Comparison of Parametric Representation for Monosyllabic Word Recognition in Continuously Spoken Sentences

By M. Padmanabhan, M. Picheny. 2017

How researchers see this paper

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

In Speech & Audio

Citations in the corpus

Librispeech: An ASR corpus based on public domain audio books (2015, methodology): “For this first decoding pass we use a triphone model discriminatively trained with Boosted MMI [12], based on MFCC [13] features processed with frame-splicing over 7 frames, followed by LDA, followed by a global semi-tied covariance (STC) transform [14] .”
Unsupervised acoustic unit discovery for speech synthesis using discrete latent-variable neural networks (2019, methodology): “For this same reason we use MFCCs, which tend to be speaker-independent [35], as input, while ﬁtlerbanks, which retain speaker-speciﬁc properties, are used as intermediate output.”
Learning Filterbanks from Raw Speech for Phone Recognition (2017, background): “Speech features such as gammatones or mel-filterbanks (MFSC, for mel-frequency spectral coefficients) were designed to match the human perceptual system [1, 2], and contain invaluable priors for speech recognition tasks.”
Natural TTS Synthesis by Conditioning Wavenet on MEL Spectrogram Predictions (2017, methodology): “Because of these properties, features derived from the mel scale have been used as an underlying representation for speech recognition for many decades [17].”
Parallel training of Deep Neural Networks with Natural Gradient and Parameter Averaging (2014, methodology): “The actual inputs to the DNN in this setup normally consist of the iVector, plus ± 7 frames of Mel frequency cepstral coeﬃcients (MFCCs) [32], without cepstral mean normalization.”
Learning Problem-agnostic Speech Representations from Multiple Self-supervised Tasks (2019, background): “Despite being proposed more than 40 years ago [40], these coefﬁcients are still the most common speech features, and it is not easy to ﬁnd alternatives that consistently outperform them.”
JHU ASpIRE system: Robust LVCSR with TDNNS, iVector adaptation and RNN-LMS (2015, methodology): “Mel-frequency cepstral coefﬁcients (MFCCs) [11], without cepstral truncation, were used as input to the neural network i.e., 40 MFCCs were computed at each time step.”
A Review on Automatic Speech Recognition Architecture and Approaches (2016, methodology): “ Delta MFCC Features – In order to capture the changes in speech from frameto-frame, the first and second derivative of the MFCC coefficients are also calculated and used [6].”
Quaternion Recurrent Neural Networks (2018, background): “…automatic speech recognition systems usually employ input sequences composed of multidimensional acoustic features, such as log Mel features, that are often enriched with their first, second and third time derivatives (Davis & Mermelstein, 1990; Furui, 1986), to integrate contextual information.”
A New Feature for Automatic Speaker Verification Anti-Spoofing: Constant Q Cepstral Coefficients (2016, methodology): “The Mel-cepstrum applies prior to cepstral analysis a frequency scale based on auditory critical bands [34].”
Constant Q cepstral coefficients: A spoofing countermeasure for automatic speaker verification (2017, methodology): “The Mel-cepstrum applies prior to cepstral analysis a frequency scale based on auditory critical bands (Davis and Mermelstein, 1980).”
ITRW on Multilingual Speech and Language Processing Stellenbosch , South Africa April 9-11 , 2006 (unknown): “The chosen parameters have been used extensively [6] and have proved to be one of the best choices for HMM-based speech recognition systems.”
A Mathematical Theory of Deep Convolutional Neural Networks for Feature Extraction (2015, background): “Furthermore, it was argued in [23], [25], [47] that the feature vector Φ 1 W ( f ) generated in the ﬁrst layer of the scattering network is very similar, in dimension d = 1 , to mel frequency cepstral coefﬁcients [48], and in dimension d = 2 to SIFT-descriptors [49], [50].”
A time delay neural network architecture for efficient modeling of long temporal contexts (2015, methodology): “Mel-frequency cepstral coefficients (MFCCs) [16], without cepstral truncation, were used as input to the neural network i.”
GELP: GAN-Excited Liner Prediction for Speech Synthesis from Mel-spectrogram (2019, methodology): “In the present Tacotron style TTS systems [8], the most commonly used acoustic feature is the log-compressed mel-spectrogram, that utilizes the classical MFCC triangular ﬁlter-bank [23].”
Semantic Speech Retrieval With a Visually Grounded Model of Untranscribed Speech (2017, background): “Mel-frequency cepstral coefficients (MFCCs) [82].”
Audio augmentation for speech recognition (2015, methodology): “Mel-frequency cepstral coefficients (MFCCs) ([13]), without cepstral truncation, were used as input to the neural network.”
A comparison of features for synthetic speech detection (2015, background): “In MFCC, the ﬁlters are placed in mel scale, having denser spacing in the low-frequency region [21].”
Power-Normalized Cepstral Coefficients (PNCC) for Robust Speech Recognition (2016, methodology): “compares the structure of conventional MFCC processing [12], PLP processing [13], [24], and the new PNCC approach which we introduce in this paper.”