AppTek participated in the subtitling and formality tracks of the IWSLT 2023 evaluation. This paper describes the details of our subtitling pipeline - speech segmentation, speech recognition, punctuation prediction and inverse text normalization, text machine translation and direct speech-to-text translation, intelligent line segmentation - and how we make use of the provided subtitling-specific data in training and fine-tuning. The evaluation results show that our final submissions are competitive, in particular outperforming the submissions by other participants by 5% absolute as measured by the SubER subtitle quality metric. For the formality track, we participate with our En-Ru and En-Pt production models, which support formality control via prefix tokens. Except for informal Portuguese, we achieve near perfect formality level accuracy while at the same time offering high general translation quality.

Position-based token-mixing approaches, such as FNet and MLPMixer, have shown to be exciting attention alternatives for computer vision and natural language understanding. The motivation is usually to remove redundant operations for higher efficiency on consumer GPUs while maintaining Transformer quality. On the hardware side, research on memristive crossbar arrays shows the possibility of efficiency gains up to two orders of magnitude by performing in-memory computation with weights stored on device. While it is impossible to store dynamic attention weights based on token-token interactions on device, position-based weights represent a concrete alternative if they only lead to minimal degradation. In this paper, we propose position-based attention as a variant of multi-head attention where the attention weights are computed from position representations. A naive replacement of token vectors with position vectors in self-attention results in a significant loss in translation quality, which can be recovered by using relative position representations and a gating mechanism. We show analytically that this gating mechanism introduces some form of word dependency and validate its effectiveness experimentally under various conditions. The resulting network, rPosNet, outperforms previous position-based approaches and matches the quality of the Transformer with relative position embedding while requiring 20% less attention parameters after training.

Video dubbing is the activity of revoicing a video while offering a viewing experience equivalent to the original video. The revoicing usually comes with a changed script, mostly in a different language, and the revoicing should reproduce the original emotions, coherent with the body language, and lip synchronized. In this project, we aim to build an AD system in three phases: (1) voice-over; (2) emotional voice-over; (3) full dubbing, while enhancing the system with human-in-the-loop capabilities for a higher quality.

This paper describes the offline and simultaneous speech translation systems developed at AppTek for IWSLT 2021. Our offline ST submission includes the direct end-to-end system and the so-called posterior tight integrated model, which is akin to the cascade system but is trained in an end-to-end fashion, where all the cascaded modules are end-to-end models themselves. For simultaneous ST, we combine hybrid automatic speech recognition with a machine translation approach whose translation policy decisions are learned from statistical word alignments. Compared to last year, we improve general quality and provide a wider range of quality/latency trade-offs, both due to a data augmentation method making the MT model robust to varying chunk sizes. Finally, we present a method for ASR output segmentation into sentences that introduces a minimal additional delay.

Translating from languages without productive grammatical gender like English into gender-marked languages is a well-known difficulty for machines. This difficulty is also due to the fact that the training data on which models are built typically reflect the asymmetries of natural languages, gender bias included. Exclusively fed with textual data, machine translation is intrinsically constrained by the fact that the input sentence does not always contain clues about the gender identity of the referred human entities. But what happens with speech translation, where the input is an audio signal? Can audio provide additional information to reduce gender bias? We present the first thorough investigation of gender bias in speech translation, contributing with: i) the release of a benchmark useful for future studies, and ii) the comparison of different technologies (cascade and end-to-end) on two language directions (English-Italian/French).

This paper describes FBK’s participation in the IWSLT 2020 offline speech translation (ST) task. The task evaluates systems’ ability to translate English TED talks audio into German texts. The test talks are provided in two versions: one contains the data already segmented with automatic tools and the other is the raw data without any segmentation. Participants can decide whether to work on custom segmentation or not. We used the provided segmentation. Our system is an end-to-end model based on an adaptation of the Transformer for speech data. Its training process is the main focus of this paper and it is based on: i) transfer learning (ASR pretraining and knowledge distillation), ii) data augmentation (SpecAugment, time stretch and synthetic data), iii)combining synthetic and real data marked as different domains, and iv) multi-task learning using the CTC loss. Finally, after the training with word-level knowledge distillation is complete, our ST models are fine-tuned using label smoothed cross entropy. Our best model scored 29 BLEU on the MuST-CEn-De test set, which is an excellent result compared to recent papers, and 23.7 BLEU on the same data segmented with VAD, showing the need for researching solutions addressing this specific data condition.

Current research on spoken language translation (SLT) has to confront with the scarcity of sizeable and publicly available training corpora. This problem hinders the adoption of neural end-to-end approaches, which represent the state of the art in the two parent tasks of SLT: automatic speech recognition and machine translation. To fill this gap, we created MuST-C, a multilingual speech translation corpus whose size and quality will facilitate the training of end-to-end systems for SLT from English into 8 languages. For each target language, MuST-C comprises at least 385 hours of audio recordings from English TED Talks, which are automatically aligned at the sentence level with their manual transcriptions and translations. Together with a description of the corpus creation methodology (scalable to add new data and cover new languages), we provide an empirical verification of its quality and SLT results computed with a state-of-the-art approach on each language direction.

Neural Machine Translation (NMT) models generally perform translation using a fixed-size lexical vocabulary, which is an important bottleneck on their generalization capability and overall translation quality. The standard approach to overcome this limitation is to segment words into subword units, typically using some external tools with arbitrary heuristics, resulting in vocabulary units not optimized for the translation task. Recent studies have shown that the same approach can be extended to perform NMT directly at the level of characters, which can deliver translation accuracy on-par with subword-based models, on the other hand, this requires relatively deeper networks. In this paper, we propose a more computationally-efficient solution for character-level NMT which implements a hierarchical decoding architecture where translations are subsequently generated at the level of words and characters. We evaluate different methods for open-vocabulary NMT in the machine translation task from English into five languages with distinct morphological typology, and show that the hierarchical decoding model can reach higher translation accuracy than the subword-level NMT model using significantly fewer parameters, while demonstrating better capacity in learning longer-distance contextual and grammatical dependencies than the standard character-level NMT model.

This paper describes FBK’s submission to the end-to-end speech translation (ST) task at IWSLT 2019. The task consists in the “direct” translation (i.e. without intermediate discrete representation) of English speech data derived from TED Talks or lectures into German texts. Our participation had a twofold goal: i) testing our latest models, and ii) eval- uating the contribution to model training of different data augmentation techniques. On the model side, we deployed our recently proposed S-Transformer with logarithmic distance penalty, an ST-oriented adaptation of the Transformer architecture widely used in machine translation (MT). On the training side, we focused on data augmentation techniques recently proposed for ST and automatic speech recognition (ASR). In particular, we exploited augmented data in different ways and at different stages of the process. We first trained an end-to-end ASR system and used the weights of its encoder to initialize the decoder of our ST model (transfer learning). Then, we used an English-German MT system trained on large data to translate the English side of the English-French training set into German, and used this newly-created data as additional training material. Finally, we trained our models using SpecAugment, an augmentation technique that randomly masks portions of the spectrograms in order to make them different at every training epoch. Our synthetic corpus and SpecAugment resulted in an improvement of 5 BLEU points over our baseline model on the test set of MuST-C En-De, reaching the score of 22.3 with a single end-to-end system.

The recent advances introduced by neural machine translation (NMT) are rapidly expanding the application fields of machine translation, as well as reshaping the quality level to be targeted. In particular, if translations have to fit some given layout, quality should not only be measured in terms of adequacy and fluency, but also length. Exemplary cases are the translation of document files, subtitles, and scripts for dubbing, where the output length should ideally be as close as possible to the length of the input text. This pa-per addresses for the first time, to the best of our knowledge, the problem of controlling the output length in NMT. We investigate two methods for biasing the output length with a transformer architecture: i) conditioning the output to a given target-source length-ratio class and ii) enriching the transformer positional embedding with length information. Our experiments show that both methods can induce the network to generate shorter translations, as well as acquiring inter- pretable linguistic skills.

Neural text simplification has gained increasing attention in the NLP community thanks to recent advancements in deep sequence-to-sequence learning. Most recent efforts with such a data-demanding paradigm have dealt with the English language, for which sizeable training datasets are currently available to deploy competitive models. Similar improvements on less resource-rich languages are conditioned either to intensive manual work to create training data, or to the design of effective automatic generation techniques to bypass the data acquisition bottleneck. Inspired by the machine translation field, in which synthetic parallel pairs generated from monolingual data yield significant improvements to neural models, in this paper we exploit large amounts of heterogeneous data to automatically select simple sentences, which are then used to create synthetic simplification pairs. We also evaluate other solutions, such as oversampling and the use of external word embeddings to be fed to the neural simplification system. Our approach is evaluated on Italian and Spanish, for which few thousand gold sentence pairs are available. The results show that these techniques yield performance improvements over a baseline sequence-to-sequence configuration.

This paper describes FBK’s submission to the end-to-end English-German speech translation task at IWSLT 2018. Our system relies on a state-of-the-art model based on LSTMs and CNNs, where the CNNs are used to reduce the temporal dimension of the audio input, which is in general much higher than machine translation input. Our model was trained only on the audio-to-text parallel data released for the task, and fine-tuned on cleaned subsets of the original training corpus. The addition of weight normalization and label smoothing improved the baseline system by 1.0 BLEU point on our validation set. The final submission also featured checkpoint averaging within a training run and ensemble decoding of models trained during multiple runs. On test data, our best single model obtained a BLEU score of 9.7, while the ensemble obtained a BLEU score of 10.24.

The requirement for neural machine translation (NMT) models to use fixed-size input and output vocabularies plays an important role for their accuracy and generalization capability. The conventional approach to cope with this limitation is performing translation based on a vocabulary of sub-word units that are predicted using statistical word segmentation methods. However, these methods have recently shown to be prone to morphological errors, which lead to inaccurate translations. In this paper, we extend the source-language embedding layer of the NMT model with a bi-directional recurrent neural network that generates compositional representations of the source words from embeddings of character n-grams. Our model consistently outperforms conventional NMT with sub-word units on four translation directions with varying degrees of morphological complexity and data sparseness on the source side.

Recurrent neural networks (RNNs) have represented for years the state of the art in neural machine translation. Recently, new architectures have been proposed, which can leverage parallel computation on GPUs better than classical RNNs. Faster training and inference combined with different sequence-to-sequence modeling also lead to performance improvements. While the new models completely depart from the original recurrent architecture, we decided to investigate how to make RNNs more efficient. In this work, we propose a new recurrent NMT architecture, called Simple Recurrent NMT, built on a class of fast and weakly-recurrent units that use layer normalization and multiple attentions. Our experiments on the WMT14 English-to-German and WMT16 English-Romanian benchmarks show that our model represents a valid alternative to LSTMs, as it can achieve better results at a significantly lower computational cost.

Neural machine translation (NMT) is the state of the art for machine translation, and it shows the best performance when there is a considerable amount of data available. When only little data exist for a language pair, the model cannot produce good representations for words, particularly for rare words. One common solution consists in reducing data sparsity by segmenting words into sub-words, in order to allow rare words to have shared representations with other words. Taking a different approach, in this paper we present a method to feed an NMT network with word embeddings trained on monolingual data, which are combined with the task-specific embeddings learned at training time. This method can leverage an embedding matrix with a huge number of words, which can therefore extend the word-level vocabulary. Our experiments on two language pairs show good results for the typical low-resourced data scenario (IWSLT in-domain dataset). Our consistent improvements over the baselines represent a positive proof about the possibility to leverage models pre-trained on monolingual data in NMT.

In this paper, we describe FBK’s neural machine translation (NMT) systems submitted at the International Workshop on Spoken Language Translation (IWSLT) 2016. The systems are based on the state-of-the-art NMT architecture that is equipped with a bi-directional encoder and an attention mechanism in the decoder. They leverage linguistic information such as lemmas and part-of-speech tags of the source words in the form of additional factors along with the words. We compare performances of word and subword NMT systems along with different optimizers. Further, we explore different ensemble techniques to leverage multiple models within the same and across different networks. Several reranking methods are also explored. Our submissions cover all directions of the MSLT task, as well as en-{de, fr} and {de, fr}-en directions of TED. Compared to previously published best results on the TED 2014 test set, our models achieve comparable results on en-de and surpass them on en-fr (+2 BLEU) and fr-en (+7.7 BLEU) language pairs.