Self-supervised learning (SSL) has helped extend speech technologies to more languages by reducing the need for labeled data. However, models are still far from supporting the world’s 7000+ languages. We propose XEUS, a Cross-lingual Encoder for Universal Speech, trained on over 1 million hours of data across 4057 languages, extending the language coverage of SSL models 4-fold. We combine 1 million hours of speech from existing publicly accessible corpora with a newly created corpus of 7400+ hours from 4057 languages, which will be publicly released. To handle the diverse conditions of multilingual speech data, we augment the typical SSL masked prediction approach with a novel dereverberation objective, increasing robustness. We evaluate XEUS on several benchmarks, and show that it consistently outperforms or achieves comparable results to state-of-the-art (SOTA) SSL models across a variety of tasks. XEUS sets a new SOTA on the ML-SUPERB benchmark: it outperforms MMS 1B and w2v-BERT 2.0 v2 by 0.8% and 4.4% respectively, despite having less parameters or pre-training data. Checkpoints, code, and data are found in https://www.wavlab.org/activities/2024/xeus/.

The Spoken Language Understanding Evaluation (SLUE) suite of benchmark tasks was recently introduced to address the need for openresources and benchmarking of complex spoken language understanding (SLU) tasks, including both classification and sequence generation tasks, on natural speech. The benchmark has demonstrated preliminary success in using pre-trained speech foundation models (SFM) for these SLU tasks. However, the community still lacks a fine-grained understanding of the comparative utility of different SFMs. Inspired by this, we ask: which SFMs offer the most benefits for these complex SLU tasks, and what is the most effective approach for incorporating these SFMs? To answer this, we perform an extensive evaluation of multiple supervised and self-supervised SFMs using several evaluation protocols: (i) frozen SFMs with a lightweight prediction head, (ii) frozen SFMs with a complex prediction head, and (iii) fine-tuned SFMs with a lightweight prediction head. Although the supervised SFMs are pre-trained on much more speech recognition data (with labels), they do not always outperform self-supervised SFMs; the latter tend to perform at least as well as, and sometimes better than, supervised SFMs, especially on the sequence generation tasks in SLUE. While there is no universally optimal way of incorporating SFMs, the complex prediction head gives the best performance for most tasks, although it increases the inference time. We also introduce an open-source toolkit and performance leaderboard, SLUE-PERB, for these tasks and modeling strategies.

This paper reports on the shared tasks organized by the 21st IWSLT Conference. The shared tasks address 7 scientific challenges in spoken language translation: simultaneous and offline translation, automatic subtitling and dubbing, speech-to-speech translation, dialect and low-resource speech translation, and Indic languages. The shared tasks attracted 17 teams whose submissions are documented in 27 system papers. 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.

This article describes the QUESPA team speech translation (ST) submissions for the Quechua to Spanish (QUE–SPA) track featured in the Evaluation Campaign of IWSLT 2024: dialectal and low-resource speech translation. Two main submission types were supported in the campaign: constrained and unconstrained. This is our second year submitting our ST systems to the IWSLT shared task and we feel that we have achieved novel performance, surpassing last year’s submissions. Again, we were able to submit six total systems of which our best (primary) constrained system consisted of an ST model based on the Fairseq S2T framework where the audio representations were created using log mel-scale filter banks as features and the translations were performed using a transformer. The system was similar to last year’s submission with slight configuration changes, allowing us to achieve slightly higher performance (2 BLEU). Contrastingly, we were able to achieve much better performance than last year on the unconstrained task using a larger pre-trained language (PLM) model for ST (without cascading) and the inclusion of parallel QUE–SPA data found on the internet. The fine-tuning of Microsoft’s SpeechT5 model in a ST setting along with the addition of new data and a data augmentation technique allowed us to achieve 19.7 BLEU. Additionally, we present the other four submissions (2 constrained and 2 unconstrained) which are part of additional efforts of hyper-parameter and configuration tuning on existent models and the inclusion of Whisper for speech recognition

This paper describes CMU’s submission to the IWSLT 2024 Simultaneous Speech Translation (SST) task for translating English speech to German text in a streaming manner. Our end-to-end speech-to-text (ST) system integrates the WavLM speech encoder, a modality adapter, and the Llama2-7B-Base model as the decoder. We employ a two-stage training approach: initially, we align the representations of speech and text, followed by full fine-tuning. Both stages are trained on MuST-c v2 data with cross-entropy loss. We adapt our offline ST model for SST using a simple fixed hold-n policy. Experiments show that our model obtains an offline BLEU score of 31.1 and a BLEU score of 29.5 under 2 seconds latency on the MuST-C-v2 tst-COMMON.

This work describes CMU’s submission to the IWSLT 2024 Offline Speech Translation (ST) Shared Task for translating English speech to German, Chinese, and Japanese text. We are the first participants to employ a long-form strategy which directly processes unsegmented recordings without the need for a separate voice-activity detection stage (VAD). We show that the Whisper automatic speech recognition (ASR) model has a hallucination problem when applied out-of-the-box to recordings containing non-speech noises, but a simple noisy fine-tuning approach can greatly enhance Whisper’s long-form robustness across multiple domains. Then, we feed English ASR outputs into fine-tuned NLLB machine translation (MT) models which are decoded using COMET-based Minimum Bayes Risk. Our VAD-free ASR+MT cascade is tested on TED talks, TV series, and workout videos and shown to outperform prior winning IWSLT submissions and large open-source models.

The application of self-supervision to speech representation learning has garnered significant interest in recent years, due to its scalability to large amounts of unlabeled data. However, much progress, both in terms of pre-training and downstream evaluation, has remained concentrated in monolingual models that only consider English. Few models consider other languages, and even fewer consider indigenous ones. In this work, benchmark the efficacy of large SSL models on 6 indigenous America languages: Quechua, Guarani , Bribri, Kotiria, Wa’ikhana, and Totonac on low-resource ASR. Our results show surprisingly strong performance by state-of-the-art SSL models, showing the potential generalizability of large-scale models to real-world data.

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.

This paper describes CMU’s submission to the IWSLT 2023 simultaneous speech translation shared task for translating English speech to both German text and speech in a streaming fashion. We first build offline speech-to-text (ST) models using the joint CTC/attention framework. These models also use WavLM front-end features and mBART decoder initialization. We adapt our offline ST models for simultaneous speech-to-text translation (SST) by 1) incrementally encoding chunks of input speech, re-computing encoder states for each new chunk and 2) incrementally decoding output text, pruning beam search hypotheses to 1-best after processing each chunk. We then build text-to-speech (TTS) models using the VITS framework and achieve simultaneous speech-to-speech translation (SS2ST) by cascading our SST and TTS models.

This article describes the QUESPA team speech translation (ST) submissions for the Quechua to Spanish (QUE–SPA) track featured in the Evaluation Campaign of IWSLT 2023: low-resource and dialect speech translation. Two main submission types were supported in the campaign: constrained and unconstrained. We submitted six total systems of which our best (primary) constrained system consisted of an ST model based on the Fairseq S2T framework where the audio representations were created using log mel-scale filter banks as features and the translations were performed using a transformer. The best (primary) unconstrained system used a pipeline approach which combined automatic speech recognition (ASR) with machine translation (MT). The ASR transcriptions for the best unconstrained system were computed using a pre-trained XLS-R-based model along with a fine-tuned language model. Transcriptions were translated using a MT system based on a fine-tuned, pre-trained language model (PLM). The four other submissions are presented in this article (2 constrained and 2 unconstrained) for comparison because they consist of various architectures. Our results show that direct ST (ASR and MT combined together) can be more effective than a PLM in a low-resource (constrained) setting for Quechua to Spanish. On the other hand, we show that fine-tuning of any type on both the ASR and MT system is worthwhile, resulting in nearly 16 BLEU for the unconstrained task.

This paper provides an overview of the first shared task on choosing beneficial instances for machine translation, conducted as part of the CoCo4MT 2023 Workshop at MTSummit. This shared task was motivated by the need to make the data annotation process for machine translation more efficient, particularly for low-resource languages for which collecting human translations may be difficult or expensive. The task involved developing methods for selecting the most beneficial instances for training a machine translation system without access to an existing parallel dataset in the target language, such that the best selected instances can then be manually translated. Two teams participated in the shared task, namely the Williams team and the AST team. Submissions were evaluated by training a machine translation model on each submission’s chosen instances, and comparing their performance with the chRF++ score. The system that ranked first is by the Williams team, that finds representative instances by clustering the training data.

The lack of resources for languages in the Americas has proven to be a problem for the creation of digital systems such as machine translation, search engines, chat bots, and more. The scarceness of digital resources for a language causes a higher impact on populations where the language is spoken by millions of people. We introduce the first official large combined corpus for deep learning of an indigenous South American low-resource language spoken by millions called Quechua. Specifically, our curated corpus is created from text gathered from the southern region of Peru where a dialect of Quechua is spoken that has not traditionally been used for digital systems as a target dialect in the past. In order to make our work repeatable by others, we also offer a public, pre-trained, BERT model called QuBERT which is the largest linguistic model ever trained for any Quechua type, not just the southern region dialect. We furthermore test our corpus and its corresponding BERT model on two major tasks: (1) named-entity recognition (NER) and (2) part-of-speech (POS) tagging by using state-of-the-art techniques where we achieve results comparable to other work on higher-resource languages. In this article, we describe the methodology, challenges, and results from the creation of QuBERT which is on par with other state-of-the-art multilingual models for natural language processing achieving between 71 and 74% F1 score on NER and 84–87% on POS tasks.

Neural Machine Translation (NMT) for Low Resource Languages (LRL) is often limited by the lack of available training data, making it necessary to explore additional techniques to improve translation quality. We propose the use of the Prefix-Root-Postfix-Encoding (PRPE) subword segmentation algorithm to improve translation quality for LRLs, using two agglutinative languages as case studies: Quechua and Indonesian. During the course of our experiments, we reintroduce a parallel corpus for Quechua-Spanish translation that was previously unavailable for NMT. Our experiments show the importance of appropriate subword segmentation, which can go as far as improving translation quality over systems trained on much larger quantities of data. We show this by achieving state-of-the-art results for both languages, obtaining higher BLEU scores than large pre-trained models with much smaller amounts of data.

We present the University of Central Florida systems for the LoResMT 2021 Shared Task, participating in the English-Irish and English-Marathi translation pairs. We focused our efforts on constrained track of the task, using transfer learning and subword segmentation to enhance our models given small amounts of training data. Our models achieved the highest BLEU scores on the fully constrained tracks of English-Irish, Irish-English, and Marathi-English with scores of 13.5, 21.3, and 17.9 respectively