What do Speech Foundation Models Learn? Analysis and Applications

Speech foundation models (SFMs) are designed to serve as general-purpose representations for a wide range of speech-processing tasks. The last five years have seen an influx of increasingly successful self-supervised and supervised pre-trained models with impressive performance on various downstream tasks. Although the zoo of SFMs continues to grow, our understanding of the knowledge they acquire lags behind. This thesis presents a lightweight analysis framework using statistical tools and training-free tasks to investigate the acoustic and linguistic knowledge encoded in SFM layers. We conduct a comparative study across multiple SFMs and statistical tools. Our study also shows that the analytical insights have concrete implications for downstream task performance. The effectiveness of an SFM is ultimately determined by its performance on speech applications. Yet it remains unclear whether the benefits extend to spoken language understanding (SLU) tasks that require a deeper understanding than widely studied ones, such as speech recognition. The limited exploration of SLU is primarily due to a lack of relevant datasets. To alleviate that, this thesis contributes tasks, specifically spoken named entity recognition (NER) and named entity localization (NEL), to the Spoken Language Understanding Evaluation benchmark. We develop SFM-based approaches for NER and NEL, and find that end-to-end (E2E) models leveraging SFMs can surpass traditional cascaded (speech recognition followed by a text model) approaches. Further, we evaluate E2E SLU models across SFMs and adaptation strategies to assess the impact on task performance. Collectively, this thesis tackles previously unanswered questions about SFMs, providing tools and datasets to further our understanding and to enable the community to make informed design choices for future model development and adoption.

Fairness of Automatic Speech Recognition: Looking Through a Philosophical Lens

Automatic Speech Recognition (ASR) systems now mediate countless human-technology interactions, yet research on their fairness implications remains surprisingly limited. This paper examines ASR bias through a philosophical lens, arguing that systematic misrecognition of certain speech varieties constitutes more than a technical limitation -- it represents a form of disrespect that compounds historical injustices against marginalized linguistic communities. We distinguish between morally neutral classification (discriminate1) and harmful discrimination (discriminate2), demonstrating how ASR systems can inadvertently transform the former into the latter when they consistently misrecognize non-standard dialects. We identify three unique ethical dimensions of speech technologies that differentiate ASR bias from other algorithmic fairness concerns: the temporal burden placed on speakers of non-standard varieties ("temporal taxation"), the disruption of conversational flow when systems misrecognize speech, and the fundamental connection between speech patterns and personal/cultural identity. These factors create asymmetric power relationships that existing technical fairness metrics fail to capture. The paper analyzes the tension between linguistic standardization and pluralism in ASR development, arguing that current approaches often embed and reinforce problematic language ideologies. We conclude that addressing ASR bias requires more than technical interventions; it demands recognition of diverse speech varieties as legitimate forms of expression worthy of technological accommodation. This philosophical reframing offers new pathways for developing ASR systems that respect linguistic diversity and speaker autonomy.

Bridging ASR and LLMs for Dysarthric Speech Recognition: Benchmarking Self-Supervised and Generative Approaches

Speech Recognition (ASR) due to phoneme distortions and high variability. While self-supervised ASR models like Wav2Vec, HuBERT, and Whisper have shown promise, their effectiveness in dysarthric speech remains unclear. This study systematically benchmarks these models with different decoding strategies, including CTC, seq2seq, and LLM-enhanced decoding (BART,GPT-2, Vicuna). Our contributions include (1) benchmarking ASR architectures for dysarthric speech, (2) introducing LLM-based decoding to improve intelligibility, (3) analyzing generalization across datasets, and (4) providing insights into recognition errors across severity levels. Findings highlight that LLM-enhanced decoding improves dysarthric ASR by leveraging linguistic constraints for phoneme restoration and grammatical correction.

Depression diagnosis from patient interviews using multimodal machine learning

Background: Depression is a major public health concern, affecting an estimated five percent of the global population. Early and accurate diagnosis is essential to initiate effective treatment, yet recognition remains challenging in many clinical contexts. Speech, language, and behavioral cues collected during patient interviews may provide objective markers that support clinical assessment. Methods: We developed a diagnostic approach that integrates features derived from patient interviews, including speech patterns, linguistic characteristics, and structured clinical information. Separate models were trained for each modality and subsequently combined through multimodal fusion to reflect the complexity of real-world psychiatric assessment. Model validity was assessed with established performance metrics, and further evaluated using calibration and decision-analytic approaches to estimate potential clinical utility. Results: The multimodal model achieved superior diagnostic accuracy compared to single-modality models, with an AUROC of 0.88 and an F1-score of 0.75. Importantly, the fused model demonstrated good calibration and offered higher net clinical benefit compared to baseline strategies, highlighting its potential to assist clinicians in identifying patients with depression more reliably. Conclusion: Multimodal analysis of patient interviews using machine learning may serve as a valuable adjunct to psychiatric evaluation. By combining speech, language, and clinical features, this approach provides a robust framework that could enhance early detection of depressive disorders and support evidence-based decision-making in mental healthcare.

AD-AVSR: Asymmetric Dual-stream Enhancement for Robust Audio-Visual Speech Recognition

Audio-visual speech recognition (AVSR) combines audio-visual modalities to improve speech recognition, especially in noisy environments. However, most existing methods deploy the unidirectional enhancement or symmetric fusion manner, which limits their capability to capture heterogeneous and complementary correlations of audio-visual data-especially under asymmetric information conditions. To tackle these gaps, we introduce a new AVSR framework termed AD-AVSR based on bidirectional modality enhancement. Specifically, we first introduce the audio dual-stream encoding strategy to enrich audio representations from multiple perspectives and intentionally establish asymmetry to support subsequent cross-modal interactions. The enhancement process involves two key components, Audio-aware Visual Refinement Module for enhanced visual representations under audio guidance, and Cross-modal Noise Suppression Masking Module which refines audio representations using visual cues, collaboratively leading to the closed-loop and bidirectional information flow. To further enhance correlation robustness, we adopt a threshold-based selection mechanism to filter out irrelevant or weakly correlated audio-visual pairs. Extensive experimental results on the LRS2 and LRS3 datasets indicate that our AD-AVSR consistently surpasses SOTA methods in both performance and noise robustness, highlighting the effectiveness of our model design.

DeCRED: Decoder-Centric Regularization for Encoder-Decoder Based Speech Recognition

This paper presents a simple yet effective regularization for the internal language model induced by the decoder in encoder-decoder ASR models, thereby improving robustness and generalization in both in- and out-of-domain settings. The proposed method, Decoder-Centric Regularization in Encoder-Decoder (DeCRED), adds auxiliary classifiers to the decoder, enabling next token prediction via intermediate logits. Empirically, DeCRED reduces the mean internal LM BPE perplexity by 36.6% relative to 11 test sets. Furthermore, this translates into actual WER improvements over the baseline in 5 of 7 in-domain and 3 of 4 out-of-domain test sets, reducing macro WER from 6.4% to 6.3% and 18.2% to 16.2%, respectively. On TEDLIUM3, DeCRED achieves 7.0% WER, surpassing the baseline and encoder-centric InterCTC regularization by 0.6% and 0.5%, respectively. Finally, we compare DeCRED with OWSM v3.1 and Whisper-medium, showing competitive WERs despite training on much less data with fewer parameters.

CarelessWhisper: Turning Whisper into a Causal Streaming Model

Automatic Speech Recognition (ASR) has seen remarkable progress, with models like OpenAI Whisper and NVIDIA Canary achieving state-of-the-art (SOTA) performance in offline transcription. However, these models are not designed for streaming (online or real-time) transcription, due to limitations in their architecture and training methodology. We propose a method to turn the transformer encoder-decoder model into a low-latency streaming model that is careless about future context. We present an analysis explaining why it is not straightforward to convert an encoder-decoder transformer to a low-latency streaming model. Our proposed method modifies the existing (non-causal) encoder to a causal encoder by fine-tuning both the encoder and decoder using Low-Rank Adaptation (LoRA) and a weakly aligned dataset. We then propose an updated inference mechanism that utilizes the fine-tune causal encoder and decoder to yield greedy and beam-search decoding, and is shown to be locally optimal. Experiments on low-latency chunk sizes (less than 300 msec) show that our fine-tuned model outperforms existing non-fine-tuned streaming approaches in most cases, while using a lower complexity. Additionally, we observe that our training process yields better alignment, enabling a simple method for extracting word-level timestamps. We release our training and inference code, along with the fine-tuned models, to support further research and development in streaming ASR.

Landmark Guided Visual Feature Extractor for Visual Speech Recognition with Limited Resource

Visual speech recognition is a technique to identify spoken content in silent speech videos, which has raised significant attention in recent years. Advancements in data-driven deep learning methods have significantly improved both the speed and accuracy of recognition. However, these deep learning methods can be effected by visual disturbances, such as lightning conditions, skin texture and other user-specific features. Data-driven approaches could reduce the performance degradation caused by these visual disturbances using models pretrained on large-scale datasets. But these methods often require large amounts of training data and computational resources, making them costly. To reduce the influence of user-specific features and enhance performance with limited data, this paper proposed a landmark guided visual feature extractor. Facial landmarks are used as auxiliary information to aid in training the visual feature extractor. A spatio-temporal multi-graph convolutional network is designed to fully exploit the spatial locations and spatio-temporal features of facial landmarks. Additionally, a multi-level lip dynamic fusion framework is introduced to combine the spatio-temporal features of the landmarks with the visual features extracted from the raw video frames. Experimental results show that this approach performs well with limited data and also improves the model's accuracy on unseen speakers.

Pitch Accent Detection improves Pretrained Automatic Speech Recognition

We show the performance of Automatic Speech Recognition (ASR) systems that use semi-supervised speech representations can be boosted by a complimentary pitch accent detection module, by introducing a joint ASR and pitch accent detection model. The pitch accent detection component of our model achieves a significant improvement on the state-of-the-art for the task, closing the gap in F1-score by 41%. Additionally, the ASR performance in joint training decreases WER by 28.3% on LibriSpeech, under limited resource fine-tuning. With these results, we show the importance of extending pretrained speech models to retain or re-learn important prosodic cues such as pitch accent.

Revealing the Role of Audio Channels in ASR Performance Degradation

Pre-trained automatic speech recognition (ASR) models have demonstrated strong performance on a variety of tasks. However, their performance can degrade substantially when the input audio comes from different recording channels. While previous studies have demonstrated this phenomenon, it is often attributed to the mismatch between training and testing corpora. This study argues that variations in speech characteristics caused by different recording channels can fundamentally harm ASR performance. To address this limitation, we propose a normalization technique designed to mitigate the impact of channel variation by aligning internal feature representations in the ASR model with those derived from a clean reference channel. This approach significantly improves ASR performance on previously unseen channels and languages, highlighting its ability to generalize across channel and language differences.