Mel-cepstral distance measure for objective speech quality assessment

By R. Kubichek. 1993

Subfields: Natural Language Processing, Speech & Audio

How researchers see this paper

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

In Natural Language Processing

1. Junichi Yamagishi
2. Tao Qin
3. Tie-Yan Liu
4. Xu Tan
5. Heiga Zen
6. Hung-yi Lee
7. Yu Zhang
8. Frank K. Soong

In Speech & Audio

1. Xu Tan
2. Junichi Yamagishi
3. Hung-yi Lee
4. Tao Qin
5. Wei-Ning Hsu
6. Tie-Yan Liu
7. Yu Zhang
8. Sheng Zhao

Citations in the corpus

• WaveGrad: Estimating Gradients for Waveform Generation (2020, methodology): “Mel cepstral distance (MCD) (Kubichek, 1993), a similar MSE metric computed using 13dimensional mel frequency cepstral coefficient features; 3.”
• Voicebox: Text-Guided Multilingual Universal Speech Generation at Scale (2023, background): “Prior studies often adopt subjective metrics like mean opinion scores (MOS) [52] which are not comparable across papers, or quantitative metrics like mel cepstral distortion [35] that assume the output is deterministic given input, which is often not realistic [54].”
• Semi-supervised Training for Improving Data Efficiency in End-to-end Speech Synthesis (2018, methodology): “The results of SxS tests are consistent to those of MCD objective tests, and both demonstrate the effectiveness of our semi-supervised framework.”
• Towards End-to-End Prosody Transfer for Expressive Speech Synthesis with Tacotron (2018, background): “Mel Cepstral Distortion (MCDK) (Kubichek, 1993): MCDK = 1 T T−1∑ t=0 √√√√ K∑ k=1 ( ct,k − c′t,k )2 Where ct,k,c′t,k are the k-th mel frequency cepstral coefficient (MFCC) of the t-th frame from the reference and predicted audio.”
• GAN Vocoder: Multi-Resolution Discriminator Is All You Need (2021, methodology): “Before computing MCD, all audio clips are normalized to have mean of 0 and standard deviation of 1 to minimize the volume differences.”
• BigVGAN: A Universal Neural Vocoder with Large-Scale Training (2022, background): “6 3) Mel-cepstral distortion (MCD) (Kubichek, 1993) with dynamic time warping which measures the difference between mel cepstra.”
• Location-Relative Attention Mechanisms for Robust Long-Form Speech Synthesis (2019, methodology): “The MCD-DTW is an objective similarity metric that uses dynamic time warping (DTW) to find the minimum mel cepstral distortion (MCD) [16] between two sequences.”
• InstructTTS: Modelling Expressive TTS in Discrete Latent Space With Natural Language Style Prompt (2023, methodology): “For speech quality, we adopt Mel-cepstral distorion (MCD) [66], structural similarity index measure (SSIM) [67] and Short-Time Objective Intelligibility (STOI) [68] to evaluate the speech quality.”
• Generating Diverse and Natural Text-to-Speech Samples Using a Quantized Fine-Grained VAE and Autoregressive Prosody Prior (2020, methodology): “The QFVAE models have higher FFE and MCD than the baseline, however their performances improve as the number of classes increased.”
• ProDiff: Progressive Fast Diffusion Model for High-Quality Text-to-Speech (2022, methodology): “For objective evaluation, ProDiff also demonstrates the outperformed performance in MCD, PESQ, and STOI, superior to all baseline models.”
• Bilateral Denoising Diffusion Models (2021, positive): “Specifically, we used the objective metrics in [2] – the log-Mel spectrogram mean squared error (LS-MSE) and the Mel-cepstral distance (MCD) [43] to assess the consistency between the original waveform and the generated waveform in the Mel-frequency domain.”
• Pretraining Techniques for Sequence-to-Sequence Voice Conversion (2020, methodology): “Similar to the calculation of MCD, DTW-based is performed and we take only the non-silent frames into account.”
• Improve GAN-based Neural Vocoder using Pointwise Relativistic LeastSquare GAN (2021, methodology): “Starting from the baseline model, MelGAN and PWGAN, which apply LSGAN and STFT loss, we added each of the proposed methods one at a time measuring MCD, FFE, and PESQ.”
• Wave-Tacotron: Spectrogram-Free End-to-End Text-to-Speech Synthesis (2020, methodology): “We also compute several objective metrics: (1) Mel cepstral distortion (MCD), the root mean squared error against the ground truth signal, computed on 13 dimensional MFCCs [30] using dynamic time warping [31] to align features computed on the synthesized audio to those computed from the ground truth”
• DiffGAN-TTS: High-Fidelity and Efficient Text-to-Speech with Denoising Diffusion GANs (2022, methodology): “We conducted an objective evaluation using the structural similarity index measure (SSIM) (Wang et al., 2004), mel-cepstral distortion (MCD)(Kubichek, 1993), F 0 root mean squared error (RMSE), and voice Cosine simi-2 https://github.com/jik876/hifi-gan larity.”
• Adaptive Wavenet Vocoder for Residual Compensation in GAN-Based Voice Conversion (2018, methodology): “Such MCD doesn’t evaluate the effect of vocoders.”
• Fast DCTTS: Efficient Deep Convolutional Text-to-Speech (2021, background): “Mel cepstral distortion(MCD) [20] is known effective in measuring perceptual distance between speech signals and expressed as”