Miles Moody

Application of noise reduction techniques for alaryngeal speech enhancement

A hybrid noise reduction method using both spectral subtraction and root cepstral subtraction procedures is applied to speech produced using an artificial larynx (electro-larynx). The noise reduction suppresses the direct path noise of the electro-larynx device and provides an improvement in mean opinion score.

Spatial bayesian learning algorithms for geographic information retrieval

An increasing amount of freely available Geographic Information System (GIS) data on the Internet has stimulated recent research into Geographic Information Retrieval (GIR). Typically, GIR looks at the problem of retrieving GIS datasets on a theme by theme basis. However in practice, themes are generally not analysed in isolation. More often than not multiple themes are required to create a map for a particular analysis task. To do this using the current GIR techniques, each theme is retrieved one by one using traditional retrieval methods and manually added to the map. To automate map creation the traditional GIR paradigm of matching a query to a single theme type must be extended to include discovering relationships between different theme types.Bayesian Inference networks can and have recently been adapted to provide a theme to theme relevance ranking scheme which can be used to automate map creation [2]. The use of Bayesian inference for GIR relies on a manually created Bayesian network. The Bayesian network contains causal probability relationships between spatial themes. The next step in using Bayesian Inference for GIR is to develop algorithms to automatically create a Bayesian network from historical data. This paper discusses a process to utilize conventional Bayesian learning algorithms in GIR. In addition, it proposes three spatial learning Bayesian network algorithms that incorporate spatial relationships between themes into the learning process. The resulting Bayesian networks were loaded into an inference engine that was used to retrieve all relevant themes given a test set of user queries. The performance of the spatial Bayesian learning algorithms were evaluated and compared to performance of conventional non-spatial Bayesian learning algorithms.This contribution will increase the performance and efficiency of knowledge extraction from GIS by allowing users to focus on interpreting data, instead of focusing on finding which data is relevant to their analysis.

Frequency offset correction for HF radio speech reception

High-frequency single-sideband (HF-SSB) radio communication suffers from highly variable propagation conditions and high levels of interference noise. One common distortion is frequency shifting of the received signal due to differences in carrier frequencies at transmitter and receiver, or to Doppler shifts due to the transmission path. This paper presents a novel and efficient technique to compensate for this frequency offset using only the received speech signal to determine and correct the frequency error. Good improvement in received speech quality is achieved, using a mean opinion score measurement.

Intelligibility of reverberant speech enhanced by inversion of room response

The effect on speech intelligibility of inverting the impulse response of a small, highly reverberant acoustical space is evaluated. Phonetically balanced (PB) speech material was recorded with a single microphone in an empty room of about 40 m/sup 3/, with reverberation time of 3 s. The room response was measured using a linear chirp, and various least-squared error inverse filters were designed. Intelligibility of the reverberated and processed speech was measured by quantitative testing with twelve listeners, and by the speech transmission index (STI). Results show improved intelligibility even using very short portions of the room impulse response, with correct responses increasing by 8% using a 93 ms impulse response. Further improvement was noted with increasing filter length. Intelligibility improvements were consistent with two types of subject response tests, but were not reflected in STI values for short filter length.<<ETX>>

Speech separation by simulating the cocktail party effect with a neural network controlled Wiener filter

A novel speech separation structure which simulates the cocktail party effect using a modified iterative Wiener filter and a multi-layer perceptron neural network is presented. The neural network is used as a speaker recognition system to control the iterative Wiener filter. The neural network is a modified perceptron with a hidden layer using feature data extracted from LPC cepstral analysis. The proposed technique has been successfully used for speech separation when the interference is competing speech or broad band noise.

Post-microphone-array speech enhancement with adaptive filters for forensic application

A multi-microphone speech enhancement system including two stages of processing has been developed. In the first stage, the array is aimed at the desired signal source and the noise signal source(s) respectively by means of the Frost algorithm or DS beamforming. In the second stage, the difference between the SNRs of the array outputs are employed to improve the SNR of the desired signal further by an adaptive filter with MRSS (matrix ratio sub space) algorithm. This algorithm is based on the orthogonal and coincidental subspace eigen analysis. The mathematical derivation and the simulation results that illustrate the advantages of the system are presented.<<ETX>>

Dynamic time-warp compensation for correlation of long sequences

Dynamic time-warping of direct-sequence spread-spectrum causes a deterioration in the process gain when time-linear correlation is applied. Applications with channels that suffer this effect are moving platforms such as to airborne communications, acoustic communications, and tape recorders. Of these, the tape recorder has the most severe time-warp. This paper describes a low powered PN-sequence which is added to the audio, such that the audible quality is not impaired. This is used for authentication of the tape recording. Tape speed errors from wow and flutter cause time based errors and a loss of the detection process gain. A technique for recovering the process gain and detecting the PN-sequence from this channel using a dynamic programming algorithm is presented.<<ETX>>