As Large Language Models (LLMs) expand beyond text, integrating speech as a native modality has given rise to SpeechLLMs, which directly process spoken language and enable speech-to-text translation (ST) and other downstream tasks, bypassing traditional transcription-based pipelines. Whether this integration improves ST quality over established cascaded architectures, however, remains an open question. We present Hearing to Translate,1 the first comprehensive test suite rigorously benchmarking 6 state-of-the-art SpeechLLMs against 16 strong direct and cascade systems that couple leading speech foundation models (SFM), with multilingual LLMs. Our analysis spans 16 benchmarks, 13 language pairs, and 9 challenging conditions, including disfluent, noisy, and long-form speech. Across this extensive evaluation, we find that cascaded systems remain the most reliable solution overall, but most recent SpeechLLMs can match or even outperform cascades in various settings while SFMs lag behind both, highlighting that integrating an LLM, either within the model or in a pipeline, is essential for high-quality speech translation.

The challenges of building speech-to-text translation (ST) systems (e.g., a relative lack of parallel speech–text data and robustness to noise in audio) are exacerbated for low-resource language pairs. In this work, we seek to improve low-resource ST by building on previous studies that regularize ST training with the connectionist temporal classification (CTC) loss. By systematically evaluating a diverse range of linguistic annotations as CTC labels across multiple auxiliary loss configurations, we improve speech translation systems for both low- and high-resource settings. These improvements over both a standard end-to-end ST system and a speech LLM indicate a need for continued research on regularizing speech representations in ST.

The five idioms (i.e., varieties) of the Romansh language are largely standardized and are taught in the schools of the respective communities in Switzerland. In this paper, we present the first parallel corpus of Romansh idioms. The corpus is based on 291 schoolbook volumes, which are comparable in content for the five idioms. We use automatic alignment methods to extract 207k multi-parallel segments from the books, with more than 2M tokens in total. A small-scale human evaluation confirms that the segments are highly parallel, making the dataset suitable for NLP applications such as machine translation between Romansh idioms. We release the parallel and unaligned versions of the dataset under a CC-BY-NC-SA license and demonstrate its utility for machine translation by training and evaluating an LLM and a supervised multilingual MT model on the dataset.

Contextual and unigram surprisal are standard predictors of reading time in computational psycholinguistics. While contextual surprisal is typically derived from neural language models, unigram surprisal is usually obtained from independent corpus counts. Although both quantities are deterministic functionals of a single underlying distribution—capturing conditional and marginal structure respectively—this practice breaks their probabilistic link and may introduce estimation error from the mismatch. We propose three principles for estimating unigram surprisal in psycholinguistic research, including the requirement that both quantities be derived from the same language model. We operationalize this principle by introducing a Monte Carlo estimator of unigram surprisal from a language model, along with a marginalized variant that has lower variance empirically. When unigram surprisal is re-estimated under this unified framework, we find evidence that predictors change in their predictive contribution to reading time. We also confirm prior findings that frequency can be a stronger predictor of processing effort than contextual surprisal under this unified framework.

Although language is predominantly spoken, empirical tests of surprisal theory rely almost entirely on written text. In practice, surprisal is estimated using language models (LMs) trained on text and applied to transcriptions, creating a text bottleneck that limits evaluation of the theory in the spoken modality. We remove this bottleneck by estimating surprisal directly from the speech signal using a speech LM. This raises a key question: should surprisal be defined over realization-specific acoustic inputs or over abstract linguistic units? Since a single unit can have many distinct acoustic realizations, each assigned a different probability by the model, defining surprisal over abstract units is difficult. To address this, we introduce an importance sampling estimator that approximates the probability of an abstract unit by marginalizing over realizations generated by a text-to-speech model. Using pupillometry data from a listening study, we evaluate surprisal theory in this text-free setting at multiple levels of granularity. Obtaining surprisal from several speech LMs, we consistently find that surprisal estimated via importance sampling better predicts pupil dilation than realization-specific surprisal for words. We also find that surprisal values for sub-lexical units are predictive of pupil dilation. While we find that surprisal can be estimated without recourse to text, and that marginalizing over distinct realizations of words is more predictive of lexical processing effort, surprisal as a whole is not a strong predictor of change in pupil size, regardless of its source.

To accurately transcribe a speech signal, automatic speech recognition (ASR) systems must show robustness to a wide range of task independent variation, such as speaker factors, recording quality, or even ädversarial noisedesigned to disrupt performance.We manipulated the dialect composition of fine-tuning data for ASR to study whether balancing the relative proportion of dialects had an impact on models robustness to two such sources of variation”:" dialect variation and adversarial perturbations. We fine-tuned XLSR-53 for Catalan ASR using four different dialect compositions, each containing the Central Catalan dialect. These were defined as 100%, 80%, 50%, and 20% Central Catalan, with the remaining portions split evenly between four other Catalan dialects. While increasing the relative proportion of dialect variants improved models’ dialect robustness, this did not have a meaningful impact on adversarial robustness. These findings suggest that while improvements to ASR can be made by diversifying the training data, such changes do not sufficiently counteract adversarial attacks, leaving the technology open to security threats.

The distinction between function and content units of the lexicon has been somewhat neglected in recent NLP work, but it could still be useful when working with low-resource languages, and, in particular, to improve cross-lingual transfer. In this paper, we investigate to what extent BPE subword tokenization can be used to identify units of the functional lexicon in a language without any annotated data. We analyze subword tokens in terms of their productivity and attempt to find thresholds that best distinguish function from content tokens. On a sample of seven diverse languages, we find that the best results are obtained with 50 BPE merges. We also show that this subword tokenization setting can be beneficial for the interlinear glossing task.

Effectively normalizing spellings in textual data poses a considerable challenge, especially for low-resource languages lacking standardized writing systems. In this study, we fine-tuned a multilingual model with data from several Occitan dialects and conducted a series of experiments to assess the model’s representations of these dialects. For evaluation purposes, we compiled a parallel lexicon encompassing four Occitan dialects.Intrinsic evaluations of the model’s embeddings revealed that surface similarity between the dialects strengthened representations. When the model was further fine-tuned for part-of-speech tagging, its performance was robust to dialectical variation, even when trained solely on part-of-speech data from a single dialect. Our findings suggest that large multilingual models minimize the need for spelling normalization during pre-processing.

This paper presents the system description of the NordicsAlps team for the AmericasNLP 2024 Machine Translation Shared Task 1. We investigate the effect of tokenization on translation quality by exploring two different tokenization schemes: byte-level and redundancy-driven tokenization. We submitted three runs per language pair. The redundancy-driven tokenization ranked first among all submissions, scoring the highest average chrF2++, chrF, and BLEU metrics (averaged across all languages). These findings demonstrate the importance of carefully tailoring the tokenization strategies of machine translation systems, particularly in resource-constrained scenarios.