Erik Visser

Arvind Krishna Sridhar, Erik Visser

Advancement in large pretrained language models has significantly improved their performance for conditional language generation tasks including summarization albeit with hallucinations. To reduce hallucinations, conventional methods proposed improving beam search or using a fact checker as a postprocessing step. In this paper, we investigate the use of the Natural Language Inference (NLI) entailment metric to detect and prevent hallucinations in summary generation. We propose an NLI-assisted beam re-ranking mechanism by computing entailment probability scores between the input context and summarization model-generated beams during saliency-enhanced greedy decoding. Moreover, a diversity metric is introduced to compare its effectiveness against vanilla beam search. Our proposed algorithm significantly outperforms vanilla beam decoding on XSum and CNN/DM datasets.

Arvind Krishna Sridhar, Yinyi Guo, Erik Visser

The Audio Question Answering (AQA) task includes audio event classification, audio captioning, and open-ended reasoning. Recently, AQA has garnered attention due to the advent of Large Audio Language Models (LALMs). Current literature focuses on constructing LALMs by integrating audio encoders with text-only Large Language Models (LLMs) through a projection module. While LALMs excel in general audio understanding, they are limited in temporal reasoning, which may hinder their commercial applications and on-device deployment. This paper addresses these challenges and limitations in audio temporal reasoning. First, we introduce a data augmentation technique for generating reliable audio temporal questions and answers using an LLM. Second, we perform a further fine-tuning of an existing baseline using curriculum learning strategy to specialize in temporal reasoning without compromising performance on fine-tuned tasks. We demonstrate the performance of our model using state-of-the-art LALMs on public audio benchmark datasets. Third, we implement our AQA model on-device locally and investigate its CPU inference for edge applications.

Arvind Krishna Sridhar, Yinyi Guo, Erik Visser et al.

There has been significant research on developing pretrained transformer architectures for multimodal-to-text generation tasks. Albeit performance improvements, such models are frequently overparameterized, hence suffer from hallucination and large memory footprint making them challenging to deploy on edge devices. In this paper, we address both these issues for the application of automated audio captioning. First, we propose a data augmentation technique for generating hallucinated audio captions and show that similarity based on an audio-text shared latent space is suitable for detecting hallucination. Then, we propose a parameter efficient inference time faithful decoding algorithm that enables smaller audio captioning models with performance equivalent to larger models trained with more data. During the beam decoding step, the smaller model utilizes an audio-text shared latent representation to semantically align the generated text with corresponding input audio. Faithful guidance is introduced into the beam probability by incorporating the cosine similarity between latent representation projections of greedy rolled out intermediate beams and audio clip. We show the efficacy of our algorithm on benchmark datasets and evaluate the proposed scheme against baselines using conventional audio captioning and semantic similarity metrics while illustrating tradeoffs between performance and complexity.

Ravi Choudhary, Arvind Krishna Sridhar, Erik Visser

In the era of loT (Internet of Things) we are surrounded by a plethora of Al enabled devices that can transcribe images, video, audio, and sensors signals into text descriptions. When such transcriptions are captured in activity reports for monitoring, life logging and anomaly detection applications, a user would typically request a summary or ask targeted questions about certain sections of the report they are interested in. Depending on the context and the type of question asked, a question answering (QA) system would need to automatically determine whether the answer covers single-span or multi-span text components. Currently available QA datasets primarily focus on single span responses only (such as SQuAD[4]) or contain a low proportion of examples with multiple span answers (such as DROP[3]). To investigate automatic selection of single/multi-span answers in the use case described, we created a new smart home environment dataset comprised of questions paired with single-span or multi-span answers depending on the question and context queried. In addition, we propose a RoBERTa[6]-based multiple span extraction question answering (MSEQA) model returning the appropriate answer span for a given question. Our experiments show that the proposed model outperforms state-of-the-art QA models on our dataset while providing comparable performance on published individual single/multi-span task datasets.

Hyungseob Lim, Kyungguen Byun, Sunkuk Moon et al.

While many recent any-to-any voice conversion models succeed in transferring some target speech's style information to the converted speech, they still lack the ability to faithfully reproduce the speaking style of the target speaker. In this work, we propose a novel method to extract rich style information from target utterances and to efficiently transfer it to source speech content without requiring text transcriptions or speaker labeling. Our proposed approach introduces an attention mechanism utilizing a self-supervised learning (SSL) model to collect the speaking styles of a target speaker each corresponding to the different phonetic content. The styles are represented with a set of embeddings called stylebook. In the next step, the stylebook is attended with the source speech's phonetic content to determine the final target style for each source content. Finally, content information extracted from the source speech and content-dependent target style embeddings are fed into a diffusion-based decoder to generate the converted speech mel-spectrogram. Experiment results show that our proposed method combined with a diffusion-based generative model can achieve better speaker similarity in any-to-any voice conversion tasks when compared to baseline models, while the increase in computational complexity with longer utterances is suppressed.

Mine Kerpicci, Van Nguyen, Shuhua Zhang et al.

Model architectures such as wav2vec 2.0 and HuBERT have been proposed to learn speech representations from audio waveforms in a self-supervised manner. When they are combined with downstream tasks such as keyword spotting and speaker verification, they provide state-of-the-art performance. However, these models use a large number of parameters, the smallest version of which has 95 million parameters. This constitutes a challenge for edge AI device deployments. In this paper, we investigate the application of knowledge distillation to speech representation learning (SRL) models followed by joint fine-tuning with multiple downstream voice-activated tasks. In our experiments on two such tasks, our approach results in nearly 75% reduction in model size while suffering only 0.1% accuracy and 0.9% equal error rate degradation compared to the full-size model. In addition, we show that fine-tuning the SRL models results in a significant performance boost compared to using frozen SRL models.

Naveen Vakada, Kartik Hegde, Arvind Krishna Sridhar et al.

Long-duration audio is increasingly common in industrial and consumer settings, yet reviewing multi-hour recordings is impractical, motivating systems that answer natural-language queries with precise temporal grounding and minimal hallucination. Existing audio-language models show promise, but long-audio question answering remains difficult due to context-length limits. We introduce LongAudio-RAG (LA-RAG), a hybrid framework that grounds Large Language Model (LLM) outputs in retrieved, timestamped acoustic event detections rather than raw audio. Multi-hour streams are converted into structured event records stored in an SQL database, and at inference time the system resolves natural-language time references, classifies intent, retrieves only the relevant events, and generates answers using this constrained evidence. To evaluate performance, we construct a synthetic long-audio benchmark by concatenating recordings with preserved timestamps and generating template-based question-answer pairs for detection, counting, and summarization tasks. Finally, we demonstrate the practicality of our approach by deploying it in a hybrid edge-cloud environment, where the audio grounding model runs on-device on IoT-class hardware while the LLM is hosted on a GPU-backed server. This architecture enables low-latency event extraction at the edge and high-quality language reasoning in the cloud. Experiments show that structured, event-level retrieval significantly improves accuracy compared to vanilla Retrieval-Augmented Generation (RAG) or text-to-SQL approaches.

Rehana Mahfuz, Yinyi Guo, Erik Visser et al.

Real-time conversational assistants for procedural tasks often depend on video input, which can be computationally expensive and compromise user privacy. For the first time, we propose a real-time conversational assistant that provides comprehensive guidance for a procedural task using only lightweight privacy-preserving modalities such as audio and IMU inputs from a user's wearable device to understand the context. This assistant proactively communicates step-by-step instructions to a user performing a furniture assembly task, and answers user questions. We construct a dataset containing conversations where the assistant guides the user in performing the task. On observing that an off-the-shelf language model is a very talkative assistant, we design a novel User Whim Agnostic (UWA) LoRA finetuning method which improves the model's ability to suppress less informative dialogues, while maintaining its tendency to communicate important instructions. This leads to >30% improvement in the F-score. Finetuning the model also results in a 16x speedup by eliminating the need to provide in-context examples in the prompt. We further describe how such an assistant is implemented on edge devices with no dependence on the cloud.

Arvind Krishna Sridhar, Yinyi Guo, Erik Visser

Spatial audio understanding aims to enable machines to interpret complex auditory scenes, particularly when sound sources move over time. In this work, we study Spatial Audio Question Answering (Spatial AQA) with a focus on movement reasoning, where a model must infer object motion, position, and directional changes directly from stereo audio. First, we introduce a movement-centric spatial audio augmentation framework that synthesizes diverse motion patterns from isolated mono audio events, enabling controlled and scalable training data generation. Second, we propose an end-to-end multimodal finetuning approach with a thinking mode, which allows audio-language models to produce explicit intermediate reasoning steps before predicting an answer. Third, we investigate the impact of query-conditioned source separation as a preprocessing stage and compare three inference regimes: no masking, an audio grounding model (AGM), and ground-truth masks. Our results show that reasoning amplifies the benefits of source separation, with thinking mode showing significant improvement of +5.1% when a single event is present in the question. These findings highlight the interplay between movement modeling, reasoning, and separation quality, offering new insights for advancing spatial audio understanding.

Arvind Krishna Sridhar, Yinyi Guo, Erik Visser

Spatial audio reasoning enables machines to interpret auditory scenes by understanding events and their spatial attributes. In this work, we focus on spatial audio understanding with an emphasis on reasoning about moving sources. First, we introduce a spatial audio encoder that processes spatial audio to detect multiple overlapping events and estimate their spatial attributes, Direction of Arrival (DoA) and source distance, at the frame level. To generalize to unseen events, we incorporate an audio grounding model that aligns audio features with semantic audio class text embeddings via a cross-attention mechanism. Second, to answer complex queries about dynamic audio scenes involving moving sources, we condition a large language model (LLM) on structured spatial attributes extracted by our model. Finally, we introduce a spatial audio motion understanding and reasoning benchmark dataset and demonstrate our framework's performance against the baseline model.

Kartik Hegde, Rehana Mahfuz, Yinyi Guo et al.

Current audio captioning systems rely heavily on supervised learning with paired audio-caption datasets, which are expensive to curate and may not reflect human preferences in real-world scenarios. To address this limitation, we propose a preference-aligned audio captioning framework based on Reinforcement Learning from Human Feedback (RLHF). To effectively capture nuanced human preferences, we train a Contrastive Language-Audio Pretraining (CLAP)-based reward model using human-labeled pairwise preference data. This reward model is integrated into a reinforcement learning framework to fine-tune any baseline captioning system without relying on ground-truth caption annotations. Extensive human evaluations across multiple datasets show that our method produces captions preferred over those from baseline models, particularly in cases where the baseline models fail to provide correct and natural captions. Furthermore, our framework achieves performance comparable to supervised approaches with ground-truth data, demonstrating its effectiveness in aligning audio captioning with human preferences and its scalability in real-world scenarios.

Miseul Kim, Soo Jin Park, Kyungguen Byun et al.

Speaker diarization systems often struggle with high intrinsic intra-speaker variability, such as shifts in emotion, health, or content. This can cause segments from the same speaker to be misclassified as different individuals, for example, when one raises their voice or speaks faster during conversation. To address this, we propose a style-controllable speech generation model that augments speech across diverse styles while preserving the target speaker's identity. The proposed system starts with diarized segments from a conventional diarizer. For each diarized segment, it generates augmented speech samples enriched with phonetic and stylistic diversity. And then, speaker embeddings from both the original and generated audio are blended to enhance the system's robustness in grouping segments with high intrinsic intra-speaker variability. We validate our approach on a simulated emotional speech dataset and the truncated AMI dataset, demonstrating significant improvements, with error rate reductions of 49% and 35% on each dataset, respectively.

Shehzeen Hussain, Van Nguyen, Shuhua Zhang et al.

Self-supervised learning methods such as wav2vec 2.0 have shown promising results in learning speech representations from unlabelled and untranscribed speech data that are useful for speech recognition. Since these representations are learned without any task-specific supervision, they can also be useful for other voice-activated tasks like speaker verification, keyword spotting, emotion classification etc. In our work, we propose a general purpose framework for adapting a pre-trained wav2vec 2.0 model for different voice-activated tasks. We develop downstream network architectures that operate on the contextualized speech representations of wav2vec 2.0 to adapt the representations for solving a given task. Finally, we extend our framework to perform multi-task learning by jointly optimizing the network parameters on multiple voice activated tasks using a shared transformer backbone. Both of our single and multi-task frameworks achieve state-of-the-art results in speaker verification and keyword spotting benchmarks. Our best performing models achieve 1.98% and 3.15% EER on VoxCeleb1 test set when trained on VoxCeleb2 and VoxCeleb1 respectively, and 98.23% accuracy on Google Speech Commands v1.0 keyword spotting dataset.

Eunjeong Koh, Fatemeh Saki, Yinyi Guo et al.

This paper presents a new learning strategy for the Sound Event Detection (SED) system to tackle the issues of i) knowledge migration from a pre-trained model to a new target model and ii) learning new sound events without forgetting the previously learned ones without re-training from scratch. In order to migrate the previously learned knowledge from the source model to the target one, a neural adapter is employed on the top of the source model. The source model and the target model are merged via this neural adapter layer. The neural adapter layer facilitates the target model to learn new sound events with minimal training data and maintaining the performance of the previously learned sound events similar to the source model. Our extensive analysis on the DCASE16 and US-SED dataset reveals the effectiveness of the proposed method in transferring knowledge between source and target models without introducing any performance degradation on the previously learned sound events while obtaining a competitive detection performance on the newly learned sound events.