This paper reports our recent exploration of the layer-by-layer learning strategy for training a multi-layer generative model of patches of speech spectrograms. The top layer of the generative model learns binary codes that can be used for efficient compression of speech and could also be used for scalable speech recognition or rapid speech content retrieval. Each layer of the generative model ...

We present two missing-feature based algorithms that recover noise-corrupted regions of spectrographic representations of speech for noise-robust speech recognition. These algorithms modify the incoming feature vector without any changes to the speech recognition system, in contrast to previously-described approaches. The first approach clusters the feature vectors representing clean speech. Mi...

Most speech recognition systems still use Mel Frequency Cepstral Coefficients (MFCC’s) or Perceptual Linear Prediction Coefficients because these preserve a lot of the information required for recognition while being much more compact than a high-resolution spectrogram. As computers get faster and methods of modeling high-dimensional data improve, however, high-resolution spectrograms or other ...

High-resolution time-frequency (TF) images of multicomponent signals are of great interest for visualization, feature extraction and estimation. The matched Gaussian multitaper spectrogram has been proposed to optimally resolve multicomponent transient functions of Gaussian shape. Hermite functions are used as multitapers and the weights of the different spectrogram functions are optimized. For...

3 Signal Processing 4 3.1 Theoretical Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 3.2 Spectrogram Computation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 3.3 Multitaper Signal Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 3.4 Visualizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ...

While generative adversarial networks (GANs) based neural text-to-speech (TTS) systems have shown significant improvement in speech synthesis, there is no TTS system to learn synthesize from text sequences with only feedback. Because feedback alone not sufficient train the generator, current models still require reconstruction loss compared ground-truth and generated mel-spectrogram directly. I...

Texture has long been recognized in computer vision as an important monocular shape cue, with texture gradients yielding information on surface orientation. A more recent trend is the analysis of images in terms of local spatial frequencies, where each pixel has associated with it its own spatial frequency distribution. This has proven to be a successful method of reasoning about and exploiting...

This paper describes the application of morphological filtering to speech spectrograms for noise robust automatic speech recognition. Speech regions of the spectrogram are identified based on the proximity of high energy regions to neighbouring high energy regions in the three-dimensional space. The process of erosion can remove noise while dilation can then restore any erroneously removed spee...

Visual perception of speech through spectrogram reading has long been a subject of research, as an aid for the deaf or hearing impaired. Attributing the lack of success in this type of visual aids mainly to the static form of information presented by the spectrograms, this paper proposes a system of dynamic visualisation for speech sounds. This system samples a high resolved, auditorybased spec...

This paper presents the addition to the Stanley’s habituation model of a previous stage based on spectrogram to detect temporal patterns in a signal and to obtain a measure of habituation to these patterns. With this addition we achieve a habituation scheme that saturates as the temporal pattern is perceived by the system and drops when the pattern changes. The detection of these temporal patte...