4

By performing the windowing with overlap we are artificially increasing our time resolution (larger granularity of features in time). This is especially useful when frame duration is long (bad time resolution, very good frequency resolution), thus yielding kind of extra 'time resolution'. Usually no one is using the rectangular window, but other types such ...


3

Ok so here is what I found. The distance is dependent on the way that the mfccs are calculated. This makes sense to me and also explains why cepstral mean normalization affects the values of the MCD. I found this implementation (https://github.com/MattShannon/mcd), which unfortunately did not support .wav files. I ran this and it gives results that are in ...


2

TL/DR: For a 2nd order transfer function of a real digital filter given as: $$H(z) = \frac{1}{1-2A_1z^{-1}-A_2z^{-2}}$$ The filter coefficient $A_1$ and $A_2$ are determined from the gain and resonant frequency to be: $$A_1 = Kcos(2\pi f_r/f_s) = Kcos(\omega_r)$$ $$A_2 = -(K^2)$$ Where: $f_r$: Resonant frequency in Hz $f_s$: Sampling frequency in Hz ...


2

Here are a bunch of noise samples used by a military speech research unit: http://spib.rice.edu/spib/select_noise.html They are free to download, and are available in both wav and Matlab binary format. If your speech material is not recorded at the same sample rate as the noise samples, it is important to resample either the speech, or noise, or both ...


1

I'm not sure that the formula posted in the question is accurate. According to the paper which introduced the measure (eq. 1) and to this paper (eqs. 1a and 1b), the sum over the cepstral coefficients (denoted by $i$) should be inside the square root; that is: $$ \frac{10 \sqrt{2}}{\ln 10} \frac{1}{T} \sum_{t=1}^T \sqrt{\sum_{i} \left(C_{ti} - \hat{C}_{ti}\...


1

Let $t_a(n)$ be the time of the n-th analysis mark ("Pitch marker"), $P_a(n)$ the period at the n-th analysis mark (it can be defined as $t_a(n+1) - t_a(n)$, $t_s(n)$ be the time of the n-th synthesis mark, $n_s(k)$ is a number whose integer part is the index of analysis segment to be overlapped-added when processing the n-th synthesis mark (it can also ...


1

The "Harvard sentences" corpus is a classic: https://en.wikipedia.org/wiki/Harvard_sentences


1

There is a correlation and it can be learned with machine learning methods. A simple google search can give you many papers, recent research Waveform Modeling and Generation Using Hierarchical Recurrent Neural Networks for Speech Bandwidth Extension Zhen-Hua Ling, Yang Ai, Yu Gu, Li-Rong Dai older paper: Techniques for the Regeneration of Wideband Speech ...


1

Have you had a look at the diphone synthesizer resources in the Festival Speech Synthesis framework? The state of the art for Text2Speech (though still not practical for production due to GPU compute requirements) is Deepmind's WaveNet. Their results are definitely smoother than concatenative or parametric methods. Good luck!


1

I guess you confuse GMM and HMM trainings. Although in both cases EM algorithm is employed, Baum-Welch is used for HMM training.


1

For observations people usually use a mixture of gaussians, not a simple gaussian. They have few advantages - EM algorithm is fast to converge and GMM approximates wide variety of distributions pretty well. Probability of the model can be computed efficiently with gaussian selection. Last, GMMs are easy to cluster for context-dependent tree for phonetic ...


1

I seriously doubt that there is a simple way to accomplish this. Most of the work that I have seen tends to go in the direction of improving the synthesis, or using better (and more) phoneme recordings to get better quality. If there were a simple way, I think someone would have found it by now rather than putting so much more effort into better ...


Only top voted, non community-wiki answers of a minimum length are eligible