libllsm | Low Level Speech Model for high quality speech analysis

libllsm is a C library. libllsm has no bugs, it has no vulnerabilities, it has a Strong Copyleft License and it has low support. You can download it from GitHub.

The following code initializes analysis parameters for libllsm. The only argument of llsm_init is the number of bands for noise excitation (the way LLSM models noise excitation is basically a multi-band extension to [2]). Given some F0 estimation stored in a float-point array f0, we first call llsm_layer0_analyze on the input signal to obtain the layer 0 representation. The harmonic parameters in the layer 0 model are in absolute phase, which is somewhat inconvenient to manipulate (consider pitch shifting or interpolation). We apply a time shift to each frame so that the phases are made relative to the first harmonic. This is called Relative Phase Shift (RPS) [3]. Then we go from layer 0 up to layer 1 by calling llsm_layer1_from_layer0. The layer 1 model contains separate information about vocal tract and source. For pitch shifting we simply need to keep the layer 1 model intact and resample the vocal tract transfer function at scaled harmonic frequencies. Note that the layer 1 model is for harmonic component only, so the layer 0 model is still relevant and it should not be discarded at this point. Next in the for loop over frames, we first scale the F0 and make an array of harmonic frequencies. Then as described above, vocal tract and lip frequency responses are subsampled at new frequencies; the vocal tract phase response is computed from amplitudes under minimum phase assumption. These parts are re-combined by multiplication and the result is written back to layer 0. The origf0 / iframe → f0 term compensates for the amplitude gain due to change in number of harmonics within audible range. At this point pitch shifting is done on layer 0. But before synthesis we first need to recover the phase progression along time axis (which is the integral of F0). Finally we call llsm_layer0_synthesize to convert from layer 0 model back to signal. The output is a structure containing harmonic and noise components of the synthesized speech. libllsm is licensed under GPLv3. I have a pending patent on LLSM-related technology. However the patent license is granted to libllsm users, free from royalty, under the terms of GPLv3. Please contact the author for an alternatively licensed version primarily for commercial purposes. Currently there’s no publication directly associated with LLSM. However there is [a poster] on the pseudo glottal inverse filtering method in layer 1 LLSM. The following are the major publications that LLSM draws inspiration from.