. In this work, three categories of MT models adapted to the medical domain were assessed: bilingual encoder-decoder MT models trained from scratch, pre-trained large multilingual encoder-decoder MT models, and multilingual decoder-only LLMs. The experimental results underline the competitiveness in translation quality of LLMs compared to encoder-decoder MT models. Finally, the ESO speech dataset, comprising roughly 1000 videos and 745 h for the training and evaluation of ASR, MT and ST models, was publicly released for the scientific community.

Abstract Streaming Machine Translation (MT) is the task of translating an unbounded input text stream in real-time. The traditional cascade approach, which combines an Automatic Speech Recognition (ASR) and an MT system, relies on an intermediate segmentation step which splits the transcription stream into sentence-like units. However, the incorporation of a hard segmentation constrains the MT system and is a source of errors. This paper proposes a Segmentation-Free framework that enables the model to translate an unsegmented source stream by delaying the segmentation decision until after the translation has been generated. Extensive experiments show how the proposed Segmentation-Free framework has better quality-latency trade-off than competing approaches that use an independent segmentation model.1

Although Long-Short Term Memory (LSTM) networks and deep Transformers are now extensively used in offline ASR, it is unclear how best offline systems can be adapted to work with them under the streaming setup. After gaining considerable experience on this regard in recent years, in this paper we show how an optimized, low-latency streaming decoder can be built in which bidirectional LSTM acoustic models, together with general interpolated language models, can be nicely integrated with minimal perfomance degradation. In brief, our streaming decoder consists of a one-pass, real-time search engine relying on a limited-duration window sliding over time and a number of ad hoc acoustic and language model pruning techniques. Extensive empirical assessment is provided on truly streaming tasks derived from the well-known LibriSpeech and TED talks datasets, as well as from TV shows on a main Spanish broadcasting station.

The rapid progress of modern AI tools for automatic speech recognition and machine translation is leading to a progressive cost reduction to produce publishable subtitles for educational videos in multiple languages. Similarly, text-to-speech technology is experiencing large improvements in terms of quality, flexibility and capabilities. In particular, state-of-the-art systems are now capable of seamlessly dealing with multiple languages and speakers in an integrated manner, thus enabling lecturer’s voice cloning in languages she/he might not even speak. This work is to report the experience gained on using such systems at the Universitat Politècnica de València (UPV), mainly as a guidance for other educational organizations willing to conduct similar studies. It builds on previous work on the UPV’s main repository of educational videos, MediaUPV, to produce multilingual subtitles at scale and low cost. Here, a detailed account is given on how this work has been extended to also allow for massive machine dubbing of MediaUPV. This includes collecting 59 h of clean speech data from UPV’s academic staff, and extending our production pipeline of subtitles with a state-of-the-art multilingual and multi-speaker text-to-speech system trained from the collected data. Our main result comes from an extensive, subjective evaluation of this system by lecturers contributing to data collection. In brief, it is shown that text-to-speech technology is not only mature enough for its application to MediaUPV, but also needed as soon as possible by students to improve its accessibility and bridge language barriers.

The cascade approach to Speech Translation (ST) is based on a pipeline that concatenates an Automatic Speech Recognition (ASR) system followed by a Machine Translation (MT) system. Nowadays, state-of-the-art ST systems are populated with deep neural networks that are conceived to work in an offline setup in which the audio input to be translated is fully available in advance. However, a streaming setup defines a completely different picture, in which an unbounded audio input gradually becomes available and at the same time the translation needs to be generated under real-time constraints. In this work, we present a state-of-the-art streaming ST system in which neural-based models integrated in the ASR and MT components are carefully adapted in terms of their training and decoding procedures in order to run under a streaming setup. In addition, a direct segmentation model that adapts the continuous ASR output to the capacity of simultaneous MT systems trained at the sentence level is introduced to guarantee low latency while preserving the translation quality of the complete ST system. The resulting ST system is thoroughly evaluated on the real-life streaming Europarl-ST benchmark to gauge the trade-off between quality and latency for each component individually as well as for the complete ST system.

Current research into spoken language translation (SLT), or speech-to-text translation, is often hampered by the lack of specific data resources for this task, as currently available SLT datasets are restricted to a limited set of language pairs. In this paper we present Europarl-ST, a novel multilingual SLT corpus containing paired audio-text samples for SLT from and into 6 European languages, for a total of 30 different translation directions. This corpus has been compiled using the debates held in the European Parliament in the period between 2008 and 2012. This paper describes the corpus creation process and presents a series of automatic speech recognition, machine translation and spoken language translation experiments that highlight the potential of this new resource. The corpus is released under a Creative Commons license and is freely accessible and downloadable.

Current state-of-the-art models based on Long-Short Term Memory (LSTM) networks have been extensively used in ASR to improve performance. However, using LSTMs under a streaming setup is not straightforward due to real-time constraints. In this paper we present a novel streaming decoder that includes a bidirectional LSTM acoustic model as well as an unidirectional LSTM language model to perform the decoding efficiently while keeping the performance comparable to that of an off-line setup. We perform a one-pass decoding using a sliding window scheme for a bidirectional LSTM acoustic model and an LSTM language model. This has been implemented and assessed under a pure streaming setup, and deployed into our production systems. We report WER and latency figures for the well-known LibriSpeech and TED-LIUM tasks, obtaining competitive WER results with low-latency responses.

Recurrent Neural Networks, in particular Long-Short TermMemory (LSTM) networks, are widely used in Automatic Speech Recognition for language modelling during decoding,usually as a mechanism for rescoring hypothesis. This paperproposes a new architecture to perform real-time one-pass de-coding using LSTM language models. To make decoding ef-ficient, the estimation of look-ahead scores was accelerated byprecomputing static look-ahead tables. These static tables wereprecomputed from a prunedn-gram model, reducing drasti-cally the computational cost during decoding. Additionally,the LSTM language model evaluation was efficiently performedusing Variance Regularization along with a strategy of lazyevaluation. The proposed one-pass decoder architecture wasevaluated on the well-known LibriSpeech and TED-LIUMv3datasets. Results showed that the proposed algorithm obtainsvery competitive WERs with 0.6 RTFs. Finally, our one-passdecoder is compared with a decoupled two-pass decoder.

In the last years, deep bidirectional recurrent neural networks (DBRNN) and DBRNN with long short-term memory cells (DBLSTM) have outperformed the most accurate classifiers for confidence estimation in automatic speech recognition. At the same time, we have recently shown that speaker adaptation of confidence measures using DBLSTM yields significant improvements over non-adapted confidence measures. In accordance with these two recent contributions to the state of the art in confidence estimation, this paper presents a comprehensive study of speaker-adapted confidence measures using DBRNN and DBLSTM models. First, we present new empirical evidences of the superiority of recurrent neural networks (RNN)-based confidence classifiers evaluated over a large speech corpus consisting of the English LibriSpeech and the Spanish poliMedia tasks. Second, we show new results on speaker-adapted confidence measures considering a multitask framework in which RNN-based confidence classifiers trained with LibriSpeech are adapted to speakers of the TED-LIUM corpus. These experiments confirm that speaker-adapted confidence measures outperform their non-adapted counterparts. Last, we describe an unsupervised adaptation method of the acoustic DBLSTM model based on confidence measures that results in better automatic speech recognition performance.

In this paper a vertical repositioning method based on the center of gravity is investigated for handwriting recognition systems and evaluated on databases containing Arabic and French handwriting. Experiments show that vertical distortion in images has a large impact on the performance of HMM based handwriting recognition systems. Recently good results were obtained with Bernoulli HMMs (BHMMs) using a preprocessing with vertical repositioning of binarized images. In order to isolate the effect of the preprocessing from the BHMM model, experiments were conducted with Gaussian HMMs and the LSTM-RNN tandem HMM approach with relative improvements of 33% WER on the Arabic and up to 62% on the French database.

Bernoulli-based models such as Bernoulli mixtures or Bernoulli HMMs (BHMMs), have been successfully applied to several handwritten text recognition (HTR) tasks which range from character recognition to continuous and isolated handwritten words. All these models belong to the generative model family and, hence, are usually trained by (joint) maximum likelihood estimation (MLE). Despite the good properties of the MLE criterion, there are better training criteria such as maximum mutual information (MMI). The MMI is a widespread criterion that is mainly employed to train discriminative models such as log-linear (or maximum entropy) models. Inspired by the Bernoulli mixture classifier, in this work a log-linear model for binary data is proposed, the so-called mixture of multi-class logistic regression. The proposed model is proved to be equivalent to the Bernoulli mixture classifier. In this way, we give a discriminative training framework for Bernoulli mixture models. The proposed discriminative training framework is applied to a well-known Indian digit recognition task.

Hidden Markov Models (HMMs) are now widely used in off-line handwriting recognition and, in particular, in Arabic handwritten word recognition. In contrast to the conventional approach, based on Gaussian mixture HMMs, we have recently proposed to directly fed columns of raw, binary pixels into Bernoulli mixture HMMs. In this work, column bit vectors are extended by means of a sliding window of adequate width to better capture image context at each horizontal position of the word image. Using these windowed Bernoulli mixture HMMs, very good results are reported on the well-known IfN/ENIT database of Arabic handwritten Tunisian town names.

Hidden Markov Models (HMMs) are now widely used in off-line handwritten word recognition. As in speech recognition, they are usually built from shared, embedded HMMs at symbol level, in which state-conditional probability density functions are modelled with Gaussian mixtures. In contrast to speech recognition, however, it is unclear which kind of real-valued features should be used and, indeed, very different features sets are in use today. In this paper, we propose to by-pass feature extraction and directly fed columns of raw, binary image pixels into embedded Bernoulli mixture HMMs, that is, embedded HMMs in which the emission probabilities are modelled with Bernoulli mixtures. The idea is to ensure that no discriminative information is filtered out during feature extraction, which in some sense is integrated into the recognition model. Empirical results are reported in which similar results are obtained with both Bernoulli and Gaussian mixtures, though Bernoulli mixtures are much simpler.