Spyros Matsoukas

Amazon AGI, Aaron Langford, Aayush Shah et al. · amazon-science

We present Amazon Nova, a new generation of state-of-the-art foundation models that deliver frontier intelligence and industry-leading price performance. Amazon Nova Pro is a highly-capable multimodal model with the best combination of accuracy, speed, and cost for a wide range of tasks. Amazon Nova Lite is a low-cost multimodal model that is lightning fast for processing images, video, documents and text. Amazon Nova Micro is a text-only model that delivers our lowest-latency responses at very low cost. Amazon Nova Canvas is an image generation model that creates professional grade images with rich customization controls. Amazon Nova Reel is a video generation model offering high-quality outputs, customization, and motion control. Our models were built responsibly and with a commitment to customer trust, security, and reliability. We report benchmarking results for core capabilities, agentic performance, long context, functional adaptation, runtime performance, and human evaluation.

Abishek Komma, Nagesh Panyam Chandrasekarasastry, Timothy Leffel et al. · amazon-science

Measurement of interaction quality is a critical task for the improvement of spoken dialog systems. Existing approaches to dialog quality estimation either focus on evaluating the quality of individual turns, or collect dialog-level quality measurements from end users immediately following an interaction. In contrast to these approaches, we introduce a new dialog-level annotation workflow called Dialog Quality Annotation (DQA). DQA expert annotators evaluate the quality of dialogs as a whole, and also label dialogs for attributes such as goal completion and user sentiment. In this contribution, we show that: (i) while dialog quality cannot be completely decomposed into dialog-level attributes, there is a strong relationship between some objective dialog attributes and judgments of dialog quality; (ii) for the task of dialog-level quality estimation, a supervised model trained on dialog-level annotations outperforms methods based purely on aggregating turn-level features; and (iii) the proposed evaluation model shows better domain generalization ability compared to the baselines. On the basis of these results, we argue that having high-quality human-annotated data is an important component of evaluating interaction quality for large industrial-scale voice assistant platforms.

Rahil Parikh, Harshavardhan Sundar, Ming Sun et al. · amazon-science

Acoustic events are sounds with well-defined spectro-temporal characteristics which can be associated with the physical objects generating them. Acoustic scenes are collections of such acoustic events in no specific temporal order. Given this natural linkage between events and scenes, a common belief is that the ability to classify events must help in the classification of scenes. This has led to several efforts attempting to do well on Acoustic Event Tagging (AET) and Acoustic Scene Classification (ASC) using a multi-task network. However, in these efforts, improvement in one task does not guarantee an improvement in the other, suggesting a tension between ASC and AET. It is unclear if improvements in AET translates to improvements in ASC. We explore this conundrum through an extensive empirical study and show that under certain conditions, using AET as an auxiliary task in the multi-task network consistently improves ASC performance. Additionally, ASC performance further improves with the AET data-set size and is not sensitive to the choice of events or the number of events in the AET data-set. We conclude that this improvement in ASC performance comes from the regularization effect of using AET and not from the network's improved ability to discern between acoustic events.

Meng Feng, Chieh-Chi Kao, Qingming Tang et al.

Standard acoustic event classification (AEC) solutions require large-scale collection of data from client devices for model optimization. Federated learning (FL) is a compelling framework that decouples data collection and model training to enhance customer privacy. In this work, we investigate the feasibility of applying FL to improve AEC performance while no customer data can be directly uploaded to the server. We assume no pseudo labels can be inferred from on-device user inputs, aligning with the typical use cases of AEC. We adapt self-supervised learning to the FL framework for on-device continual learning of representations, and it results in improved performance of the downstream AEC classifiers without labeled/pseudo-labeled data available. Compared to the baseline w/o FL, the proposed method improves precision up to 20.3\% relatively while maintaining the recall. Our work differs from prior work in FL in that our approach does not require user-generated learning targets, and the data we use is collected from our Beta program and is de-identified, to maximally simulate the production settings.

Satyapriya Krishna, Andy Zou, Rahul Gupta et al.

The safety and alignment of Large Language Models (LLMs) are critical for their responsible deployment. Current evaluation methods predominantly focus on identifying and preventing overtly harmful outputs. However, they often fail to address a more insidious failure mode: models that produce benign-appearing outputs while operating on malicious or deceptive internal reasoning. This vulnerability, often triggered by sophisticated system prompt injections, allows models to bypass conventional safety filters, posing a significant, underexplored risk. To address this gap, we introduce the Deceptive Reasoning Exposure Suite (D-REX), a novel dataset designed to evaluate the discrepancy between a model's internal reasoning process and its final output. D-REX was constructed through a competitive red-teaming exercise where participants crafted adversarial system prompts to induce such deceptive behaviors. Each sample in D-REX contains the adversarial system prompt, an end-user's test query, the model's seemingly innocuous response, and, crucially, the model's internal chain-of-thought, which reveals the underlying malicious intent. Our benchmark facilitates a new, essential evaluation task: the detection of deceptive alignment. We demonstrate that D-REX presents a significant challenge for existing models and safety mechanisms, highlighting the urgent need for new techniques that scrutinize the internal processes of LLMs, not just their final outputs.

Han Li, Sunghyun Park, Aswarth Dara et al.

Current state-of-the-art large-scale conversational AI or intelligent digital assistant systems in industry comprises a set of components such as Automatic Speech Recognition (ASR) and Natural Language Understanding (NLU). For some of these systems that leverage a shared NLU ontology (e.g., a centralized intent/slot schema), there exists a separate skill routing component to correctly route a request to an appropriate skill, which is either a first-party or third-party application that actually executes on a user request. The skill routing component is needed as there are thousands of skills that can either subscribe to the same intent and/or subscribe to an intent under specific contextual conditions (e.g., device has a screen). Ensuring model robustness or resilience in the skill routing component is an important problem since skills may dynamically change their subscription in the ontology after the skill routing model has been deployed to production. We show how different modeling design choices impact the model robustness in the context of skill routing on a state-of-the-art commercial conversational AI system, specifically on the choices around data augmentation, model architecture, and optimization method. We show that applying data augmentation can be a very effective and practical way to drastically improve model robustness.

Mao Li, Bo Yang, Joshua Levy et al.

Speech emotion recognition (SER) is a key technology to enable more natural human-machine communication. However, SER has long suffered from a lack of public large-scale labeled datasets. To circumvent this problem, we investigate how unsupervised representation learning on unlabeled datasets can benefit SER. We show that the contrastive predictive coding (CPC) method can learn salient representations from unlabeled datasets, which improves emotion recognition performance. In our experiments, this method achieved state-of-the-art concordance correlation coefficient (CCC) performance for all emotion primitives (activation, valence, and dominance) on IEMOCAP. Additionally, on the MSP- Podcast dataset, our method obtained considerable performance improvements compared to baselines.

Sunghyun Park, Han Li, Ameen Patel et al.

Natural Language Understanding (NLU) is an established component within a conversational AI or digital assistant system, and it is responsible for producing semantic understanding of a user request. We propose a scalable and automatic approach for improving NLU in a large-scale conversational AI system by leveraging implicit user feedback, with an insight that user interaction data and dialog context have rich information embedded from which user satisfaction and intention can be inferred. In particular, we propose a general domain-agnostic framework for curating new supervision data for improving NLU from live production traffic. With an extensive set of experiments, we show the results of applying the framework and improving NLU for a large-scale production system and show its impact across 10 domains.

Yixin Gao, Yuriy Mishchenko, Anish Shah et al.

Wake word (WW) spotting is challenging in far-field not only because of the interference in signal transmission but also the complexity in acoustic environments. Traditional WW model training requires large amount of in-domain WW-specific data with substantial human annotations therefore it is hard to build WW models without such data. In this paper we present data-efficient solutions to address the challenges in WW modeling, such as domain-mismatch, noisy conditions, limited annotation, etc. Our proposed system is composed of a multi-condition training pipeline with a stratified data augmentation, which improves the model robustness to a variety of predefined acoustic conditions, together with a semi-supervised learning pipeline to accurately extract the WW and confusable examples from untranscribed speech corpus. Starting from only 10 hours of domain-mismatched WW audio, we are able to enlarge and enrich the training dataset by 20-100 times to capture the acoustic complexity. Our experiments on real user data show that the proposed solutions can achieve comparable performance of a production-grade model by saving 97\% of the amount of WW-specific data collection and 86\% of the bandwidth for annotation.

Praveen Kumar Bodigutla, Aditya Tiwari, Josep Valls Vargas et al.

Dialogue level quality estimation is vital for optimizing data driven dialogue management. Current automated methods to estimate turn and dialogue level user satisfaction employ hand-crafted features and rely on complex annotation schemes, which reduce the generalizability of the trained models. We propose a novel user satisfaction estimation approach which minimizes an adaptive multi-task loss function in order to jointly predict turn-level Response Quality labels provided by experts and explicit dialogue-level ratings provided by end users. The proposed BiLSTM based deep neural net model automatically weighs each turn's contribution towards the estimated dialogue-level rating, implicitly encodes temporal dependencies, and removes the need to hand-craft features. On dialogues sampled from 28 Alexa domains, two dialogue systems and three user groups, the joint dialogue-level satisfaction estimation model achieved up to an absolute 27% (0.43->0.70) and 7% (0.63->0.70) improvement in linear correlation performance over baseline deep neural net and benchmark Gradient boosting regression models, respectively.

Longshaokan Wang, Maryam Fazel-Zarandi, Aditya Tiwari et al.

Speech-based virtual assistants, such as Amazon Alexa, Google assistant, and Apple Siri, typically convert users' audio signals to text data through automatic speech recognition (ASR) and feed the text to downstream dialog models for natural language understanding and response generation. The ASR output is error-prone; however, the downstream dialog models are often trained on error-free text data, making them sensitive to ASR errors during inference time. To bridge the gap and make dialog models more robust to ASR errors, we leverage an ASR error simulator to inject noise into the error-free text data, and subsequently train the dialog models with the augmented data. Compared to other approaches for handling ASR errors, such as using ASR lattice or end-to-end methods, our data augmentation approach does not require any modification to the ASR or downstream dialog models; our approach also does not introduce any additional latency during inference time. We perform extensive experiments on benchmark data and show that our approach improves the performance of downstream dialog models in the presence of ASR errors, and it is particularly effective in the low-resource situations where there are constraints on model size or the training data is scarce.

Bowen Shi, Ming Sun, Krishna C. Puvvada et al.

We study few-shot acoustic event detection (AED) in this paper. Few-shot learning enables detection of new events with very limited labeled data. Compared to other research areas like computer vision, few-shot learning for audio recognition has been under-studied. We formulate few-shot AED problem and explore different ways of utilizing traditional supervised methods for this setting as well as a variety of meta-learning approaches, which are conventionally used to solve few-shot classification problem. Compared to supervised baselines, meta-learning models achieve superior performance, thus showing its effectiveness on generalization to new audio events. Our analysis including impact of initialization and domain discrepancy further validate the advantage of meta-learning approaches in few-shot AED.

Praveen Kumar Bodigutla, Lazaros Polymenakos, Spyros Matsoukas

An automated metric to evaluate dialogue quality is vital for optimizing data driven dialogue management. The common approach of relying on explicit user feedback during a conversation is intrusive and sparse. Current models to estimate user satisfaction use limited feature sets and employ annotation schemes with limited generalizability to conversations spanning multiple domains. To address these gaps, we created a new Response Quality annotation scheme, introduced five new domain-independent feature sets and experimented with six machine learning models to estimate User Satisfaction at both turn and dialogue level. Response Quality ratings achieved significantly high correlation (0.76) with explicit turn-level user ratings. Using the new feature sets we introduced, Gradient Boosting Regression model achieved best (rating [1-5]) prediction performance on 26 seen (linear correlation ~0.79) and one new multi-turn domain (linear correlation 0.67). We observed a 16% relative improvement (68% -> 79%) in binary ("satisfactory/dissatisfactory") class prediction accuracy of a domain-independent dialogue-level satisfaction estimation model after including predicted turn-level satisfaction ratings as features.

Maryam Fazel-Zarandi, Longshaokan Wang, Aditya Tiwari et al.

Training dialog policies for speech-based virtual assistants requires a plethora of conversational data. The data collection phase is often expensive and time consuming due to human involvement. To address this issue, a common solution is to build user simulators for data generation. For the successful deployment of the trained policies into real world domains, it is vital that the user simulator mimics realistic conditions. In particular, speech-based assistants are heavily affected by automatic speech recognition and language understanding errors, hence the user simulator should be able to simulate similar errors. In this paper, we review the existing error simulation methods that induce errors at audio, phoneme, text, or semantic level; and conduct detailed comparisons between the audio-level and text-level methods. In the process, we improve the existing text-level method by introducing confidence score prediction and out-of-vocabulary word mapping. We also explore the impact of audio-level and text-level methods on learning a simple clarification dialog policy to recover from errors to provide insight on future improvement for both approaches.

Praveen Kumar Bodigutla, Longshaokan Wang, Kate Ridgeway et al.

An automated metric to evaluate dialogue quality is vital for optimizing data driven dialogue management. The common approach of relying on explicit user feedback during a conversation is intrusive and sparse. Current models to estimate user satisfaction use limited feature sets and rely on annotation schemes with low inter-rater reliability, limiting generalizability to conversations spanning multiple domains. To address these gaps, we created a new Response Quality annotation scheme, based on which we developed turn-level User Satisfaction metric. We introduced five new domain-independent feature sets and experimented with six machine learning models to estimate the new satisfaction metric. Using Response Quality annotation scheme, across randomly sampled single and multi-turn conversations from 26 domains, we achieved high inter-annotator agreement (Spearman's rho 0.94). The Response Quality labels were highly correlated (0.76) with explicit turn-level user ratings. Gradient boosting regression achieved best correlation of ~0.79 between predicted and annotated user satisfaction labels. Multi Layer Perceptron and Gradient Boosting regression models generalized to an unseen domain better (linear correlation 0.67) than other models. Finally, our ablation study verified that our novel features significantly improved model performance.

Bowen Shi, Ming Sun, Chieh-Chi Kao et al.

Acoustic Event Detection (AED), aiming at detecting categories of events based on audio signals, has found application in many intelligent systems. Recently deep neural network significantly advances this field and reduces detection errors to a large scale. However how to efficiently execute deep models in AED has received much less attention. Meanwhile state-of-the-art AED models are based on large deep models, which are computational demanding and challenging to deploy on devices with constrained computational resources. In this paper, we present a simple yet effective compression approach which jointly leverages knowledge distillation and quantization to compress larger network (teacher model) into compact network (student model). Experimental results show proposed technique not only lowers error rate of original compact network by 15% through distillation but also further reduces its model size to a large extent (2% of teacher, 12% of full-precision student) through quantization.

Bowen Shi, Ming Sun, Chieh-Chi Kao et al.

In this paper, we present a compression approach based on the combination of low-rank matrix factorization and quantization training, to reduce complexity for neural network based acoustic event detection (AED) models. Our experimental results show this combined compression approach is very effective. For a three-layer long short-term memory (LSTM) based AED model, the original model size can be reduced to 1% with negligible loss of accuracy. Our approach enables the feasibility of deploying AED for resource-constraint applications.

Bowen Shi, Ming Sun, Chieh-Chi Kao et al.

This paper presents our work of training acoustic event detection (AED) models using unlabeled dataset. Recent acoustic event detectors are based on large-scale neural networks, which are typically trained with huge amounts of labeled data. Labels for acoustic events are expensive to obtain, and relevant acoustic event audios can be limited, especially for rare events. In this paper we leverage an Internet-scale unlabeled dataset with potential domain shift to improve the detection of acoustic events. Based on the classic tri-training approach, our proposed method shows accuracy improvement over both the supervised training baseline, and semisupervised self-training set-up, in all pre-defined acoustic event detection tasks. As our approach relies on ensemble models, we further show the improvements can be distilled to a single model via knowledge distillation, with the resulting single student model maintaining high accuracy of teacher ensemble models.

Sanchit Agarwal, Rahul Goel, Tagyoung Chung et al.

Typical spoken language understanding systems provide narrow semantic parses using a domain-specific ontology. The parses contain intents and slots that are directly consumed by downstream domain applications. In this work we discuss expanding such systems to handle compound entities and intents by introducing a domain-agnostic shallow parser that handles linguistic coordination. We show that our model for parsing coordination learns domain-independent and slot-independent features and is able to segment conjunct boundaries of many different phrasal categories. We also show that using adversarial training can be effective for improving generalization across different slot types for coordination parsing.

Stanislav Peshterliev, John Kearney, Abhyuday Jagannatha et al.

We explore active learning (AL) for improving the accuracy of new domains in a natural language understanding (NLU) system. We propose an algorithm called Majority-CRF that uses an ensemble of classification models to guide the selection of relevant utterances, as well as a sequence labeling model to help prioritize informative examples. Experiments with three domains show that Majority-CRF achieves 6.6%-9% relative error rate reduction compared to random sampling with the same annotation budget, and statistically significant improvements compared to other AL approaches. Additionally, case studies with human-in-the-loop AL on six new domains show 4.6%-9% improvement on an existing NLU system.

Chengwei Su, Rahul Gupta, Shankar Ananthakrishnan et al.

Large scale Natural Language Understanding (NLU) systems are typically trained on large quantities of data, requiring a fast and scalable training strategy. A typical design for NLU systems consists of domain-level NLU modules (domain classifier, intent classifier and named entity recognizer). Hypotheses (NLU interpretations consisting of various intent+slot combinations) from these domain specific modules are typically aggregated with another downstream component. The re-ranker integrates outputs from domain-level recognizers, returning a scored list of cross domain hypotheses. An ideal re-ranker will exhibit the following two properties: (a) it should prefer the most relevant hypothesis for the given input as the top hypothesis and, (b) the interpretation scores corresponding to each hypothesis produced by the re-ranker should be calibrated. Calibration allows the final NLU interpretation score to be comparable across domains. We propose a novel re-ranker strategy that addresses these aspects, while also maintaining domain specific modularity. We design optimization loss functions for such a modularized re-ranker and present results on decreasing the top hypothesis error rate as well as maintaining the model calibration. We also experiment with an extension involving training the domain specific re-rankers on datasets curated independently by each domain to allow further asynchronization. %The proposed re-ranker design showcases the following: (i) improved NLU performance over an unweighted aggregation strategy, (ii) cross-domain calibrated performance and, (iii) support for use cases involving training each re-ranker on datasets curated by each domain independently.

Sri Harish Mallidi, Roland Maas, Kyle Goehner et al.

In this work, we propose a classifier for distinguishing device-directed queries from background speech in the context of interactions with voice assistants. Applications include rejection of false wake-ups or unintended interactions as well as enabling wake-word free follow-up queries. Consider the example interaction: $"Computer,~play~music", "Computer,~reduce~the~volume"$. In this interaction, the user needs to repeat the wake-word ($Computer$) for the second query. To allow for more natural interactions, the device could immediately re-enter listening state after the first query (without wake-word repetition) and accept or reject a potential follow-up as device-directed or background speech. The proposed model consists of two long short-term memory (LSTM) neural networks trained on acoustic features and automatic speech recognition (ASR) 1-best hypotheses, respectively. A feed-forward deep neural network (DNN) is then trained to combine the acoustic and 1-best embeddings, derived from the LSTMs, with features from the ASR decoder. Experimental results show that ASR decoder, acoustic embeddings, and 1-best embeddings yield an equal-error-rate (EER) of $9.3~\%$, $10.9~\%$ and $20.1~\%$, respectively. Combination of the features resulted in a $44~\%$ relative improvement and a final EER of $5.2~\%$.

Anuj Goyal, Angeliki Metallinou, Spyros Matsoukas

Fast expansion of natural language functionality of intelligent virtual agents is critical for achieving engaging and informative interactions. However, developing accurate models for new natural language domains is a time and data intensive process. We propose efficient deep neural network architectures that maximally re-use available resources through transfer learning. Our methods are applied for expanding the understanding capabilities of a popular commercial agent and are evaluated on hundreds of new domains, designed by internal or external developers. We demonstrate that our proposed methods significantly increase accuracy in low resource settings and enable rapid development of accurate models with less data.

Ming Sun, Anirudh Raju, George Tucker et al.

We propose a max-pooling based loss function for training Long Short-Term Memory (LSTM) networks for small-footprint keyword spotting (KWS), with low CPU, memory, and latency requirements. The max-pooling loss training can be further guided by initializing with a cross-entropy loss trained network. A posterior smoothing based evaluation approach is employed to measure keyword spotting performance. Our experimental results show that LSTM models trained using cross-entropy loss or max-pooling loss outperform a cross-entropy loss trained baseline feed-forward Deep Neural Network (DNN). In addition, max-pooling loss trained LSTM with randomly initialized network performs better compared to cross-entropy loss trained LSTM. Finally, the max-pooling loss trained LSTM initialized with a cross-entropy pre-trained network shows the best performance, which yields $67.6\%$ relative reduction compared to baseline feed-forward DNN in Area Under the Curve (AUC) measure.