Value alignment, at the intersection of moral philosophy and AI safety, is dedicated to ensuring that artificially intelligent (AI) systems align with a certain set of values. One challenge facing value alignment researchers is accurately translating these values into a machine readable format. In the case of reinforcement learning (RL), a popular method within value alignment, this requires designing a reward function which accurately defines the value of all state-action pairs. It is common for programmers to hand-set and manually tune these values. In this paper, we examine the challenges of hand-programming values into reward functions for value alignment, and propose mathematical models as an alternative grounding for reward function design in ethical scenarios. Experimental results demonstrate that our modelled-ethics approach offers a more consistent alternative and outperforms our hand-programmed reward functions.

Rainbow Deep Q-Network (DQN) demonstrated combining multiple independent enhancements could significantly boost a reinforcement learning (RL) agent’s performance. In this paper, we present “Beyond The Rainbow” (BTR), a novel algorithm that integrates six improvements from across the RL literature to Rainbow DQN, establishing a new state-of-the-art for RL using a desktop PC, with a human-normalized interquartile mean (IQM) of 7.6 on Atari-60. Beyond Atari, we demonstrate BTR’s capability to handle complex 3D games, successfully training agents to play Super Mario Galaxy, Mario Kart, and Mortal Kombat with minimal algorithmic changes. Designing BTR with computational efficiency in mind, agents can be trained using a high-end desktop PC on 200 million Atari frames within 12 hours. Additionally, we conduct detailed ablation studies of each component, analyzing the performance and impact using numerous measures.

Multi-channel signals captured by spatially separated sensors often contain a high level of data redundancy. A compact signal representation enables more efficient storage and processing, which has been exploited for data compression, noise reduction, and speech and image coding. This paper focuses on the compact representation of speech signals acquired by spherical microphone arrays. A polynomial matrix eigenvalue decomposition (PEVD) can spatially decorrelate signals over a range of time lags and is known to achieve optimum multi-channel data compaction. However, the complexity of PEVD algorithms scales at best cubically with the number of channel signals, e.g., the number of microphones comprised in a spherical array used for processing. In contrast, the spherical harmonic transform (SHT) provides a compact spatial representation of the 3-dimensional sound field measured by spherical microphone arrays, referred to as eigenbeam signals, at a cost that rises only quadratically with the number of microphones. Yet, the SHT's spatially orthogonal basis functions cannot completely decorrelate sound field components over a range of time lags. In this work, we propose to exploit the compact representation offered by the SHT to reduce the number of channels used for subsequent PEVD processing. In the proposed framework for signal representation, we show that the diagonality factor improves by up to 7 dB over the microphone signal representation with a significantly lower computation cost. Moreover, when applying this framework to speech enhancement and source separation, the proposed method improves metrics known as short-time objective intelligibility (STOI) and source-to-distortion ratio (SDR) by up to 0.2 and 20 dB, respectively.

We generalise the problem of reward modelling (RM) for reinforcement learning (RL) to handle non-Markovian rewards. Existing work assumes that human evaluators observe each step in a trajectory independently when providing feedback on agent behaviour. In this work, we remove this assumption, extending RM to capture temporal dependencies in human assessment of trajectories. We show how RM can be approached as a multiple instance learning (MIL) problem, where trajectories are treated as bags with return labels, and steps within the trajectories are instances with unseen reward labels. We go on to develop new MIL models that are able to capture the time dependencies in labelled trajectories. We demonstrate on a range of RL tasks that our novel MIL models can reconstruct reward functions to a high level of accuracy, and can be used to train high-performing agent policies.

Abstract Improving the ability to interact through voice with a robot is still a challenge especially in real environments where multiple speakers coexist. This work has evaluated a proposal based on improving the intelligibility of the voice information that feeds an existing ASR service in the network and in conditions similar to those that could occur in a care centre for the elderly. The results indicate the feasibility and improvement of a proposal based on the use of an embedded microphone array and the use of a simple beamforming and masking technique. The system has been evaluated with 12 people and results obtained for time responsiveness indicate that the system would allow natural interaction with voice. It is shown to be necessary to incorporate a system to properly employ the masking algorithm, through the intelligent and stable estimation of the interfering signals. In addition, this approach allows to fix as sources of interest other speakers not located in the vicinity of the robot.

An essential part of any diarization system is the task of speaker segmentation which is important for many applications including speaker indexing and automatic speech recognition (ASR) in multi-speaker environments. Segmentation of overlapping speech has recently been a key focus of this work. In this paper we explore the use of a new multimodal approach for overlapping speaker segmentation that tracks both the fundamental frequency (F0) of the speaker and the speaker’s direction of arrival (DOA) simultaneously. Our proposed multiple hypothesis tracking system, which simultaneously tracks both features, shows an improvement in segmentation performance when compared to tracking these features separately. An illustrative example of overlapping speech demonstrates the effectiveness of our proposed system. We also undertake a statistical analysis on 12 meetings from the AMI corpus and show an improvement in the HIT rate of 14.1% on average against a commonly used deep learning bidirectional long short term memory network (BLSTM) approach.

Speech enhancement algorithms using polynomial matrix eigenvalue decomposition (PEVD) have been shown to be effective for noisy and reverberant speech. However, these algorithms do not scale well in complexity with the number of channels used in the processing. For a spherical microphone array sampling an order-limited sound field, the spherical harmonics provide a compact representation of the microphone signals in the form of eigenbeams. We propose a PEVD algorithm that uses only the lower dimension eigenbeams for speech enhancement at a significantly lower computation cost. The proposed algorithm is shown to significantly reduce complexity while maintaining full performance. Informal listening examples have also indicated that the processing does not introduce any noticeable artefacts.

This paper demonstrates how the harmonic structure of voiced speech can be exploited to segment multiple overlapping speakers in a speaker diarization task. We explore how a change in the speaker can be inferred from a change in pitch. We show that voiced harmonics can be useful in detecting when more than one speaker is talking, such as during overlapping speaker activity. A novel system is proposed to track multiple harmonics simultaneously, allowing for the determination of onsets and end-points of a speaker's utterance in the presence of an additional active speaker. This system is bench-marked against a segmentation system from the literature that employs a bidirectional long short term memory network (BLSTM) approach and requires training. Experimental results highlight that the proposed approach outperforms the BLSTM baseline approach by 12.9% in terms of HIT rate for speaker segmentation. We also show that the estimated pitch tracks of our system can be used as features to the BLSTM to achieve further improvements of 1.21% in terms of coverage and 2.45% in terms of purity.

Speech enhancement is important for applications such as telecommunications, hearing aids, automatic speech recognition and voice-controlled systems. Enhancement algorithms aim to reduce interfering noise and reverberation while minimizing any speech distortion. In this work for speech enhancement, we propose to use polynomial matrices to model the spatial, spectral and temporal correlations between the speech signals received by a microphone array and polynomial matrix eigenvalue decomposition (PEVD) to decorrelate in space, time and frequency simultaneously. We then propose a blind and unsupervised PEVD-based speech enhancement algorithm. Simulations and informal listening examples involving diverse reverberant and noisy environments have shown that our method can jointly suppress noise and reverberation, thereby achieving speech enhancement without introducing processing artefacts into the enhanced signal.

The enhancement of noisy speech is important for applications involving human-to-human interactions, such as telecommunications and hearing aids, as well as human-to-machine interactions, such as voice-controlled systems and robot audition. In this work, we focus on reverberant environments. It is shown that, by exploiting the lack of correlation between speech and the late reflections, further noise reduction can be achieved. This is verified using simulations involving actual acoustic impulse responses and noise from the ACE corpus. The simulations show that even without using a noise estimator, our proposed method simultaneously achieves noise reduction, and enhancement of speech quality and intelligibility, in reverberant environments over a wide range of SNRs. Furthermore, informal listening examples highlight that our approach does not introduce any significant processing artefacts such as musical noise.

The ability to localize and track acoustic events is a fundamental prerequisite for equipping machines with the ability to be aware of and engage with humans in their surrounding environment. However, in realistic scenarios, audio signals are adversely affected by reverberation, noise, interference, and periods of speech inactivity. In dynamic scenarios, where the sources and microphone platforms may be moving, the signals are additionally affected by variations in the source-sensor geometries. In practice, approaches to sound source localization and tracking are often impeded by missing estimates of active sources, estimation errors, as well as false estimates. The aim of the LOCAlization and TrAcking (LOCATA) Challenge is an open-access framework for the objective evaluation and benchmarking of broad classes of algorithms for sound source localization and tracking. This article provides a review of relevant localization and tracking algorithms and, within the context of the existing literature, a detailed evaluation and dissemination of the LOCATA submissions. The evaluation highlights achievements in the field, open challenges, and identifies potential future directions.

Reverberation is present in our workplaces, our homes, concert halls and theatres. This article investigates how deep learning can use the effect of reverberation on speech to classify a recording in terms of the room in which it was recorded. Existing approaches in the literature rely on domain expertise to manually select acoustic parameters as inputs to classifiers. Estimation of these parameters from reverberant speech is adversely affected by estimation errors, impacting the classification accuracy. In order to overcome the limitations of previously proposed methods, this paper shows how DNNs can perform the classification by operating directly on reverberant speech spectra and a CRNN with an attention-mechanism is proposed for the task. The relationship is investigated between the reverberant speech representations learned by the DNNs and acoustic parameters. For evaluation, AIRs are used from the ACE-challenge dataset that were measured in 7 real rooms. The classification accuracy of the CRNN classifier in the experiments is 78% when using 5 hours of training data and 90% when using 10 hours.

This paper shows that time varying pitch properties can be used advantageously within the segmentation step of a multi-talker diarization system. First a study is conducted to verify that changes in pitch are strong indicators of changes in the speaker. It is then highlighted that an individual's pitch is smoothly varying and, therefore, can be predicted by means of a Kalman filter. Subsequently it is shown that if the pitch is not predictable then this is most likely due to a change in the speaker. Finally, a novel system is proposed that uses this approach of pitch prediction for speaker change detection. This system is then evaluated against a commonly used MFCC segmentation system. The proposed system is shown to increase the speaker change detection rate from 43.3% to 70.5% on meetings in the AMI corpus. Therefore, there are two equally weighted contributions in this paper: 1. We address the question of whether a change in pitch is a reliable estimator of a speaker change in multi-talk meeting audio. 2. We develop a method to extract such speaker changes and test them on a widely available meeting corpus.

In this letter, we address the problem of simultaneously tracking several moving audio sources, namely the problem of estimating source trajectories from a sequence of observed features. We propose to use the von Mises distribution to model audio-source directions of arrival with circular random variables. This leads to a Bayesian filtering formulation, which is intractable because of the combinatorial explosion of associating observed variables with latent variables, over time. We propose a variational approximation of the filtering distribution. We infer a variational expectation-maximization algorithm that is both computationally tractable and time efficient. We propose an audio-source birth method that favors smooth source trajectories and which is used both to initialize the number of active sources and to detect new sources. We perform experiments with the recently released LOCATA dataset comprising two moving sources and a moving microphone array mounted onto a robot.

Distributed acoustic tracking estimates the trajectories of source positions using an acoustic sensor network. As it is often difficult to estimate the source-sensor range from individual nodes, the source positions have to be inferred from the direction-of-arrival (DoA) estimates. Due to reverberation and noise, the sound field becomes increasingly diffuse with increasing source-sensor distance, leading to a decreased Direction of Arrival (DoA)-estimation accuracy. To distinguish between accurate and uncertain DoA estimates, this letter proposes to incorporate the coherent-to-diffuse ratio as a measure of DoA reliability for single-source tracking. It is shown that the source positions, therefore, can be probabilistically triangulated by exploiting the spatial diversity of all nodes.

An algorithm is presented that enables devices equipped with microphones, such as robots, to move within their environment in order to explore, adapt to, and interact with sound sources of interest. Acoustic scene mapping creates a three-dimensional (3D) representation of the positional information of sound sources across time and space. In practice, positional source information is only provided by Direction-of-Arrival (DoA) estimates of the source directions; the source-sensor range is typically difficult to obtain. DoA estimates are also adversely affected by reverberation, noise, and interference, leading to errors in source location estimation and consequent false DoA estimates. Moreover, many acoustic sources, such as human talkers, are not continuously active, such that periods of inactivity lead to missing DoA estimates. Withal, the DoA estimates are specified relative to the observer's sensor location and orientation. Accurate positional information about the observer therefore is crucial. This paper proposes Acoustic Simultaneous Localization and Mapping (aSLAM), which uses acoustic signals to simultaneously map the 3D positions of multiple sound sources while passively localizing the observer within the scene map. The performance of aSLAM is analyzed and evaluated using a series of realistic simulations. Results are presented to show the impact of the observer motion and sound source localization accuracy.

Intuitive spoken dialogues are a prerequisite for human-robot interaction. In many practical situations, robots must be able to identify and focus on sources of interest in the presence of interfering speakers. Techniques such as spatial filtering and blind source separation are therefore often used, but rely on accurate knowledge of the source location. In practice, sound emitted in enclosed environments is subject to reverberation and noise. Hence, sound source localization must be robust to both diffuse noise due to late reverberation, as well as spurious detections due to early reflections. For improved robustness against reverberation, this paper proposes a novel approach for sound source tracking that constructively exploits the spatial diversity of a microphone array installed in a moving robot. In previous work, we developed speaker localization approaches using expectation-maximization (EM) approaches and using Bayesian approaches. In this paper we propose to combine the EM and Bayesian approach in one framework for improved robustness against reverberation and noise.

Several speech processing and audio data-mining applications rely on a description of the acoustic environment as a feature vector for classification. The discriminative properties of the feature domain play a crucial role in the effectiveness of these methods. In this work, we consider three environment identification tasks and the task of acoustic model selection for speech recognition. A set of acoustic parameters and Machine Learning algorithms for feature selection are used and an analysis is performed on the resulting feature domains for each task. In our experiments, a classification accuracy of 100% is achieved for the majority of tasks and the Word Error Rate is reduced by 20.73 percentage points for Automatic Speech Recognition when using the resulting domains. Experimental results indicate a significant dissimilarity in the parameter choices for the composition of the domains, which highlights the importance of the feature selection process for individual applications.

Direction of arrival (DOA) estimation is a fundamental problem in acoustic signal processing. It is used in a diverse range of applications, including spatial filtering, speech dereverberation, source separation and diarization. Intensity vector-based DOA estimation is attractive, especially for spherical sensor arrays, because it is computationally efficient. Two such methods are presented that operate on a spherical harmonic decomposition of a sound field observed using a spherical microphone array. The first uses pseudointensity vectors (PIVs) and works well in acoustic environments where only one sound source is active at any time. The second uses subspace pseudointensity vectors (SSPIVs) and is targeted at environments where multiple simultaneous soures and significant levels of reverberation make the problem more challenging. Analytical models are used to quantify the effects of an interfering source, diffuse noise, and sensor noise on PIVs and SSPIVs. The accuracy of DOA estimation using PIVs and SSPIVs is compared against the state of the art in simulations including realistic reverberation and noise for single and multiple, stationary and moving sources. Finally, robust performance of the proposed methods is demonstrated by using speech recordings in a real acoustic environment.

Acoustic scene mapping creates a representation of positions of audio sources such as talkers within the surrounding environment of a microphone array. By allowing the array to move, the acoustic scene can be explored in order to improve the map. Furthermore, the spatial diversity of the kinematic array allows for estimation of the source-sensor distance in scenarios where source directions of arrival are measured. As sound source localization is performed relative to the array position, mapping of acoustic sources requires knowledge of the absolute position of the microphone array in the room. If the array is moving, its absolute position is unknown in practice. Hence, Simultaneous Localization and Mapping (SLAM) is required in order to localize the microphone array position and map the surrounding sound sources. In realistic environments, microphone arrays receive a convolutive mixture of direct-path speech signals, noise and reflections due to reverberation. A key challenge of Acoustic SLAM (a-SLAM) is robustness against reverberant clutter measurements and missing source detections. This paper proposes a novel bearing-only a-SLAM approach using a Single-Cluster Probability Hypothesis Density filter. Results demonstrate convergence to accurate estimates of the array trajectory and source positions.

Reverberation and noise cause significant deterioration of audio quality and intelligibility to signals recorded in acoustic environments. Noise is usually modeled as a common signal observed in the room and independent of room acoustics. However, this simplistic model cannot necessarily capture the effects of separate noise sources at different locations in the room. This paper proposes a noise model that considers distinct noise sources whose individual acoustic impulse responses are separated into source-sensor specific and common acoustical resonances. Further to noise, the signal is distorted by reverberation. Using parametric models of the system, recursive expressions of the filtering distribution can be derived. Based on these results, a sequential Monte Carlo approach for online dereverberation and enhancement is proposed. Simulation results for speech are presented to verify the effectiveness of the model and method.

Reverberation and noise cause significant deterioration of audio quality and intelligibility to signals recorded in acoustic environments. Bayesian dereverberation infers knowledge about the system by exploiting the statistical properties of speech and the acoustic channel. In Bayesian frameworks, the signal can be processed either sequentially using online methods or in a batch using offline methods. This paper compares the two approaches for blind speech dereverberation by means of a previously proposed batch approach and a novel sequential approach. Results show that while both methods have different advantages, online processing leads to a more flexible solution.

In this paper, a semi-analytical approach for the performance analysis of the random time slot (TS) hopping (RTSH) algorithm applied to code division multiple access - time division duplex(CDMA-TDD) systems will be given. TDD systems are subject to two independent interference scenarios, giving rise to interference diversity, which is exploited by the RTSH algorithm. Depending on the actual slot assignment, the system either experiences same-entity interference (MS (mobile station)-to-MS and BS (base station)-to-BS interference) or other- entity interference (MS-to-BS and BS-to-MS interference). The RTSH algorithm results in a random switching between these two scenarios, each of which will result in a different level of interference. Thereby, constant severe interference is avoided. It has been shown that the RTSH algorithm results in lowest interference for channel asymmetries in favor of the downlink (DL).