Soham Deshmukh

Hira Dhamyal, Benjamin Elizalde, Soham Deshmukh et al.

Emotions lie on a broad continuum and treating emotions as a discrete number of classes limits the ability of a model to capture the nuances in the continuum. The challenge is how to describe the nuances of emotions and how to enable a model to learn the descriptions. In this work, we devise a method to automatically create a description (or prompt) for a given audio by computing acoustic properties, such as pitch, loudness, speech rate, and articulation rate. We pair a prompt with its corresponding audio using 5 different emotion datasets. We trained a neural network model using these audio-text pairs. Then, we evaluate the model using one more dataset. We investigate how the model can learn to associate the audio with the descriptions, resulting in performance improvement of Speech Emotion Recognition and Speech Audio Retrieval. We expect our findings to motivate research describing the broad continuum of emotion

Soham Deshmukh, Benjamin Elizalde, Huaming Wang

Audio-Text retrieval takes a natural language query to retrieve relevant audio files in a database. Conversely, Text-Audio retrieval takes an audio file as a query to retrieve relevant natural language descriptions. Most of the literature train retrieval systems with one audio captioning dataset, but evaluating the benefit of training with multiple datasets is underexplored. Moreover, retrieval systems have to learn the alignment between elaborated sentences describing audio content of variable length ranging from a few seconds to several minutes. In this work, we propose a new collection of web audio-text pairs and a new framework for retrieval. First, we provide a new collection of about five thousand web audio-text pairs that we refer to as WavText5K. When used to train our retrieval system, WavText5K improved performance more than other audio captioning datasets. Second, our framework learns to connect language and audio content by using a text encoder, two audio encoders, and a contrastive learning objective. Combining both audio encoders helps to process variable length audio. The two contributions beat state of the art performance for AudioCaps and Clotho on Text-Audio retrieval by a relative 2% and 16%, and Audio-Text retrieval by 6% and 23%.

Muhammad Ahmed Shah, Roshan Sharma, Hira Dhamyal et al.

It has been shown that Large Language Model (LLM) alignments can be circumvented by appending specially crafted attack suffixes with harmful queries to elicit harmful responses. To conduct attacks against private target models whose characterization is unknown, public models can be used as proxies to fashion the attack, with successful attacks being transferred from public proxies to private target models. The success rate of attack depends on how closely the proxy model approximates the private model. We hypothesize that for attacks to be transferrable, it is sufficient if the proxy can approximate the target model in the neighborhood of the harmful query. Therefore, in this paper, we propose \emph{Local Fine-Tuning (LoFT)}, \textit{i.e.}, fine-tuning proxy models on similar queries that lie in the lexico-semantic neighborhood of harmful queries to decrease the divergence between the proxy and target models. First, we demonstrate three approaches to prompt private target models to obtain similar queries given harmful queries. Next, we obtain data for local fine-tuning by eliciting responses from target models for the generated similar queries. Then, we optimize attack suffixes to generate attack prompts and evaluate the impact of our local fine-tuning on the attack's success rate. Experiments show that local fine-tuning of proxy models improves attack transferability and increases attack success rate by $39\%$, $7\%$, and $0.5\%$ (absolute) on target models ChatGPT, GPT-4, and Claude respectively.

Hira Dhamyal, Benjamin Elizalde, Soham Deshmukh et al.

Emotions lie on a continuum, but current models treat emotions as a finite valued discrete variable. This representation does not capture the diversity in the expression of emotion. To better represent emotions we propose the use of natural language descriptions (or prompts). In this work, we address the challenge of automatically generating these prompts and training a model to better learn emotion representations from audio and prompt pairs. We use acoustic properties that are correlated to emotion like pitch, intensity, speech rate, and articulation rate to automatically generate prompts i.e. 'acoustic prompts'. We use a contrastive learning objective to map speech to their respective acoustic prompts. We evaluate our model on Emotion Audio Retrieval and Speech Emotion Recognition. Our results show that the acoustic prompts significantly improve the model's performance in EAR, in various Precision@K metrics. In SER, we observe a 3.8% relative accuracy improvement on the Ravdess dataset.

Soham Deshmukh, Charles Lee

Intent detection is a key part of any Natural Language Understanding (NLU) system of a conversational assistant. Detecting the correct intent is essential yet difficult for email conversations where multiple directives and intents are present. In such settings, conversation context can become a key disambiguating factor for detecting the user's request from the assistant. One prominent way of incorporating context is modeling past conversation history like task-oriented dialogue models. However, the nature of email conversations (long form) restricts direct usage of the latest advances in task-oriented dialogue models. So in this paper, we provide an effective transfer learning framework (EMToD) that allows the latest development in dialogue models to be adapted for long-form conversations. We show that the proposed EMToD framework improves intent detection performance over pre-trained language models by 45% and over pre-trained dialogue models by 30% for task-oriented email conversations. Additionally, the modular nature of the proposed framework allows plug-and-play for any future developments in both pre-trained language and task-oriented dialogue models.

Soham Deshmukh, Satvik Dixit, Rita Singh et al. · cmu

Multimodal Audio-Language Models (ALMs) can understand and reason over both audio and text. Typically, reasoning performance correlates with model size, with the best results achieved by models exceeding 8 billion parameters. However, no prior work has explored enabling small audio-language models to perform reasoning tasks, despite the potential applications for edge devices. To address this gap, we introduce Mellow, a small Audio-Language Model specifically designed for reasoning. Mellow achieves state-of-the-art performance among existing small audio-language models and surpasses several larger models in reasoning capabilities. For instance, Mellow scores 52.11 on MMAU, comparable to SoTA Qwen2 Audio (which scores 52.5) while using 50 times fewer parameters and being trained on 60 times less data (audio hrs). To train Mellow, we introduce ReasonAQA, a dataset designed to enhance audio-grounded reasoning in models. It consists of a mixture of existing datasets (30% of the data) and synthetically generated data (70%). The synthetic dataset is derived from audio captioning datasets, where Large Language Models (LLMs) generate detailed and multiple-choice questions focusing on audio events, objects, acoustic scenes, signal properties, semantics, and listener emotions. To evaluate Mellow's reasoning ability, we benchmark it on a diverse set of tasks, assessing on both in-distribution and out-of-distribution data, including audio understanding, deductive reasoning, and comparative reasoning. Finally, we conduct extensive ablation studies to explore the impact of projection layer choices, synthetic data generation methods, and language model pretraining on reasoning performance. Our training dataset, findings, and baseline pave the way for developing small ALMs capable of reasoning.

Soham Deshmukh, Shuo Han, Rita Singh et al.

Understanding and explaining differences between audio recordings is crucial for fields like audio forensics, quality assessment, and audio generation. This involves identifying and describing audio events, acoustic scenes, signal characteristics, and their emotional impact on listeners. This paper stands out as the first work to comprehensively study the task of explaining audio differences and then propose benchmark, baselines for the task. First, we present two new datasets for audio difference explanation derived from the AudioCaps and Clotho audio captioning datasets. Using Large Language Models (LLMs), we generate three levels of difference explanations: (1) concise descriptions of audio events and objects, (2) brief sentences about audio events, acoustic scenes, and signal properties, and (3) comprehensive explanations that include semantics and listener emotions. For the baseline, we use prefix tuning where audio embeddings from two audio files are used to prompt a frozen language model. Our empirical analysis and ablation studies reveal that the naive baseline struggles to distinguish perceptually similar sounds and generate detailed tier 3 explanations. To address these limitations, we propose ADIFF, which introduces a cross-projection module, position captioning, and a three-step training process to enhance the model's ability to produce detailed explanations. We evaluate our model using objective metrics and human evaluation and show our model enhancements lead to significant improvements in performance over naive baseline and SoTA Audio-Language Model (ALM) Qwen Audio. Lastly, we conduct multiple ablation studies to study the effects of cross-projection, language model parameters, position captioning, third stage fine-tuning, and present our findings. Our benchmarks, findings, and strong baseline pave the way for nuanced and human-like explanations of audio differences.

Massa Baali, Shuo Han, Syed Abdul Hannan et al.

Speaker recognition systems are often limited to classification tasks and struggle to generate detailed speaker characteristics or provide context-rich descriptions. These models primarily extract embeddings for speaker identification but fail to capture demographic attributes such as dialect, gender, and age in a structured manner. This paper introduces CoLMbo, a Speaker Language Model (SLM) that addresses these limitations by integrating a speaker encoder with prompt-based conditioning. This allows for the creation of detailed captions based on speaker embeddings. CoLMbo utilizes user-defined prompts to adapt dynamically to new speaker characteristics and provides customized descriptions, including regional dialect variations and age-related traits. This innovative approach not only enhances traditional speaker profiling but also excels in zero-shot scenarios across diverse datasets, marking a significant advancement in the field of speaker recognition.

Satvik Dixit, Soham Deshmukh, Bhiksha Raj · cmu

Audio Question Answering (AQA) is a key task for evaluating Audio-Language Models (ALMs), yet assessing open-ended responses remains challenging. Existing metrics used for AQA such as BLEU, METEOR and BERTScore, mostly adapted from NLP and audio captioning, rely on surface similarity and fail to account for question context, reasoning, and partial correctness. To address the gap in literature, we make three contributions in this work. First, we introduce AQEval to enable systematic benchmarking of AQA metrics. It is the first benchmark of its kind, consisting of 10k model responses annotated by multiple humans for their correctness and relevance. Second, we conduct a comprehensive analysis of existing AQA metrics on AQEval, highlighting weak correlation with human judgment, especially for longer answers. Third, we propose a new metric - AURA score, to better evaluate open-ended model responses. On AQEval, AURA achieves state-of-the-art correlation with human ratings, significantly outperforming all baselines. Through this work, we aim to highlight the limitations of current AQA evaluation methods and motivate better metrics. We release both the AQEval benchmark and the AURA metric to support future research in holistic AQA evaluation.

Ruijie Zhou, Soham Deshmukh, Jeremiah Greer et al.

Current research in dialogue systems is focused on conversational assistants working on short conversations in either task-oriented or open domain settings. In this paper, we focus on improving task-based conversational assistants online, primarily those working on document-type conversations (e.g., emails) whose contents may or may not be completely related to the assistant's task. We propose "NARLE" a deep reinforcement learning (RL) framework for improving the natural language understanding (NLU) component of dialogue systems online without the need to collect human labels for customer data. The proposed solution associates user emotion with the assistant's action and uses that to improve NLU models using policy gradients. For two intent classification problems, we empirically show that using reinforcement learning to fine tune the pre-trained supervised learning models improves performance up to 43%. Furthermore, we demonstrate the robustness of the method to partial and noisy implicit feedback.

Soham Deshmukh, Bhiksha Raj, Rita Singh

While multitask and transfer learning has shown to improve the performance of neural networks in limited data settings, they require pretraining of the model on large datasets beforehand. In this paper, we focus on improving the performance of weakly supervised sound event detection in low data and noisy settings simultaneously without requiring any pretraining task. To that extent, we propose a shared encoder architecture with sound event detection as a primary task and an additional secondary decoder for a self-supervised auxiliary task. We empirically evaluate the proposed framework for weakly supervised sound event detection on a remix dataset of the DCASE 2019 task 1 acoustic scene data with DCASE 2018 Task 2 sounds event data under 0, 10 and 20 dB SNR. To ensure we retain the localisation information of multiple sound events, we propose a two-step attention pooling mechanism that provides a time-frequency localisation of multiple audio events in the clip. The proposed framework with two-step attention outperforms existing benchmark models by 22.3%, 12.8%, 5.9% on 0, 10 and 20 dB SNR respectively. We carry out an ablation study to determine the contribution of the auxiliary task and two-step attention pooling to the SED performance improvement.

Soham Deshmukh, Mahmoud Al Ismail, Rita Singh

In the pathogenesis of COVID-19, impairment of respiratory functions is often one of the key symptoms. Studies show that in these cases, voice production is also adversely affected -- vocal fold oscillations are asynchronous, asymmetrical and more restricted during phonation. This paper proposes a method that analyzes the differential dynamics of the glottal flow waveform (GFW) during voice production to identify features in them that are most significant for the detection of COVID-19 from voice. Since it is hard to measure this directly in COVID-19 patients, we infer it from recorded speech signals and compare it to the GFW computed from physical model of phonation. For normal voices, the difference between the two should be minimal, since physical models are constructed to explain phonation under assumptions of normalcy. Greater differences implicate anomalies in the bio-physical factors that contribute to the correctness of the physical model, revealing their significance indirectly. Our proposed method uses a CNN-based 2-step attention model that locates anomalies in time-feature space in the difference of the two GFWs, allowing us to infer their potential as discriminative features for classification. The viability of this method is demonstrated using a clinically curated dataset of COVID-19 positive and negative subjects.

Mahmoud Al Ismail, Soham Deshmukh, Rita Singh

Phonation, or the vibration of the vocal folds, is the primary source of vocalization in the production of voiced sounds by humans. It is a complex bio-mechanical process that is highly sensitive to changes in the speaker's respiratory parameters. Since most symptomatic cases of COVID-19 present with moderate to severe impairment of respiratory functions, we hypothesize that signatures of COVID-19 may be observable by examining the vibrations of the vocal folds. Our goal is to validate this hypothesis, and to quantitatively characterize the changes observed to enable the detection of COVID-19 from voice. For this, we use a dynamical system model for the oscillation of the vocal folds, and solve it using our recently developed ADLES algorithm to yield vocal fold oscillation patterns directly from recorded speech. Experimental results on a clinically curated dataset of COVID-19 positive and negative subjects reveal characteristic patterns of vocal fold oscillations that are correlated with COVID-19. We show that these are prominent and discriminative enough that even simple classifiers such as logistic regression yields high detection accuracies using just the recordings of isolated extended vowels.

Soham Deshmukh, Bhiksha Raj, Rita Singh

Weakly Labelled learning has garnered lot of attention in recent years due to its potential to scale Sound Event Detection (SED) and is formulated as Multiple Instance Learning (MIL) problem. This paper proposes a Multi-Task Learning (MTL) framework for learning from Weakly Labelled Audio data which encompasses the traditional MIL setup. To show the utility of proposed framework, we use the input TimeFrequency representation (T-F) reconstruction as the auxiliary task. We show that the chosen auxiliary task de-noises internal T-F representation and improves SED performance under noisy recordings. Our second contribution is introducing two step Attention Pooling mechanism. By having 2-steps in attention mechanism, the network retains better T-F level information without compromising SED performance. The visualisation of first step and second step attention weights helps in localising the audio-event in T-F domain. For evaluating the proposed framework, we remix the DCASE 2019 task 1 acoustic scene data with DCASE 2018 Task 2 sounds event data under 0, 10 and 20 db SNR resulting in a multi-class Weakly labelled SED problem. The proposed total framework outperforms existing benchmark models over all SNRs, specifically 22.3 %, 12.8 %, 5.9 % improvement over benchmark model on 0, 10 and 20 dB SNR respectively. We carry out ablation study to determine the contribution of each auxiliary task and 2-step Attention Pooling to the SED performance improvement. The code is publicly released

Soham Deshmukh, Rahul Rade, Faruk Kazi

Cyber threat intelligence is one of the emerging areas of focus in information security. Much of the recent work has focused on rule-based methods and detection of network attacks using Intrusion Detection algorithms. In this paper we propose a framework for inspecting and modelling the behavioural aspect of an attacker to obtain better insight predictive power on his future actions. For modelling we propose a novel semi-supervised algorithm called Fusion Hidden Markov Model (FHMM) which is more robust to noise, requires comparatively less training time, and utilizes the benefits of ensemble learning to better model temporal relationships in data. This paper evaluates the performances of FHMM and compares it with both traditional algorithms like Markov Chain, Hidden Markov Model (HMM) and recently developed Deep Recurrent Neural Network (Deep RNN) architectures. We conduct the experiments on dataset consisting of real data attacks on a Cowrie honeypot system. FHMM provides accuracy comparable to deep RNN architectures at significant lower training time. Given these experimental results, we recommend using FHMM for modelling discrete temporal data for significantly faster training and better performance than existing methods.