While large multilingual automatic speech recognition (ASR) models achieve remarkable performance, the internal mechanisms of the end-to-end pipeline, particularly concerning fairness and efficacy across languages, remain underexplored. This paper introduces a fine-grained analysis of Whisper’s multilingual decoder, examining its sub-token hypotheses during transcription across languages with various resource levels. Our method traces the beam search path, capturing sub-token guesses and their associated probabilities. Results reveal that higher resource languages benefit from higher likelihood of the correct token being top-ranked, greater confidence, lower predictive entropy, and more diverse alternative candidates. Lower resource languages fare worse on these metrics, but also exhibit distinct clustering patterns in sub-token usage sometimes influenced by typology in our PCA and t-SNE analysis. This sub-token probing uncovers systematic decoding disparities masked by aggregate error rates and points towards targeted interventions to ameliorate the imbalanced development of speech technology.

Code-switching, using multiple languages in a single utterance, is a common means of communication.In the language documentation process, speakers may code-switch between the target language and a language of broader communication; however, how to handle this mixed speech data is not always clearly addressed for speech research and specifically for a corpus phonetics pipeline.This paper investigates best practices for conducting phone-level forced alignment of code-switched field data using the Urum speech dataset from DoReCo. This dataset comprises 117 minutes of narrative utterances, of which 42% contain code-switched Urum–Russian speech.We demonstrate that the inclusion of Russian speech and Russian pretrained acoustic models can aid the alignment of Urum phones.Beyond using boundary alignment precision and accuracy metrics, we also discovered that the method of acoustic modeling impacted a downstream corpus phonetics investigation of code-switched Urum–Russian.

The development of Automatic Speech Recognition (ASR) has yielded impressive results, but its use in linguistic fieldwork remains limited. Recordings collected in fieldwork contexts present unique challenges, including spontaneous speech, environmental noise, and severely constrained datasets from under-documented languages. In this paper, we benchmark the performance of two fine-tuned multilingual ASR models, MMS and XLS-R, on five typologically diverse low-resource languages with control of training data duration. Our findings show that MMS is best suited when extremely small amounts of training data are available, whereas XLS-R shows parity performance once training data exceed one hour. We provide linguistically grounded analysis for further provide insights towards practical guidelines for field linguists, highlighting reproducible ASR adaptation approaches to mitigate the transcription bottleneck in language documentation.

This study investigates the impact of pitch flattening on automatic speech recognition (ASR) performance across tonal and non-tonal languages. Using vocoder-based signal processing techniques, we created pitch-flattened versions of speech recordings and compared ASR performance against original recordings. Results reveal that tonal languages experience substantially larger performance degradation than non-tonal languages. Analysis of tone confusion matrices shows systematic patterns of misidentification where contour tones collapse toward level tones when pitch information is removed. Calculation of tone’s functional load at syllable and word levels demonstrates that syllable-level functional load strongly predicts ASR vulnerability to pitch flattening, while word-level patterns reflect each language’s morphological structure. These findings illuminate the differential importance of pitch information across languages and suggest that ASR systems for languages with high syllable-level functional load require more robust pitch modeling.

Recent advancements in multilingual models for automatic speech recognition (ASR) have been able to achieve a high accuracy for languages with extremely limited resources. This study examines ASR modeling for the Mvskoke language, an indigenous language of America. The parameter efficiency of adapter training is contrasted with training entire models, and it is demonstrated how performance varies with different amounts of data. Additionally, the models are evaluated with trigram language model decoding, and the outputs are compared across different types of speech recordings. Results show that training an adapter is both parameter efficient and gives higher accuracy for a relatively small amount of data.

Significant research has focused on speaker recognition, determining which speaker is speaking in a segment of audio. However, few experiments have investigated speaker recognition for very low-resource or endangered languages. Furthermore, speaker recognition has the potential to support language documentation and revitalization efforts, making recordings more accessible to researchers and communities. Since endangered language datasets are too small to build competitive speaker representations from scratch, we investigate the application of large-scale pre-built speaker recognition models to bridge this gap. This paper compares four speaker recognition models on six diverse endangered language data sets. Comparisons contrast three recent neural network-based x-vector models and an earlier baseline i-vector model. Experiments demonstrate significantly stronger performance for some of the studied models. Further analysis highlights differences in effectiveness tied to the lengths of test audio segments and amount of data used for speaker modeling.

Cross-lingual emotion detection allows us to analyze global trends, public opinion, and social phenomena at scale. We participated in the Explainability of Cross-lingual Emotion Detection (EXALT) shared task, achieving an F1-score of 0.6046 on the evaluation set for the emotion detection sub-task. Our system outperformed the baseline by more than 0.16 F1-score absolute, and ranked second amongst competing systems. We conducted experiments using fine-tuning, zero-shot learning, and few-shot learning for Large Language Model (LLM)-based models as well as embedding-based BiLSTM and KNN for non-LLM-based techniques. Additionally, we introduced two novel methods: the Multi-Iteration Agentic Workflow and the Multi-Binary-Classifier Agentic Workflow. We found that LLM-based approaches provided good performance on multilingual emotion detection. Furthermore, ensembles combining all our experimented models yielded higher F1-scores than any single approach alone.

This paper introduces our submitted systems for WASSA 2024 Shared Task 2: Cross-Lingual Emotion Detection. We implemented a BERT-based classifier and an in-context learning-based system. Our best-performing model, using English Chain of Thought prompts with trigger words, reached 3rd overall with an F1 score of 0.6015. Following the motivation of the shared task, we further analyzed the scalability and transferability of the monolingual English dataset on cross-lingual tasks. Our analysis demonstrates the importance of data quality over quantity. We also found that augmented multilingual data does not necessarily perform better than English monolingual data in cross-lingual tasks. We open-sourced the augmented data and source code of our system for future research.

This paper describes our task 1 submission for the WASSA 2024 shared task on Explainability for Cross-lingual Emotion in Tweets. Our task is to predict the correct emotion label (Anger, Sadness, Fear, Joy, Love, and Neutral) for a dataset of English, Dutch, French, Spanish, and Russian tweets, while training exclusively on English emotion labeled data, to reveal what kind of emotion detection information is transferable cross-language (Maladry et al., 2024). To that end, we used an ensemble of models with a GPT-4 decider. Our ensemble consisted of a few-shot GPT-4 prompt system and a TwHIN-BERT system fine-tuned on the EXALT and additional English data. We ranked 8th place under the name WU_TLAXE with an F1 Macro score of 0.573 on the test set. We also experimented with an English-only TwHIN-BERT model by translating the other languages into English for inference, which proved to be worse than the other models.

We introduce a Masked Segmental Language Model (MSLM) for joint language modeling and unsupervised segmentation. While near-perfect supervised methods have been developed for segmenting human-like linguistic units in resource-rich languages such as Chinese, many of the world’s languages are both morphologically complex, and have no large dataset of “gold” segmentations for supervised training. Segmental Language Models offer a unique approach by conducting unsupervised segmentation as the byproduct of a neural language modeling objective. However, current SLMs are limited in their scalability due to their recurrent architecture. We propose a new type of SLM for use in both unsupervised and lightly supervised segmentation tasks. The MSLM is built on a span-masking transformer architecture, harnessing a masked bidirectional modeling context and attention, as well as adding the potential for model scalability. In a series of experiments, our model outperforms the segmentation quality of recurrent SLMs on Chinese, and performs similarly to the recurrent model on English.

Prosody is essential in human interaction, enabling people to show interest, establish rapport, efficiently convey nuances of attitude or intent, and so on. Some applications that exploit prosodic knowledge have recently shown superhuman performance, and in many respects our ability to effectively model prosody is rapidly advancing. This tutorial will overview the computational modeling of prosody, including recent advances and diverse actual and potential applications.

Basic-level categories have been shown to be both psychologically significant and useful in a wide range of practical applications. We build a rule-based system to identify basic-level categories in WordNet, achieving 77% accuracy on a test set derived from prior psychological experiments. With additional annotations we found our system also has low precision, in part due to the existence of many categories that do not fit into the three classes (superordinate, basic-level, and subordinate) relied on in basic-level category research.

We introduce a simple method for extracting non-arbitrary form-meaning representations from a collection of semantic vectors. We treat the problem as one of feature selection for a model trained to predict word vectors from subword features. We apply this model to the problem of automatically discovering phonesthemes, which are submorphemic sound clusters that appear in words with similar meaning. Many of our model-predicted phonesthemes overlap with those proposed in the linguistics literature, and we validate our approach with human judgments.

This paper addresses the problem of building conceptual resources for multilingual applications. We describe new techniques for large-scale construction of a Chinese-English lexicon for verbs, using thematic-role information to create links between Chinese and English conceptual information. We then present an approach to compensating for gaps in the existing resources. The resulting lexicon is used for multilingual applications such as machine translation and cross-language information retrieval.