Timothy R. Anderson

Analysis/synthesis-based microphone array speech enhancer with variable signal distortion

A microphone array speech enhancement algorithm based on analysis/synthesis filtering that allows for variable signal distortion. The algorithm is used to suppress additive noise and interference. The processing structure consists of delaying the received signals so that the desired signal components add coherently, filtering each of the delayed signals through an analysis filter bank, summing the corresponding channel outputs from the sensors, applying a gain function to the channel outputs, and combining the weighted channel outputs using a synthesis filter. The structure uses two different gain functions, both of which are based on cross correlations of the channel signals from the two sensors. The first gain yields the GEQ-I array, which performs best for the case of a desired speech signal corrupted by uncorrelated white background noise. The second gain yields the GEQ-II array, which performs best for the case where there are more signals than microphones. The GEQ-II gain allows for a trade-off on a channel-dependent basis of additional signal degradation in exchange for additional noise and interference suppression.

Supervised and Unsupervised Speaker Adaptation in the NIST 2005 Speaker Recognition Evaluation

Starting in 2004, the annual NIST speaker recognition evaluation (SRE) has added an optional unsupervised speaker adaptation track where test files are processed sequentially and one may update the target model. In this paper, various model adaptation factors are investigated for MAP adaptation using a supervised (ideal) adaptation scheme. Once the best performing model adaptation factor is found, unsupervised adaptation experiments are run using a threshold to determine when to update the target model. Three NIST training conditions, 10sec4w, 1conv4w, and 8conv4w, all with the 1conv4w test condition are used for experiments with the NIST 2005 SRE. MinDCF values for the three training conditions are reduced by 60.9% for 10sec4w, 48.3% for 1conv4w, and 33.3% for 8conv4w using the supervised adaptation compared to the baseline. For the unsupervised adaptation, minDCF values were reduced by 16.7%, 21.6%, and 20.5% for the respective training conditions

Speech processing using an auditory model and neural networks: A preprocessing comparison.

A neural network that employs unsupervised learning was used on the output of a neurophysiologically based model of the auditory periphery [K. L. Payton, J. Acoust. Soc. Am. 83, 145–162 (1988)] to perform phoneme recognition [T. R. Anderson, J. Acoust. Soc. Am. Suppl. 1 87, S107 (1990)]. Continuous speech from ten speakers (three female and seven male) taken from the TIMIT database [W. Fisher, V. Zue, J. Bernstein, and D. Pallet, J. Acoust. Soc. Am. Suppl. 1 81, S92 (1987)] was processed through the auditory model and through an FFT based filter bank for this experiment. Each representation used a window length of 16 ms with an 8‐ms overlap. The average firing rate of 20 channels from the auditory model and the FFT filter‐bank output was used to train two separate Kohonen self‐organizing feature maps. Results indicate that the auditory model performs significantly better than the FFT. However, the two representations make different types of broad class errors. Results in the form of average recognition, s...

The Perception of Synthetic Speech in Noise

Although much information about synthetic speech has been acquired over past decades, we have been unable to find in the literature a systematic examination of the perception of synthetic speech in noise. Simpson [1] has reported that synthetic speech altitude callouts to airline pilots in widebody jet cockpit noise at a S/N of -10 dB for the first time were 99.1% intelligible and that synthetic speech voice warnings to helicopter pilots in simulated helicopter noise at a S/N ratio of -22 dB were 99.2% intelligible [2]. Nusbaum [3] has reported that perceptual confusions for synthetic CV and VC syllables were quite different than confusions observed for natural speech degraded by noise. Pisoni (personal communication) indicates that one of two synthetic speech systems with very high levels of segmental intelligibility in quiet, showed greater decrements in the intelligibility of CV syllables in noise than did the other system. Clark [4] reported little difference in the intelligibility of vowels in noise for synthetic and natural speech, whereas natural CV syllables were clearly superior to synthetic CV syllables under all noise conditions.

The AFRL-MITLL WMT16 News-Translation Task Systems

This paper describes the AFRL-MITLL statistical machine translation systems and the improvements that were developed during the WMT16 evaluation campaign. New techniques applied this year include Neural Machine Translation, a unique selection process for language modelling data, additional out-of-vocabulary transliteration techniques, and morphology generation.

Improved Phrase Translation Modeling Using MAP Adaptation

In this paper, we explore several methods of improving the estimation of translation model probabilities for phrase-based statistical machine translation given in-domain data sparsity. We introduce a hierarchical variant of maximum a posteriori (MAP) adaptation for domain adaptation with an arbitrary number of out-of-domain models. We note that domain adaptation can have a smoothing effect, and we explore the interaction between smoothing and the incorporation of out-of-domain data. We find that the relative contributions of smoothing and interpolation depend on the datasets used. For both the IWSLT 2011 and WMT 2011 English-French datasets, the MAP adaptation method we present improves on a baseline system by 1.5+ BLEU points.

An artificial neural network model of human sound localization

An artificial neural network was trained to identify the location of sound sources using the head‐related transfer functions (HRTFs) of Wightman and Kistler [J. Acoust. Soc. Am. 85, 858–867 (1989)]. The simulated signals were either filtered clicks or pure tones, with speaker placements separated in steps of 15 deg in azimuth of 18 deg in elevation. After the signals were passed through the HRTFs, the inputs to the nets were computed as the difference of left ear and right ear phase spectra or the difference of the power at the output of left and right ear third‐octave or twelfth‐octave filter banks. Back propagation was used to train the nets. Separate nets were trained for each signal type and for each type of input data. Better than 90% correct identification of the source speakers location can be achieved in either the horizontal or median planes. The results for the horizontal plane are compared to the predictions of the duplex theory of sound localization. [Work supported by AFOSR‐91‐0289 and AFOSR‐...

Neural network models of sound localization based on monaural information and based on binaural information

Neural networks were trained with back propagation to localize ‘‘virtual’’ sounds that could originate from any of 24 azimuths (−165° to +180°) and 6 elevations (−36° to +54°). The sounds (clicks) were filtered with head related transfer functions and transformed into 22‐point quarter‐octave spectra. The networks were composed of 22 to 44 input nodes, followed by 50 hidden nodes and 30 output nodes (24 azimuth nodes and 6 elevation nodes). With the ‘‘binaural’’ configuration, the interaural delay and the interaural difference spectrum were provided as inputs to the network. With the ‘‘monaural’’ configuration, separate networks were trained with the left‐ear and the right‐ear spectra; a third, arbitrator, network learned to localize based on the output of these two monaural networks (i.e., the activation levels of their azimuth and elevation nodes). The monaural configuration did not allow for binaural interaction in the traditional sense. Both configurations achieved performance comparable to humans, sug...

Binaural sound localization using neural networks.

Artificial neural networks (ANNs) were used as a tool to localize sound sources from simulated, binaural signals. The sound sources for the experiments were restricted to a circle of radius 9 ft, centered about the head and lying on the horizontal circle. Sound source positions were randomly selected from one of the 360, one‐deg increments on the circle. Classes for the ANNs were created by dividing the circle into equally sized wedges, much like slices of a pie. The number of classes used in the experiments varied from 4 to 36. Two types of sound source signals were considered: tones and Gaussian noise. Three different feature sets were tried. Results will be presented that compare the performance of the three feature sets for each sound source type. The best feature set produced similar results in terms of localization accuracy on tones and Gaussian noise (over 91% for 18 classes). Observations were made of phenomena which also occur in human psychological experiments such as front–back confusions and i...

Speech processing using an auditory model and neural networks

A neural network that employs unsupervised learning was used on the output of a neurophysiologically based model of the auditory periphery [K. L. Payton, J. Acoust. Soc. Am. 83, 145–162 (1988)] to perform phoneme recognition. The auditory model incorporates steps describing the conversion from the acoustic pressure‐wave signal at the eardrum to the time course activity in auditory neurons. The model can process arbitrary time domain waveforms and yields the predicated neural firing rate. Continuous speech from ten speakers (three female and seven male) taken from the TIMIT database [W. Fisher et al., J. Acoust. Soc. Am. Suppl. 1 81, S92 (1987)] was processed through the auditory model for this experiment. The average firing rate of 20 channels from the auditory model was used to train a Kohonen self‐organizing feature map. The resulting context‐independent phoneme recognition performance (30% correct) was comparable to that of the SPHINX System [K. F. Lee and H. W. Hon, IEEE Trans. Acoust. Speech Signal P...