The Home Office Research Studies are reports on research undertaken by or on behalf of the Home Office. They cover the range of subjects for which the Home Secre t a ry has re s p o n s i b i l i t y. Other publications produced by the Research, Development and Statistics include Findings, Statistical Bulletins and Statistical Papers. RDS is part of the Home Office. RDS staff are embedded within delivery groups working closely with front-line staff. The HO Chief Scientific Advisor, who is also Director of RDS, oversees p rofessional development for RDS teams, quality assurance and strategic R & D issues. The Home Off i c e 's purpose is to build a safe, just and tolerant society in which the rights and responsibilities of individuals, families and communities are properly balanced and the p rotection and security of the public are maintained. RDS includes staff within the Government Statistical Service (GSS). One of the GSS aims is to i n f o rm Parliament and the members of the public about the state of the nation and provide a window on the work and perf o rmance of government, allowing the impact of govern m e n t policies and actions to be assessed. T h e re f o re-R e s e a rch Development and Statistics in the Home Office improves policy making, decision taking and practice in support of the Home Office purpose and aims, to provide the public and Parliament with information necessary for informed debate and to publish information for future use.

There is information in speech sounds about the length of the vocal tract; specifically, as a child grows, the resonators in the vocal tract grow and the formant frequencies of the vowels decrease. It has been hypothesized that the auditory system applies a scale transform to all sounds to segregate size information from resonator shape information, and thereby enhance both size perception and speech recognition [Irino and Patterson, Speech Commun. 36, 181-203 (2002)]. This paper describes size discrimination experiments and vowel recognition experiments designed to provide evidence for an auditory scaling mechanism. Vowels were scaled to represent people with vocal tracts much longer and shorter than normal, and with pitches much higher and lower than normal. The results of the discrimination experiments show that listeners can make fine judgments about the relative size of speakers, and they can do so for vowels scaled well beyond the normal range. Similarly, the recognition experiments show good performance for vowels in the normal range, and for vowels scaled well beyond the normal range of experience. Together, the experiments support the hypothesis that the auditory system automatically normalizes for the size information in communication sounds.

No abstract

Variational methods are a key component of the approximate inference and learning toolbox. These methods fill an important middle ground, retaining distributional information about uncertainty in latent variables, unlike maximum a posteriori methods (MAP), and yet requiring fewer computational resources than Monte Carlo Markov Chain methods. In particular the variational Expectation Maximisation (vEM) and variational Bayes algorithms, both involving variational optimisation of a free energy, are widely used in time-series modelling. Here, we investigate the success of vEM in simple probabilistic time-series models. First we consider the inference step of vEM, and show that a consequence of the wellknown compactness property is a failure to propagate uncertainty in time, thus limiting the usefulness of the retained distributional information. In particular, the uncertainty may appear to be smallest precisely when the approximation is poorest. Second, we consider parameter learning and analytically reveal systematic biases in the parameters found by vEM. Surprisingly, simpler variational approximations (such a mean-field) can lead to less bias than more complicated structured approximations.

Prosodic rhythm in speech [the alternation of "Strong" (S) and "weak" (w) syllables] is cued, among others, by slow rates of amplitude modulation (AM) within the speech envelope. However, it is unclear exactly which envelope modulation rates and statistics are the most important for the rhythm percept. Here, the hypothesis that the phase relationship between "Stress" rate (∼2 Hz) and "Syllable" rate (∼4 Hz) AMs provides a perceptual cue for speech rhythm is tested. In a rhythm judgment task, adult listeners identified AM tone-vocoded nursery rhyme sentences that carried either trochaic (S-w) or iambic patterning (w-S). Manipulation of listeners' rhythm perception was attempted by parametrically phase-shifting the Stress AM and Syllable AM in the vocoder. It was expected that a 1π radian phase-shift (half a cycle) would reverse the perceived rhythm pattern (i.e., trochaic → iambic) whereas a 2π radian shift (full cycle) would retain the perceived rhythm pattern (i.e., trochaic → trochaic). The results confirmed these predictions. Listeners judgments of rhythm systematically followed Stress-Syllable AM phase-shifts, but were unaffected by phase-shifts between the Syllable AM and the Sub-beat AM (∼14 Hz) in a control condition. It is concluded that the Stress-Syllable AM phase relationship is an envelope-based modulation statistic that supports speech rhythm perception.

Model-free deep reinforcement learning (RL) methods have been successful in a wide variety of simulated domains. However, a major obstacle facing deep RL in the real world is their high sample complexity. Batch policy gradient methods offer stable learning, but at the cost of high variance, which often requires large batches. TD-style methods, such as off-policy actor-critic and Q-learning, are more sample-efficient but biased, and often require costly hyperparameter sweeps to stabilize. In this work, we aim to develop methods that combine the stability of policy gradients with the efficiency of off-policy RL. We present Q-Prop, a policy gradient method that uses a Taylor expansion of the off-policy critic as a control variate. Q-Prop is both sample efficient and stable, and effectively combines the benefits of on-policy and off-policy methods. We analyze the connection between Q-Prop and existing model-free algorithms, and use control variate theory to derive two variants of Q-Prop with conservative and aggressive adaptation. We show that conservative Q-Prop provides substantial gains in sample efficiency over trust region policy optimization (TRPO) with generalized advantage estimation (GAE), and improves stability over deep deterministic policy gradient (DDPG), the stateof-the-art on-policy and off-policy methods, on OpenAI Gym's MuJoCo continuous control environments.

SUMMARYA study is undertaken of a particular method of updating numerical prediction models. The method constrains the time development of a subset of the model fields toward a prescribed space-time estimate of those fields, while the other model variables are allowed to evolve in an explicitly unconstrained fashion. It represents an attempt to update the model by a form of dynamical relaxation.An analysis of the method is carried out in the context of the primitive equations linearized about an isothermal basic state of no motion. It is shown that, for a particular form of the scheme and the availability of :he time history of the mass field (or the wind field) on an f-plane, all the time-and space-scale error fields in the initial specification suffer an amplitude reduction of a t least the order on the timescale of one day. On an equatorial p-plane it is shown that the same rate of amplitude reduction can be achieved if an accurate time history is known for both the mass field and the zonal-wind field. Numerical experiments performed with the nonlinear shallow-water equations for a mid-latitude ,&plane geometry support these results and demonstrate that the technique compares favourably with the conventional direct insertion update method. Consideration is also given to the possible effects of errors in the prescribed fields and it is shown that !he relaxation scheme can to some extent be tuned to offset the effect of this particular source of error.This study of the relaxation update scheme, although not comprehensive, is nevertheless sufficient to indicate its potential. However, it is stressed that a trenchant assessment of the scheme's usefulness should be based at least in part upon its performance under more testing and realistic conditions.

No abstract