ljuvela/glottdnn

GlottDNN vocoder and tools for training DNN excitation models

glottdnn's Issues

TODO: CreateExcitation: Exactness of PSOLA

The pitch-synchronous overlap-add (PSOLA) windowing used in CreateExcitation should be explored to produce a uniform windowing effect.

Ideas:
-Create a "ApplyPsolaWindow()" function that takes in the pitch period of previous and next frames, and then creates the left-hand and right-hand -sides of the window accordingly. This should work as such with Rect-windowed frames
-Problem: Can we apply this principle to cosine windowed pulses? Or should we try to have always rect-windowed pulses as initial excitation?

TODO: F0 Post filtering

TODO: Filter Excitation: GetFilteringGain()

-Should GetFilteringGain() include target gain of frame?

TODO: CreateExcitation: Pulses-as-features excitation

TODO: DNN training block

Theano data provider Python script
Merlin compatibility?

TODO: FilterExcitation: Add warped filtering

TODO: "Fix LSF" function

TODO: HnrAnalysis: Explore better ways to obtain noise floor estimates

(1) Cepstral processing?
(2) Frequency domain "sinusoidal synthesis" (construct harmonic peaks + amplitudes, convolve with analysis window)?
(3) Something else? What?

HNR modification fails for low F0 values

If input F0 to Synthesis contains low but non-zero values (around 10Hz or so), the program will crash and give the following error:

gel: interp.c:84: ERROR: x values must be monotonically increasing
HNR modification ...Default GSL error handler invoked.

This is caused by multiples of F0 in harmonics analysis falling into a single FFT bin (hence x values not monotonically increasing).

User fix: manually clip your F0 values to be above 40Hz.

Developer fix: clip F0 values in Synthesis and add a warning message.

Extracting excitation signal

In the config file, under # File extensions for parameters (optional) section, there is .exc.wav extension. I think that represents the excitation signal. But when I'm running the Analysis command, I'm not getting any excitation signal. Can you please tell how to extract the Excitation signal.?

TODO: Analysis: Fix QMF analysis

TODO: Moving average filtering

Failsafe for all-zero LSF input features in Synthesis

If LSF type input features within a frame are all zero, the Synthesis program should interpret this as a neutral filter (and maybe give a warning).

All-zero LSFs commonly arise from zero-padding during modeling. Although they are not valid LSFs, they should be treated gracefully. Currently, Synthesis tries to fix the LSFs, which results in sharply resonant filters and loud outputs.

TODO: HarmonicModification: What is the best way to generate noise

(1) Additive noise with set magnitude spectrum and random phase (current implementation)
(2) Additive noise with a set energy and shaping with the spectral envelope
(3) Modulated noise (modulation in the spectral domain = convolution in the time domain) (legacy implementation)
(4) Something else? What?

Error when LPC_ORDER_VT or LPC_ORDER_GLOT are set to an odd number

When LPC_ORDER_VT or LPC_ORDER_GLOT are set to an odd number, I get the following when running the Analysis script

Spectral analysis ... done.
HNR Analysis ... done.
gsl: zsolve.c:60: ERROR: leading term of polynomial must be non-zero
Default GSL error handler invoked.
Aborted

TODO: ParameterSmoothing

-Moving average filtering, median filtering, or both?

TODO: Analysis: Fix WLP function

-Problems with stabilization function(?)

TODO: Add Leja ordering to high-order LSFs.

TODO: CreateExcitation: DNN excitation

-Class for the DNN
-Read DNN

TODO: PostFilter POWER_SPECTRUM_WIN coefficient

Should this be FS dependent? If so, how?

Conda environment

Hello,
I've installed the dependencies using a conda environment :

conda install -c conda-forge libsndfile --yes
conda install -c conda-forge gsl --yes
conda install -c conda-forge libconfig --yes

However, when I try make, I get the errors "libconfig.h++: No such file or directory", as well as "sndfile.hh: No such file or directory".

Any ideas anybody? Is it possible to use a conda environment with GlottDNN? Thanks a lot.

Kind regards, Philipp Aichinger