Mel-generalized cepstral analysis - a unified approach to speech spectral estimation

By Keiichi Tokuda, Takao Kobayashi, Takashi Masuko, Satoshi Imai. 1994

The generalized cepstral analysis method is viewed as a unified approach to the cepstral method and the linear prediction method, in which the model spectrum varies continuously from all-pole to cepstral according to the value of a parameter γ. Since the human ear has high resolution at low frequencies, introducing similar characteristics to the model spectrum, we can represent speech spectrum more efficiently. From this point of view, this paper proposes a spectral estimation method which uses the spectral model represented by mel-generalized cepstral coefficients. The effectiveness of mel-generalized cepstral analysis is demonstrated by an experiment of HMM-based isolated word recognition.

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

A Survey on Neural Speech Synthesis (2021, methodology): “2) Simplifying acoustic features, where the complicated acoustic features such as MGC, BAP and F0 used in SPSS are simplified into mel-spectrograms.”
Neural Source-Filter Waveform Models for Statistical Parametric Speech Synthesis (2019, methodology): “Two types of spectral features were prepared: Mel-generalized cepstral coefficients (MGCs) [52] of order 60 extracted using the WORLD vocoder [53] and Mel-spectrogram of order 80.”
Neural Source-filter-based Waveform Model for Statistical Parametric Speech Synthesis (2018, methodology): “Acoustic features, including 60 dimensions of Melgeneralized cepstral coefficients (MGCs) [18] and 1 dimension of F0, were extracted from the 48 kHz waveforms at a frame shift of 5 ms using WORLD [19].”
A Survey on Audio Diffusion Models: Text To Speech Synthesis and Enhancement in Generative AI (2023, unknown): “Common acoustic features include mel-cepstral coefficients [18], Mel-generalized coefficients [114], F0 [34] and band aperiodicity [33].”
Multi-distribution deep belief network for speech synthesis (2013, background): “Speech parameters for synthesis include the spectrum and the log-F0 with assumed Gaussian distribution, and the voicedunvoiced switches which are essentially binary.”
Tacotron-Based Acoustic Model Using Phoneme Alignment for Practical Neural Text-to-Speech Systems (2019, methodology): “Both mel-spectrograms and vocoder features (VF) constructed from the fundamental frequency and melcepstra [55] were also evaluated as acoustic features for conventional pipeline SPSS systems.”
A Neural Parametric Singing Synthesizer (2017, methodology): “The dimensionality of the harmonic component is reduced to 60 co-efﬁcients by truncated frequency warping in the cepstral domain (Tokuda et al., 1994), using an all-pole warping coefﬁcient α = 0 .”
GlotNet—A Raw Waveform Model for the Glottal Excitation in Statistical Parametric Speech Synthesis (2019, methodology): “Various acoustic features have been utilized for WaveNet conditioning, including mel-filterbank energies [14] and mel-generalized cepstral coefficients (MGC) [15] in combination with fundamental frequency (F0) [16].”
Thousands of Voices for HMM-Based Speech Synthesis–Analysis and Application of TTS Systems Built on Various ASR Corpora (2010, methodology): “The mel-generalized cepstra are similar to PLP features in terms of spectral representation [64].”
Cyclic Spectral Modeling for Unsupervised Unit Discovery into Voice Conversion with Excitation and Waveform Modeling (2020, methodology): “Specifically, 49th dimensional mel-cepstrum parameters [28] including 0-th Figure 2: Flow of the proposed frameworks to perform discovery of subword units using the latent space representations of CycleVAE/CycleVQVAE spectral model, and into voice conversion with additional excitation module and Wav”
Speech Synthesis Based on Hidden Markov Models (2013, background): “In addition to the mel-cepstral coefficients, various spectral representations, such as line-spectral pairs (LSPs) [43], mel-generalized cepstral coefficients [44], and various excitation parameters (e.”
Voice Morphing that improves TTS quality using an optimal dynamic frequency warping-and-weighting transform (2016, background): “Following [2], the speech parameterization consists of 40-dimensional MCEP (including power), log f0 and a 7-dimensional band aperiodicity.”
FoundationTTS: Text-to-Speech for ASR Customization with Generative Language Model (2023, background): “Human speech naturally has a random sampling effect like white noise in source-filter vocoders [24, 25, 26, 27, 28] that status linguistic or acoustic condition is not enough to simulate.”
A Vector Quantized Variational Autoencoder (VQ-VAE) Autoregressive Neural $F_0$ Model for Statistical Parametric Speech Synthesis (2020, methodology): “Speech waveforms for listening tests were synthesized using the WORLD vocoder [49] given natural Mel-generalized cepstral coefficients [50] (60 coefficients per frame) and band ape-”
A spectral envelope estimation method based on F0-adaptive multi-frame integration analysis (2012, methodology): “LPC has been extended to melgeneralized cepstral analysis for treating various spectral representations [7], and the cepstrum-based methods have also been improved [8–10] and combined with the LPC method [11].”
Statistical approach to vocal tract transfer function estimation based on factor analyzed trajectory HMM (2008, methodology): “We evaluated the VTTF estimation accuracy of mel-cepstral analysis [2], STRAIGHT analysis [3], and the proposed analysis by comparing the VTTFs estimated from the re-synthesized speech samples with true VTTFs, i.”