Token-based multitasking frameworks like TokenVerse require all training utterances to have labels for all tasks, hindering their ability to leverage partially annotated datasets and scale effectively. We propose TokenVerse++, which introduces learnable vectors in the acoustic embedding space of the XLSR-Transducer ASR model for dynamic task activation. This core mechanism enables training with utterances labeled for only a subset of tasks, a key advantage over TokenVerse. We demonstrate this by successfully integrating a dataset with partial labels, specifically for ASR and an additional task, language identification, improving overall performance. TokenVerse++ achieves results on par with or exceeding TokenVerse across multiple tasks, establishing it as a more practical multitask alternative without sacrificing ASR performance.

Self-supervised pretrained models exhibit competitive performance in automatic speech recognition (ASR) on finetuning, even with limited in-domain supervised data. However, popular pretrained models are not suitable for streaming ASR because they are trained with full attention context. In this paper, we introduce XLSR-Transducer, where the XLSR-53 model is used as encoder in transducer setup. Our experiments on the AMI dataset reveal that the XLSR-Transducer achieves 4% absolute WER improvement over Whisper large-v2 and 8% over a Zipformer transducer model trained from scratch. To enable streaming capabilities, we investigate different attention masking patterns in the self-attention computation of transformer layers within the XLSR-53 model. We validate XLSR-Transducer on AMI and 5 languages from CommonVoice under low-resource scenarios. Finally, with the introduction of attention sinks, we reduce the left context by half while achieving a relative 12% improvement in WER.

In real-world speech data processing, the scarcity of annotated data and the abundance of unlabelled speech data present a significant challenge. To address this, we propose an efficient data selection pipeline for fine-tuning ASR models by generating pseudo-labels using WhisperX pipeline and selecting efficient labels for fine-tuning. In our work, we propose a domain classifier system developed with a computationally inexpensive TFIDF and classical machine learning algorithm. Later, we filter data from the classifier output using a novel metric that assesses word ratio and perplexity distribution. The filtered pseudo labels are then used for fine-tuning standard encoder-decoder Whisper models and Zipformer. Our proposed data selection pipeline reduces the dataset size by approximately 1/100thwhile maintaining performance comparable to the full dataset, outperforming random domain-independent selection strategies.

Fine-tuning pretrained ASR models for specific domains is challenging when labeled data is scarce. But unlabeled audio and labeled data from related domains are often available. We propose an incremental semi-supervised learning pipeline that first integrates a small in-domain labeled set and an auxiliary dataset from a closely related domain, achieving a relative improvement of 4% over no auxiliary data. Filtering based on multi-model consensus or named entity recognition (NER) is then applied to select and iteratively refine pseudo-labels, showing slower performance saturation compared to random selection. Evaluated on the multi-domain Wow call center and Fisher English corpora, it outperforms single-step fine-tuning. Consensus-based filtering outperforms other methods, providing up to 22.3% relative improvement on Wow and 24.8% on Fisher over single-step fine-tuning with random selection. NER is the second-best filter, providing competitive performance at a lower computational cost.

Fine-tuning pretrained ASR models for specific domains is challenging for small organizations with limited labeled data and computational resources. Here, we explore different data selection pipelines and propose a robust approach that improves ASR adaptation by filtering pseudo-labels generated using Whisper (encoder-decoder) and Zipformer (transducer) models. Our approach integrates multiple selection strategies -- including word error rate (WER) prediction, named entity recognition (NER), and character error rate (CER) analysis -- to extract high-quality training segments. We evaluate our method on Whisper and Zipformer using a 7500-hour baseline, comparing it to a CER-based approach relying on hypotheses from three ASR systems. Fine-tuning on 7500 hours of pseudo-labeled call center data achieves 12.3% WER, while our filtering reduces the dataset to 100 hours (1.4%) with similar performance; a similar trend is observed on Fisher English.

Shashi Kumar, Srikanth Madikeri, Juan Pablo Zuluaga Gomez, Iuliia Thorbecke, Esaú Villatoro-tello, Sergio Burdisso, Petr Motlicek, Karthik Pandia D S, Aravind Ganapathiraju. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Traditionally, automatic speech recognition (ASR) and speaker change detection (SCD) systems have been independently trained to generate comprehensive transcripts accompanied by speaker turns. Recently, joint training of ASR and SCD systems, by inserting speaker turn tokens in the ASR training text, has been shown to be successful. In this work, we present a multitask alternative to the joint training approach. Results obtained on the mix-headset audios of AMI corpus show that the proposed multitask training yields an absolute improvement of 1.8% in coverage and purity based F1 score on SCD task without ASR degradation. We also examine the trade-offs between the ASR and SCD performance when trained using multitask criteria. Additionally, we validate the speaker change information in the embedding spaces obtained after different transformer layers of a self-supervised pre-trained model, such as XLSR-53, by integrating an SCD classifier at the output of specific transformer layers. Results reveal that the use of different embedding spaces from XLSR-53 model for multitask ASR and SCD is advantageous.1

Spoken Language Understanding (SLU) technologies have greatly improved due to the effective pretraining of speech representations. A common requirement of industry-based solutions is the portability to deploy SLU models in voice-assistant devices. Thus, distilling knowledge from large text-based language models has become an attractive solution for achieving good performance and guaranteeing portability. In this paper, we introduce a novel architecture that uses a cross-modal attention mechanism to extract bin-level contextual embeddings from a word-confusion network (WNC) encoding such that these can be directly compared and aligned with traditional text-based contextual embeddings. This alignment is achieved using a recently proposed tokenwise constrastive loss function. We validate our architecture’s effectiveness by fine-tuning our WCN-based pretrained model to do intent classification (IC) on the well-known SLURP dataset. Obtained accuracy on the IC task (81%), depicts a 9.4% relative improvement compared to a recent/equivalent E2E method.

Sergio Burdisso, Dairazalia Sanchez-cortes, Esaú Villatoro-tello, Petr Motlicek. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

In this paper, we perform an exhaustive evaluation of different representations to address the intent classification problem in a Spoken Language Understanding (SLU) setup. We benchmark three types of systems to perform the SLU intent detection task: 1) text-based, 2) lattice-based, and a novel 3) multimodal approach. Our work provides a comprehensive analysis of what could be the achievable performance of different state-of-the-art SLU systems under different circumstances, e.g., automatically- vs. manually-generated transcripts. We evaluate the systems on the publicly available SLURP spoken language resource corpus. Our results indicate that using richer forms of Automatic Speech Recognition (ASR) outputs, namely word-consensus-networks, allows the SLU system to improve in comparison to the 1-best setup (5.5% relative improvement). However, crossmodal approaches, i.e., learning from acoustic and text embeddings, obtains performance similar to the oracle setup, a relative improvement of 17.8% over the 1-best configuration, being a recommended alternative to overcome the limitations of working with automatically generated transcripts.

We propose a simple approach for weighting selfconnecting edges in a Graph Convolutional Network (GCN) and show its impact on depression detection from transcribed clinical interviews.To this end, we use a GCN for modeling non-consecutive and long-distance semantics to classify the transcriptions into depressed or control subjects.The proposed method aims to mitigate the limiting assumptions of locality and the equal importance of self-connections vs. edges to neighboring nodes in GCNs, while preserving attractive features such as low computational cost, data agnostic, and interpretability capabilities.We perform an exhaustive evaluation in two benchmark datasets.Results show that our approach consistently outperforms the vanilla GCN model as well as previously reported results, achieving an F1=0.84 on both datasets.Finally, a qualitative analysis illustrates the interpretability capabilities of the proposed approach and its alignment with previous findings in psychology.

In this paper, we evaluate different alternatives to process richer forms of Automatic Speech Recognition (ASR) output based on lattice expansion algorithms for Spoken Document Retrieval (SDR). Typically, SDR systems employ ASR transcripts to index and retrieve relevant documents. However, ASR errors negatively affect the retrieval performance. Multiple alternative hypotheses can also be used to augment the input to document retrieval to compensate for the erroneous one-best hypothesis. In Weighted Finite State Transducer-based ASR systems, using the n-best output (i.e. the top "n'' scoring hypotheses) for the retrieval task is common, since they can easily be fed to a traditional Information Retrieval (IR) pipeline. However, the n-best hypotheses are terribly redundant, and do not sufficiently encapsulate the richness of the ASR output, which is represented as an acyclic directed graph called the lattice. In particular, we utilize the lattice's constrained minimum path cover to generate a minimum set of hypotheses that serve as input to the reranking phase of IR. The novelty of our proposed approach is the incorporation of the lattice as an input for neural reranking by considering a set of hypotheses that represents every arc in the lattice. The obtained hypotheses are encoded through sentence embeddings using BERT-based models, namely SBERT and RoBERTa, and the final ranking of the retrieved segments is obtained with a max-pooling operation over the computed scores among the input query and the hypotheses set. We present our evaluation on the publicly available AMI meeting corpus. Our results indicate that the proposed use of hypotheses from the expanded lattice improves the SDR performance significantly over the n-best ASR output.

Depression disorder is one of the major causes of disability in the world that can lead to tragic outcomes. In this paper, we propose a method for using an approximation to a mental lexicon to model the communication process of depressed and non-depressed participants in spontaneous North American English clinical interviews. Our approach, inspired by the Lexical Availability theory, identifies the most relevant vocabulary of the interviewed participant, and use it as features in a classification process. We performed an in-depth evaluation on the DAIC-WOZ [20] and the E-DAIC [11] clinical datasets. Obtained results indicate that our approach can compete against recent contextual embeddings when modeling and identifying depression. We show the generalization capabilities of our algorithm using outside data, reaching a macro F1 = 0.83 and F1 = 0.80 in the DAIC-WOZ and E-DAIC datasets respectively. An analysis of our method’s interpretability allows understanding how the classifier is making its decisions. During this process, we observed strong connections between our obtained results and previous research from the psychological field.

Mental disorders, e.g. depression and dementia, are categorized as priority conditions according to the World Health Organization (WHO). When diagnosing, psychologists employ structured questionnaires/interviews, and different cognitive tests. Although accurate, there is an increasing necessity of developing digital mental health support technologies to alleviate the burden faced by professionals. In this paper, we propose a multi-modal approach for modeling the communication process employed by patients being part of a clinical interview or a cognitive test. The language-based modality, inspired by the Lexical Availability (LA) theory from psycho-linguistics, identifies the most accessible vocabulary of the interviewed subject and use it as features in a classification process. The acoustic-based modality is processed by a Convolutional Neural Network (CNN) trained on signals of speech that predominantly contained voice source characteristics. In the end, a late fusion technique, based on majority voting, assigns the final classification. Results show the complementarity of both modalities, reaching an overall Macro-F1 of 84% and 90% for Depression and Alzheimer's dementia respectively.

Abstract This paper introduces a novel ranking refinement approach based on relevance feedback for the task of document retrieval. We focus on the problem of ranking refinement since recent evaluation results from Information Retrieval (IR) systems indicate that current methods are effective retrieving most of the relevant documents for different sets of queries, but they have severe difficulties to generate a pertinent ranking of them. Motivated by these results, we propose a novel method to re-rank the list of documents returned by an IR system. The proposed method is based on a Markov Random Field (MRF) model that classifies the retrieved documents as relevant or irrelevant. The proposed MRF combines: (i) information provided by the base IR system, (ii) similarities among documents in the retrieved list, and (iii) relevance feedback information. Thus, the problem of ranking refinement is reduced to that of minimising an energy function that represents a trade-off between document relevance and inter-document similarity. Experiments were conducted using resources from four different tasks of the Cross Language Evaluation Forum (CLEF) forum as well as from one task of the Text Retrieval Conference (TREC) forum. The obtained results show the feasibility of the method for re-ranking documents in IR and also depict an improvement in mean average precision compared to a state of the art retrieval machine.