spectral similarity distance

[cassiotbatista]

without depending o #4 which may take a while an idea for quick simulation and quality assessment of the TTS acoustic model could be add some similarity measure at the spectrogram level

• https://dsp.stackexchange.com/questions/47885/quantify-the-difference-between-two-audio-spectrograms

ground truth would have to be generated with the same parameters used for FastSpeech 2 tho, which can be a burden finding out an exact match

[cassiotbatista]

hmm MSD might be just it https://github.com/AI-Unicamp/TTS-Objective-Metrics

[cassiotbatista]

also my old codes https://gitlab.com/fb-tts/fb-uttcopy/dne-uttcopy/-/tree/master/metrics