Human evaluation is a critical component in machine translation system development and has received much attention in text translation research. However, little prior work exists on the topic of human evaluation for speech translation, which adds additional challenges such as noisy data and segmentation mismatches. We take the first steps to fill this gap by conducting a comprehensive human evaluation of the results of several shared tasks from the last International Workshop on Spoken Language Translation (IWSLT 2023). We propose an effective evaluation strategy based on automatic resegmentation and direct assessment with segment context. Our analysis revealed that: 1) the proposed evaluation strategy is robust and scores well-correlated with other types of human judgements; 2) automatic metrics are usually, but not always, well-correlated with direct assessment scores; and 3) COMET as a slightly stronger automatic metric than chrF, despite the segmentation noise introduced by the resegmentation step systems. We release the collected human-annotated data in order to encourage further investigation.

Code-switching (CS), i.e. mixing different languages in a single sentence, is a common phenomenon in communication and can be challenging in many Natural Language Processing (NLP) settings. Previous studies on CS speech have shown promising results for end-to-end speech translation (ST), but have been limited to offline scenarios and to translation to one of the languages present in the source monolingual transcription). In this paper, we focus on two essential yet unexplored areas for real-world CS speech translation: streaming settings, and translation to a third language (i.e., a language not included in the source). To this end, we extend the Fisher and Miami test and validation datasets to include new targets in Spanish and German. Using this data, we train a model for both offline and streaming ST and we establish baseline results for the two settings mentioned earlier.

Self-training has been shown to be helpful in addressing data scarcity for many domains, including vision, speech, and language. Specifically, self-training, or pseudo-labeling, labels unsupervised data and adds that to the training pool. In this work, we investigate and use pseudo-labeling for a recently proposed novel setup: joint transcription and translation of speech, which suffers from an absence of sufficient parallel data resources. We show that under such data-deficient circumstances, the unlabeled data can significantly vary in domain from the supervised data, which results in pseudo-label quality degradation. We investigate two categories of remedies that require no additional supervision and target the domain mismatch: pseudo-label filtering and data augmentation. We show that pseudo-label analysis and processing in this way results in additional gains on top of the vanilla pseudo-labeling setup providing a total improvement of up to 0.4% absolute WER and 2.1 BLEU points for En–De and 0.6% absolute WER and 2.2 BLEU points for En–Zh.

Behavioral testing in NLP allows fine-grained evaluation of systems by examining their linguistic capabilities through the analysis of input-output behavior. Unfortunately, existing work on behavioral testing in Machine Translation (MT) is currently restricted to largely handcrafted tests covering a limited range of capabilities and languages. To address this limitation, we propose to use Large Language Models (LLMs) to generate a diverse set of source sentences tailored to test the behavior of MT models in a range of situations. We can then verify whether the MT model exhibits the expected behavior through matching candidate sets that are also generated using LLMs. Our approach aims to make behavioral testing of MT systems practical while requiring only minimal human effort. In our experiments, we apply our proposed evaluation framework to assess multiple available MT systems, revealing that while in general pass-rates follow the trends observable from traditional accuracy-based metrics, our method was able to uncover several important differences and potential bugs that go unnoticed when relying only on accuracy.

This paper reports on the shared tasks organized by the 20th IWSLT Conference. The shared tasks address 9 scientific challenges in spoken language translation: simultaneous and offline translation, automatic subtitling and dubbing, speech-to-speech translation, multilingual, dialect and low-resource speech translation, and formality control. The shared tasks attracted a total of 38 submissions by 31 teams. The growing interest towards spoken language translation is also witnessed by the constantly increasing number of shared task organizers and contributors to the overview paper, almost evenly distributed across industry and academia.

Code switching (CS) refers to the phenomenon of interchangeably using words and phrases from different languages. CS can pose significant accuracy challenges to NLP, due to the often monolingual nature of the underlying systems. In this work, we focus on CS in the context of English/Spanish conversations for the task of speech translation (ST), generating and evaluating both transcript and translation. To evaluate model performance on this task, we create a novel ST corpus derived from existing public data sets. We explore various ST architectures across two dimensions: cascaded (transcribe then translate) vs end-to-end (jointly transcribe and translate) and unidirectional (source -> target) vs bidirectional (source <-> target). We show that our ST architectures, and especially our bidirectional end-to-end architecture, perform well on CS speech, even when no CS training data is used.

The evaluation campaign of the 19th International Conference on Spoken Language Translation featured eight shared tasks: (i) Simultaneous speech translation, (ii) Offline speech translation, (iii) Speech to speech translation, (iv) Low-resource speech translation, (v) Multilingual speech translation, (vi) Dialect speech translation, (vii) Formality control for speech translation, (viii) Isometric speech translation. A total of 27 teams participated in at least one of the shared tasks. This paper details, for each shared task, the purpose of the task, the data that were released, the evaluation metrics that were applied, the submissions that were received and the results that were achieved.

Using end-to-end models for speech translation (ST) has increasingly been the focus of the ST community. These models condense the previously cascaded systems by directly converting sound waves into translated text. However, cascaded models have the advantage of including automatic speech recognition output, useful for a variety of practical ST systems that often display transcripts to the user alongside the translations. To bridge this gap, recent work has shown initial progress into the feasibility for end-to-end models to produce both of these outputs. However, all previous work has only looked at this problem from the consecutive perspective, leaving uncertainty on whether these approaches are effective in the more challenging streaming setting. We develop an end-to-end streaming ST model based on a re-translation approach and compare against standard cascading approaches. We also introduce a novel inference method for the joint case, interleaving both transcript and translation in generation and removing the need to use separate decoders. Our evaluation across a range of metrics capturing accuracy, latency, and consistency shows that our end-to-end models are statistically similar to cascading models, while having half the number of parameters. We also find that both systems provide strong translation quality at low latency, keeping 99% of consecutive quality at a lag of just under a second.

Over its three decade history, speech translation has experienced several shifts in its primary research themes; moving from loosely coupled cascades of speech recognition and machine translation, to exploring questions of tight coupling, and finally to end-to-end models that have recently attracted much attention. This paper provides a brief survey of these developments, along with a discussion of the main challenges of traditional approaches which stem from committing to intermediate representations from the speech recognizer, and from training cascaded models separately towards different objectives. Recent end-to-end modeling techniques promise a principled way of overcoming these issues by allowing joint training of all model components and removing the need for explicit intermediate representations. However, a closer look reveals that many end-to-end models fall short of solving these issues, due to compromises made to address data scarcity. This paper provides a unifying categorization and nomenclature that covers both traditional and recent approaches and that may help researchers by highlighting both trade-offs and open research questions.

Variational Neural Machine Translation (VNMT) is an attractive framework for modeling the generation of target translations, conditioned not only on the source sentence but also on some latent random variables. The latent variable modeling may introduce useful statistical dependencies that can improve translation accuracy. Unfortunately, learning informative latent variables is non-trivial, as the latent space can be prohibitively large, and the latent codes are prone to be ignored by many translation models at training time. Previous works impose strong assumptions on the distribution of the latent code and limit the choice of the NMT architecture. In this paper, we propose to apply the VNMT framework to the state-of-the-art Transformer and introduce a more flexible approximate posterior based on normalizing flows. We demonstrate the efficacy of our proposal under both in-domain and out-of-domain conditions, significantly outperforming strong baselines.

The conventional paradigm in speech translation starts with a speech recognition step to generate transcripts, followed by a translation step with the automatic transcripts as input. To address various shortcomings of this paradigm, recent work explores end-to-end trainable direct models that translate without transcribing. However, transcripts can be an indispensable output in practical applications, which often display transcripts alongside the translations to users. We make this common requirement explicit and explore the task of jointly transcribing and translating speech. Although high accuracy of transcript and translation are crucial, even highly accurate systems can suffer from inconsistencies between both outputs that degrade the user experience. We introduce a methodology to evaluate consistency and compare several modeling approaches, including the traditional cascaded approach and end-to-end models. We find that direct models are poorly suited to the joint transcription/translation task, but that end-to-end models that feature a coupled inference procedure are able to achieve strong consistency. We further introduce simple techniques for directly optimizing for consistency, and analyze the resulting trade-offs between consistency, transcription accuracy, and translation accuracy.1

Spoken language translation applications for speech suffer due to conversational speech phenomena, particularly the presence of disfluencies. With the rise of end-to-end speech translation models, processing steps such as disfluency removal that were previously an intermediate step between speech recognition and machine translation need to be incorporated into model architectures. We use a sequence-to-sequence model to translate from noisy, disfluent speech to fluent text with disfluencies removed using the recently collected ‘copy-edited’ references for the Fisher Spanish-English dataset. We are able to directly generate fluent translations and introduce considerations about how to evaluate success on this task. This work provides a baseline for a new task, implicitly removing disfluencies in end-to-end translation of conversational speech.

Lattices are an efficient and effective method to encode ambiguity of upstream systems in natural language processing tasks, for example to compactly capture multiple speech recognition hypotheses, or to represent multiple linguistic analyses. Previous work has extended recurrent neural networks to model lattice inputs and achieved improvements in various tasks, but these models suffer from very slow computation speeds. This paper extends the recently proposed paradigm of self-attention to handle lattice inputs. Self-attention is a sequence modeling technique that relates inputs to one another by computing pairwise similarities and has gained popularity for both its strong results and its computational efficiency. To extend such models to handle lattices, we introduce probabilistic reachability masks that incorporate lattice structure into the model and support lattice scores if available. We also propose a method for adapting positional embeddings to lattice structures. We apply the proposed model to a speech translation task and find that it outperforms all examined baselines while being much faster to compute than previous neural lattice models during both training and inference.

Previous work on end-to-end translation from speech has primarily used frame-level features as speech representations, which creates longer, sparser sequences than text. We show that a naive method to create compressed phoneme-like speech representations is far more effective and efficient for translation than traditional frame-level speech features. Specifically, we generate phoneme labels for speech frames and average consecutive frames with the same label to create shorter, higher-level source sequences for translation. We see improvements of up to 5 BLEU on both our high and low resource language pairs, with a reduction in training time of 60%. Our improvements hold across multiple data sizes and two language pairs.

Paraphrases, rewordings of the same semantic meaning, are useful for improving generalization and translation. Unlike previous works that only explore paraphrases at the word or phrase level, we use different translations of the whole training data that are consistent in structure as paraphrases at the corpus level. We treat paraphrases as foreign languages, tag source sentences with paraphrase labels, and train on parallel paraphrases in the style of multilingual Neural Machine Translation (NMT). Our multi-paraphrase NMT that trains only on two languages outperforms the multilingual baselines. Adding paraphrases improves the rare word translation and increases entropy and diversity in lexical choice. Adding the source paraphrases boosts performance better than adding the target ones, while adding both lifts performance further. We achieve a BLEU score of 57.2 for French-to-English translation using 24 corpus-level paraphrases of the Bible, which outperforms the multilingual baselines and is +34.7 above the single-source single-target NMT baseline.

Speech translation has traditionally been approached through cascaded models consisting of a speech recognizer trained on a corpus of transcribed speech, and a machine translation system trained on parallel texts. Several recent works have shown the feasibility of collapsing the cascade into a single, direct model that can be trained in an end-to-end fashion on a corpus of translated speech. However, experiments are inconclusive on whether the cascade or the direct model is stronger, and have only been conducted under the unrealistic assumption that both are trained on equal amounts of data, ignoring other available speech recognition and machine translation corpora. In this paper, we demonstrate that direct speech translation models require more data to perform well than cascaded models, and although they allow including auxiliary data through multi-task training, they are poor at exploiting such data, putting them at a severe disadvantage. As a remedy, we propose the use of end- to-end trainable models with two attention mechanisms, the first establishing source speech to source text alignments, the second modeling source to target text alignment. We show that such models naturally decompose into multi-task–trainable recognition and translation tasks and propose an attention-passing technique that alleviates error propagation issues in a previous formulation of a model with two attention stages. Our proposed model outperforms all examined baselines and is able to exploit auxiliary training data much more effectively than direct attentional models.

This paper describes KIT’s submission to the IWSLT 2018 Translation task. We describe a system participating in the baseline condition and a system participating in the end-to-end condition. The baseline system is a cascade of an ASR system, a system to segment the ASR output and a neural machine translation system. We investigate the combination of different ASR systems. For the segmentation and machine translation components, we focused on transformer-based architectures.

In today’s globalized world we have the ability to communicate with people across the world. However, in many situations the language barrier still presents a major issue. For example, many foreign students coming to KIT to study are initially unable to follow a lecture in German. Therefore, we offer an automatic simultaneous interpretation service for students. To fulfill this task, we have developed a low-latency translation system that is adapted to lectures and covers several language pairs. While the switch from traditional Statistical Machine Translation to Neural Machine Translation (NMT) significantly improved performance, to integrate NMT into the speech translation framework required several adjustments. We have addressed the run-time constraints and different types of input. Furthermore, we utilized one-shot learning to easily add new topic-specific terms to the system. Besides better performance, NMT also enabled us increase our covered languages through multilingual NMT. % Combining these techniques, we are able to provide an adapted speech translation system for several European languages.

We work on translation from rich-resource languages to low-resource languages. The main challenges we identify are the lack of low-resource language data, effective methods for cross-lingual transfer, and the variable-binding problem that is common in neural systems. We build a translation system that addresses these challenges using eight European language families as our test ground. Firstly, we add the source and the target family labels and study intra-family and inter-family influences for effective cross-lingual transfer. We achieve an improvement of +9.9 in BLEU score for English-Swedish translation using eight families compared to the single-family multi-source multi-target baseline. Moreover, we find that training on two neighboring families closest to the low-resource language is often enough. Secondly, we construct an ablation study and find that reasonably good results can be achieved even with considerably less target data. Thirdly, we address the variable-binding problem by building an order-preserving named entity translation model. We obtain 60.6% accuracy in qualitative evaluation where our translations are akin to human translations in a preliminary study.

In this paper, we present KIT’s multilingual neural machine translation (NMT) systems for the IWSLT 2017 evaluation campaign machine translation (MT) and spoken language translation (SLT) tasks. For our MT task submissions, we used our multi-task system, modified from a standard attentional neural machine translation framework, instead of building 20 individual NMT systems. We investigated different architectures as well as different data corpora in training such a multilingual system. We also suggested an effective adaptation scheme for multilingual systems which brings great improvements compared to monolingual systems. For the SLT track, in addition to a monolingual neural translation system used to generate correct punctuations and true cases of the data prior to training our multilingual system, we introduced a noise model in order to make our system more robust. Results show that our novel modifications improved our systems considerably on all tasks.

This paper describes our German and English Speech-to-Text (STT) systems for the 2017 IWSLT evaluation campaign. The campaign focuses on the transcription of unsegmented lecture talks. Our setup includes systems using both the Janus and Kaldi frameworks. We combined the outputs using both ROVER [1] and confusion network combination (CNC) [2] to achieve a good overall performance. The individual subsystems are built by using different speaker-adaptive feature combination (e.g., lMEL with i-vector or bottleneck speaker vector), acoustic models (GMM or DNN) and speaker adaptation (MLLR or fMLLR). Decoding is performed in two stages, where the GMM and DNN systems are adapted on the combination of the first stage outputs using MLLR, and fMLLR. The combination setup produces a final hypothesis that has a significantly lower WER than any of the individual sub-systems. For the English lecture task, our best combination system has a WER of 8.3% on the tst2015 development set while our other combinations gained 25.7% WER for German lecture tasks.

Translating noisy inputs, such as the output of a speech recognizer, is a difficult but important challenge for neural machine translation. One way to increase robustness of neural models is by introducing artificial noise to the training data. In this paper, we experiment with appropriate forms of such noise, exploring a middle ground between general-purpose regularizers and highly task-specific forms of noise induction. We show that with a simple generative noise model, moderate gains can be achieved in translating erroneous speech transcripts, provided that type and amount of noise are properly calibrated. The optimal amount of noise at training time is much smaller than the amount of noise in our test data, indicating limitations due to trainability issues. We note that unlike our baseline model, models trained on noisy data are able to generate outputs of proper length even for noisy inputs, while gradually reducing output length for higher amount of noise, as might also be expected from a human translator. We discuss these findings in details and give suggestions for future work.

The input to a neural sequence-to-sequence model is often determined by an up-stream system, e.g. a word segmenter, part of speech tagger, or speech recognizer. These up-stream models are potentially error-prone. Representing inputs through word lattices allows making this uncertainty explicit by capturing alternative sequences and their posterior probabilities in a compact form. In this work, we extend the TreeLSTM (Tai et al., 2015) into a LatticeLSTM that is able to consume word lattices, and can be used as encoder in an attentional encoder-decoder model. We integrate lattice posterior scores into this architecture by extending the TreeLSTM’s child-sum and forget gates and introducing a bias term into the attention mechanism. We experiment with speech translation lattices and report consistent improvements over baselines that translate either the 1-best hypothesis or the lattice without posterior scores.

Speech that contains multimedia content can pose a serious challenge for real-time automatic speech recognition (ASR) for two reasons: (1) The ASR produces meaningless output, hurting the readability of the transcript. (2) The search space of the ASR is blown up when multimedia content is encountered, resulting in large delays that compromise real-time requirements. This paper introduces a segmenter that aims to remove these problems by detecting music and noise segments in real-time and replacing them with silence. We propose a two step approach, consisting of frame classification and smoothing. First, a classifier detects speech and multimedia on the frame level. In the second step the smoothing algorithm considers the temporal context to prevent rapid class fluctuations. We investigate in frame classification and smoothing settings to obtain an appealing accuracy-latency-tradeoff. The proposed segmenter yields increases the transcript quality of an ASR system by removing on average 39 % of the errors caused by non-speech in the audio stream, while maintaining a real-time applicable delay of 270 milliseconds.

In this paper, we present the KIT systems of the IWSLT 2016 machine translation evaluation. We participated in the machine translation (MT) task as well as the spoken language language translation (SLT) track for English→German and German→English translation. We use attentional neural machine translation (NMT) for all our submissions. We investigated different methods to adapt the system using small in-domain data as well as methods to train the system on these small corpora. In addition, we investigated methods to combine NMT systems that encode the input as well as the output differently. We combine systems using different vocabularies, reverse translation systems, multi-source translation system. In addition, we used pre-translation systems that facilitate phrase-based machine translation systems. Results show that applying domain adaptation and ensemble technique brings a crucial improvement of 3-4 BLEU points over the baseline system. In addition, system combination using n-best lists yields further 1-2 BLEU points.

This paper describes our German and English Speech-to-Text (STT) systems for the 2016 IWSLT evaluation campaign. The campaign focuses on the transcription of unsegmented TED talks. Our setup includes systems using both the Janus and Kaldi frameworks. We combined the outputs using both ROVER [1] and confusion network combination (CNC) [2] to archieve a good overall performance. The individual subsystems are built by using different speaker-adaptive feature combination (e.g., lMEL with i-vector or bottleneck speaker vector), acoustic models (GMM or DNN) and speaker adaption (MLLR or fMLLR). Decoding is performed in two stages, where the GMM and DNN systems are adapted on the combination of the first stage outputs using MLLR, and fMLLR. The combination setup produces a final hypothesis that has a significantly lower WER than any of the individual subsystems. For the English TED task, our best combination system has a WER of 7.8% on the development set while our other combinations gained 21.8% and 28.7% WERs for the English and German MSLT tasks.

Computer-assisted transcription promises high-quality speech transcription at reduced costs. This is achieved by limiting human effort to transcribing parts for which automatic transcription quality is insufficient. Our goal is to improve the human transcription quality via appropriate user interface design. We focus on iterative interfaces that allow humans to solve tasks based on an initially given suggestion, in this case an automatic transcription. We conduct a user study that reveals considerable quality gains for three variations of iterative interfaces over a non-iterative from-scratch transcription interface. Our iterative interfaces included post-editing, confidence-enhanced post-editing, and a novel retyping interface. All three yielded similar quality on average, but we found that the proposed retyping interface was less sensitive to the difficulty of the segment, and superior when the automatic transcription of the segment contained relatively many errors. An analysis using mixed-effects models allows us to quantify these and other factors and draw conclusions over which interface design should be chosen in which circumstance.

Evaluating the quality of output from language processing systems such as machine translation or speech recognition is an essential step in ensuring that they are sufficient for practical use. However, depending on the practical requirements, evaluation approaches can differ strongly. Often, reference-based evaluation measures (such as BLEU or WER) are appealing because they are cheap and allow rapid quantitative comparison. On the other hand, practitioners often focus on manual evaluation because they must deal with frequently changing domains and quality standards requested by customers, for which reference-based evaluation is insufficient or not possible due to missing in-domain reference data (Harris et al., 2016). In this paper, we attempt to bridge this gap by proposing a framework for lightly supervised quality estimation. We collect manually annotated scores for a small number of segments in a test corpus or document, and combine them with automatically predicted quality scores for the remaining segments to predict an overall quality estimate. An evaluation shows that our framework estimates quality more reliably than using fully automatic quality estimation approaches, while keeping annotation effort low by not requiring full references to be available for the particular domain.

In this paper, we study the problem of manually correcting automatic annotations of natural language in as efficient a manner as possible. We introduce a method for automatically segmenting a corpus into chunks such that many uncertain labels are grouped into the same chunk, while human supervision can be omitted altogether for other segments. A tradeoff must be found for segment sizes. Choosing short segments allows us to reduce the number of highly confident labels that are supervised by the annotator, which is useful because these labels are often already correct and supervising correct labels is a waste of effort. In contrast, long segments reduce the cognitive effort due to context switches. Our method helps find the segmentation that optimizes supervision efficiency by defining user models to predict the cost and utility of supervising each segment and solving a constrained optimization problem balancing these contradictory objectives. A user study demonstrates noticeable gains over pre-segmented, confidence-ordered baselines on two natural language processing tasks: speech transcription and word segmentation.

This paper describes our German, Italian and English Speech-to-Text (STT) systems for the 2014 IWSLT TED ASR track. Our setup uses ROVER and confusion network combination from various subsystems to achieve a good overall performance. The individual subsystems are built by using different front-ends, (e.g., MVDR-MFCC or lMel), acoustic models (GMM or modular DNN) and phone sets and by training on various subsets of the training data. Decoding is performed in two stages, where the GMM systems are adapted in an unsupervised manner on the combination of the first stage outputs using VTLN, MLLR, and cMLLR. The combination setup produces a final hypothesis that has a significantly lower WER than any of the individual subsystems.

This paper describes our English Speech-to-Text (STT) systems for the 2013 IWSLT TED ASR track. The systems consist of multiple subsystems that are combinations of different front-ends, e.g. MVDR-MFCC based and lMel based ones, GMM and NN acoustic models and different phone sets. The outputs of the subsystems are combined via confusion network combination. Decoding is done in two stages, where the systems of the second stage are adapted in an unsupervised manner on the combination of the first stage outputs using VTLN, MLLR, and cMLLR.

This paper describes our English Speech-to-Text (STT) systems for the 2012 IWSLT TED ASR track evaluation. The systems consist of 10 subsystems that are combinations of different front-ends, e.g. MVDR based and MFCC based ones, and two different phone sets. The outputs of the subsystems are combined via confusion network combination. Decoding is done in two stages, where the systems of the second stage are adapted in an unsupervised manner on the combination of the first stage outputs using VTLN, MLLR, and cM-LLR.

This paper describes the KIT-NAIST (Contrastive) English speech recognition system for the IWSLT 2012 Evaluation Campaign. In particular, we participated in the ASR track of the IWSLT TED task. The system was developed by Karlsruhe Institute of Technology (KIT) and Nara Institute of Science and Technology (NAIST) teams in collaboration within the interACT project. We employ single system decoding with fully continuous and semi-continuous models, as well as a three-stage, multipass system combination framework built with the Janus Recognition Toolkit. On the IWSLT 2010 test set our single system introduced in this work achieves a WER of 17.6%, and our final combination achieves a WER of 14.4%.