For most of the blind separation methods of convolutive mixtures, the parameters of unmixing filter are derived in frequency domain. This leads to a seldom mentioned but important problem that generally the independence assumption between source signals collapses in frequency domain because of the inadequate samples. There exists correlation at each frequency bin. Sometimes it is too high to be neglected and consequently degrades the performance of all the BSS methods in various degrees. In this paper, we propose a recursive algorithm for lowering the unfavorable effect from the correlation, and combine it with the TDD-based blind separation method proposed by S. Ikeda and N. Murata. The bin mixtures are separated into the components of the sources as practical instead of the independent bins as achieved by the conventional method. The signal-to-noise ratio is greatly increased at certain bins, which results in a much better separation.
Five kinds of loudness functions were examined to determine a suitable function for drawing equal-loudness-level contours based on experimental data. First, equal-loudness relations were measured between 125-Hz and 1-kHz pure tones from 70 phons down to 5 phons and thresholds of hearing were measured at both frequencies. Then, five model equations expressing the equal-loudness relation were derived from the five loudness functions, and these equations were fitted to the data. The results showed that three of them could well explain the equal-loudness relation down to 5 phons. Among the three functions, the loudness function proposed by Zwislocki and Hellman [J. Acoust. Soc. Am., 32, 924 (1960)] and Lochner and Burger [J. Acoust. Soc. Am., 33, 1705–1707 (1961)] was regarded as the most appropriate function for drawing equal-loudness-level contours, because the number of parameters of the function was fewer than that in the other two and the threshold of hearing could be used for the fitting as a datum of the equal-loudness relation, resulting in a more stable estimation of the parameters of the model equation.
This paper proposes a new method, using neural networks, of adapting phone HMMs to noisy speech. The neural networks are designed to map clean speech HMMs to noise-adapted HMMs, using noise HMMs and signal-to-noise ratios (SNRs) as inputs. The neural network is trained by minimizing the mean square error between the output HMMs and the target noise-adapted HMMs. In an evaluation, the proposed method was used to recognize noisy broadcast-news speech in speaker-dependent and speaker-independent modes. The trained networks were found to be effective in recognizing new speakers under new noise and various SNR conditions.
The high frequency components of an auditory stimulus are considered to be the primary contribution to median plane localization. However, a number of studies have demonstrated that the low frequency components of a stimulus are also important in median plane localization. Asano et al. concluded that important cues for front-back discrimination seem to exist in the frequency range below 2 kHz. In the present paper, localization tests were performed in order to examine the contribution of low frequency components to median plane localization. In these tests, the higher (above 4,800 Hz) and lower (below 4,800 Hz) frequency components, respectively, of a wide-band white noise were simultaneously presented from different directions so that the individual components provided different directional information. The results of these tests reveal that: (1) when the source is a wide-band signal, the higher frequency components are dominant in median plane localization, whereas the lower frequency components do not contribute significantly to the localization, and (2) important cues for front-back discrimination do not exist in the low frequency range.
In many measurement situations of wind turbine noise immission, the effects of background noise and reflected sound waves from building facades on the measurements are a major problem in addition to the wind-induced noise on the microphone, resulting in a low signal-to-noise ratio of measurements. The use of a vertical board mounted with a microphone was suggested in the document of IEA recommended practices in case of such a problem. Acoustic tests of the board had been carried out on two types of ground to investigate whether performances of the board outdoors are as expected. Results of the tests revealed that one-third octave band measurements were affected by the diffraction of sound waves incident on the board and agreed well with the previous results for a lower to mid frequency range, while the measurements were found to have irregularities as large as ±2 dB for higher frequencies. The A-weighted measurements were found to be within +6 dB ±0.5 dB compared to that of the free-standing microphone. The shielding effect of the board turned out to depend both on the angle of incidence to the board and microphone position and was found to be as large as 10 dBA.