Abstract--Speech Foundation Models have gained significant attention recently. Prior works have shown that the fusion of representations from multiple layers of the same model or the fusion of multiple models can improve performance on downstream tasks. We unify these two fusion strategies by proposing an interface module that enables fusion across multiple upstream speech models while integrating information across their layers. We conduct extensive experiments on different self-supervised and supervised models across various speech tasks, including ASR and paralinguistic analysis, and demonstrate that our method outperforms prior fusion approaches. We further analyze its scalability concerning model size and count, highlighting the importance of selecting appropriate upstream models. Our results show that the proposed interface provides an additional performance boost when given a suitable upstream model selection, making it a promising approach for utilizing Speech Foundation Models. Personal use of this material is permitted.

Speech foundation models have significantly advanced various speech-related tasks by providing exceptional representation capabilities. However, their high-dimensional output features often create a mismatch with downstream task models, which typically require lower-dimensional inputs. A common solution is to apply a dimensionality reduction (DR) layer, but this approach increases parameter overhead, computational costs, and risks losing valuable information. To address these issues, we propose Nested Res2Net (Nes2Net), a lightweight back-end architecture designed to directly process high-dimensional features without DR layers. The nested structure enhances multi-scale feature extraction, improves feature interaction, and preserves high-dimensional information. We first validate Nes2Net on CtrSVDD, a singing voice deepfake detection dataset, and report a 22% performance improvement and an 87% back-end computational cost reduction over the state-of-the-art baseline. Additionally, extensive testing across four diverse datasets: ASVspoof 2021, ASVspoof 5, PartialSpoof, and In-the-Wild, covering fully spoofed speech, adversarial attacks, partial spoofing, and real-world scenarios, consistently highlights Nes2Net's superior robustness and generalization capabilities. The code package and pre-trained models are available at https://github.com/Liu-Tianchi/Nes2Net.

ABSTRACT Large speech foundation models achieve strong performance across many domains, but they often require adaptation to handle local needs such as code-switching, where speakers mix languages within the same utterance. To address this challenge, we explore Bayesian factorized adapters for speech foundation models, which place priors near zero to achieve sparser adaptation matrices and thereby retain general performance while adapting to specific domains. We apply our approach to the Whisper model and evaluate on different multilingual code-switching scenarios. Our results show only minimal adaptation loss while significantly reducing catastrophic forgetting of the base model. Compared to LoRA, our method achieves a backward gain of 54% with only a 4% drop on the new domain.

We propose an unsupervised adaptation framework, Self-T Aught Recognizer (ST AR), which leverages unlabeled data to enhance the robustness of automatic speech recognition (ASR) systems in diverse target domains, such as noise and accents. ST AR is developed for prevalent speech foundation models based on Transformer-related architecture with auto-regressive decoding (e.g., Whisper, Canary; SeamlessM4T).

Automatic Speech Recognition (ASR) systems in real-world settings need to handle imperfect audio, often degraded by hardware limitations or environmental noise, while accommodating diverse user groups. In human-robot interaction (HRI), these challenges intersect to create a uniquely challenging recognition environment. We evaluate four state-of-the-art ASR systems on eight publicly available datasets that capture six dimensions of difficulty: domain-specific, accented, noisy, age-variant, impaired, and spontaneous speech. Our analysis demonstrates significant variations in performance, hallucination tendencies, and inherent biases, despite similar scores on standard benchmarks. These limitations have serious implications for HRI, where recognition errors can interfere with task performance, user trust, and safety.

Specifically, we report comprehensive benchmark evaluations on dialects and regional language varieties in English, Arabic, Mandarin and Cantonese, Tibetan, Indic languages, Thai, Spanish, French, German, Brazilian Portuguese, and Italian. Our study used over 2 million training utterances from 30 publicly available speech corpora that are provided with dialectal information. We evaluate the performance of several widely used speech foundation models in classifying speech dialects. We assess the robustness of the dialectal models under noisy conditions and present an error analysis that highlights modeling results aligned with geographic continuity. In addition to benchmarking dialect classification, we demonstrate several downstream applications enabled by Voxlect . Specifically, we show that Voxlect can be applied to augment existing speech recognition datasets with dialect information, enabling a more detailed analysis of ASR performance across dialectal variations. Voxlect is also used as a tool to evaluate the performance of speech generation systems.

The SENCOTEN language, spoken on the Saanich peninsula of southern Vancouver Island, is in the midst of vigorous language revitalization efforts to turn the tide of language loss as a result of colonial language policies. To support these on-the-ground efforts, the community is turning to digital technology. Automatic Speech Recognition (ASR) technology holds great promise for accelerating language documentation and the creation of educational resources. However, developing ASR systems for SENCOTEN is challenging due to limited data and significant vocabulary variation from its polysynthetic structure and stress-driven metathesis. To address these challenges, we propose an ASR-driven documentation pipeline that leverages augmented speech data from a text-to-speech (TTS) system and cross-lingual transfer learning with Speech Foundation Models (SFMs). An n-gram language model is also incorporated via shallow fusion or n-best restoring to maximize the use of available data. Experiments on the SENCOTEN dataset show a word error rate (WER) of 19.34% and a character error rate (CER) of 5.09% on the test set with a 57.02% out-of-vocabulary (OOV) rate. After filtering minor cedilla-related errors, WER improves to 14.32% (26.48% on unseen words) and CER to 3.45%, demonstrating the potential of our ASR-driven pipeline to support SENCOTEN language documentation.

Cross-lingual alignment in pretrained language models (LMs) has enabled efficient transfer in text-based LMs. Such an alignment has also been observed in speech foundation models. However, it remains an open question whether findings and methods from text-based cross-lingual alignment apply to speech. Building on prior work on spoken translation retrieval, we perform pronunciation-controlled experiments to observe if cross-lingual alignment can indeed occur in such models on a semantic basis, instead of relying on phonetic similarities. Our findings indicate that even in the absence of phonetic cues, spoken translation retrieval accuracy remains relatively stable. We follow up with a controlled experiment on a word-level dataset of cross-lingual synonyms and near-homophones, confirming the existence of both phonetic and semantic knowledge in the encoder. Finally, we qualitatively examine the transcriptions produced by early exiting the encoder, where we observe that speech translation produces semantic errors that are characterized by phonetic similarities to corresponding words in the source language. We apply this insight from early exiting to speech recognition in seven low-resource languages unsupported by the Whisper model, and achieve improved accuracy in all languages examined, particularly for languages with transparent orthographies.

We propose an unsupervised adaptation framework, Self-TAught Recognizer (STAR), which leverages unlabeled data to enhance the robustness of automatic speech recognition (ASR) systems in diverse target domains, such as noise and accents. STAR is developed for prevalent speech foundation models based on Transformer-related architecture with auto-regressive decoding (e.g., Whisper, Canary). Specifically, we propose a novel indicator that empirically integrates step-wise information during decoding to assess the token-level quality of pseudo labels without ground truth, thereby guiding model updates for effective unsupervised adaptation. Experimental results show that STAR achieves an average of 13.5% relative reduction in word error rate across 14 target domains, and it sometimes even approaches the upper-bound performance of supervised adaptation. Surprisingly, we also observe that STAR prevents the adapted model from the common catastrophic forgetting problem without recalling source-domain data. Furthermore, STAR exhibits high data efficiency that only requires less than one-hour unlabeled data, and seamless generality to alternative large speech models and speech translation tasks.