RASTA processing of speech

By Hynek Hermansky, Nelson Morgan. 1994

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

Citations in the corpus

Time-Frequency Masking in the Complex Domain for Speech Dereverberation and Denoising (2017, neutral): “These features include amplitude modulation spectrogram (AMS) [23], relative spectral transform and perceptual linear prediction (RASTA-PLP) [14], [15], mel-frequency cepstral coefficients (MFCC), as well as their deltas.”
An overview of text-independent speaker recognition: From features to supervectors (2010, background): “RelAtive SpecTrAl (RASTA) filtering (Hermansky and Morgan, 1994; Malayath et al., 2000) applies a bandpass filter in the log-spectral or cepstral domain.”
Continuous latent variable models for dimensionality reduction and sequential data reconstruction (2001, methodology): “We extracted 12th-order perceptual linear prediction (PLP) coefficients (Hermansky, 1990) from the waveform (plus the log-gain, which we use in some experiments) with the RASTA program (Hermansky and Morgan, 1994); we used a window size of 25.”
Learning filter banks within a deep neural network framework (2013, result): “Given the constraints that we want to normalize the power spectral features, but want the output of the filter bank stage to be positive, we explore an idea that is similar to that done in RASTA processing [17].”
MVA Processing of Speech Features (2007, methodology): “In the case of corruption, techniques such as Kalman filters [6] and microphone arrays [7] are often used to recover clean speech samples.”
Joint noise adaptive training for robust automatic speech recognition (2014, background): “• 13 dimensional RASTA filtered perceptual linear predictive cepstral coefficients (RASTA-PLPs) [21].”
Learning words from sights and sounds: a computational model (2002, methodology): “Acoustic input is first converted into a spectral representation using the Relative Spectral-Perceptual Linear Prediction (RASTA-PLP) algorithm (Hermansky & Morgan, 1994).”
rVAD: An Unsupervised Segment-Based Robust Voice Activity Detection Method (2019, methodology): “rsal background model (GMM-UBM) since GMM based methods are known to outperform the i-vector technique for SV with short utterances. MFCCs of 57 dimensions (including static, and ) with RASTA ltering [75] are extracted from speech signals with a 25 ms hamming window and a 10 ms frame shift. After VA”
ABX-Discriminability Measures and Applications. (Mesures de discriminabilité ABX et applications) (2016, methodology): “Next, one of 16 possible representations is derived by making a succession of 4 binary choices: use a linear or a Mel frequency scale; weight frequency channels according to human’s equal-loudness contour or not; cubic root compress the dynamic range of frequency channels or not; apply log-domain RA”
Connectionist speech recognition of Broadcast News (2002, methodology): “The development of MSG features originated in experience with RASTA processing (Hermansky and Morgan, 1994), which also used an auditory-like spectral analysis, and achieved robustness though band-pass ﬁltering of subband energy envelopes.”
Monitoring temporal change of bird communities with dissimilarity acoustic indices (2014, methodology): “For all field recordings, 100 Mel-frequency cepstral coefficients (MFCC) were calculated using the R package tuneR (Ligges 2011), with a frame shift of 10 ms, a window length of 25 ms and a liftering exponent of 0 6, leading to 256 frequential bands Mel spectrograms (Hermansky & Morgan 1994).”
Improving Robustness of Deep Neural Network Acoustic Models via Speech Separation and Joint Adaptive Training (2015, methodology): “The following feature set is extracted at every time frame from the noisy input signal: • 13 dimensional relative spectral ﬁ ltered perceptual linear predictive cepstral coef ﬁ cients (RASTA-PLP) [10].”
Robust GMM Based Gender Classification using Pitch and RASTA-PLP Parameters of Speech (2006, methodology): “So that, the relative spectral perceptual linear predictive (RASTA-PLP) coefficients [8] combined with pitch are chosen for the parameters of our proposed gender classification system.”
Exploring Monaural Features for Classification-Based Speech Segregation (2013, background): “RASTA filtering [13] is often coupled with PLP for robust speech recognition.”
Developments and directions in speech recognition and understanding, Part 1 [DSP Education] (2009, methodology): “…motivated mel-frequency cepstral coefficients (MFCC) [10], [29] and perceptual linear prediction (PLP) coefficients [21], as well as normalizations via cepstral mean subtraction (CMS) [16], [44], relative spectral (RASTA) filtering [20], and vocal tract length normalization (VTLN) [13].”
Updated MINDS Report on Speech Recognition and Understanding, Part 2 (2009, background): “Many factors unrelated to the information being sought by the models also significantly influence the signal presented to the system.”
Capture, Learning, and Synthesis of 3D Speaking Styles (2019, methodology): “Brand [7] uses a mix of Linear Predictive Coding (LPC) and RASTA-PLP [30] audio features and an HMM to output a sequence of facial motion vectors.”
Discriminative Training for Large-Vocabulary Speech Recognition Using Minimum Classification Error (2007, methodology): “The feature vectors consist of 38 or 39 components: MFCCs (after cepstral mean subtraction), deltas, and delta–deltas, computed every 10 ms for 25-ms windows over the speech signal.”