voice-morphing

Voice morphing currently based on gender detector/age regressor & spectral features modification [SFFT, Griphin-Lim argoritmn].
According to https://www.researchgate.net/publication/269562531_The_Aging_Voice (and amount of another medical research papers), male and female voice changes with age in different ways.

Articulation feature (AF)	Male	Female
pitch	higher	lower
volume	lower	lower
tremor	higher	higher

1. Baseline
   Train female/male voice classifier on TIMIT dataset and then manually change voice parametres, e.g. pitch/volume/tremor. To analyze how to change pitch/volume/tremor, we need a speech factorizer. Age regressor and Gender Detector can help to annotate large amounts of data for future training. As a baseline, firstly I trained gender classifier and age regressor to choose voice morphing strategies. I tuned hyperpams for pitch shifter (SFFT+Griphin-Lim, in future can be replaced with vocoder e.g. WaveGlow) taking into consideration predicted age from age_regressor.
2. Future experiments
   a. Replace Griphin-Lim with vocoder
   b. Create telegram-bot for testing
   c. Mine pseudolabeled data via trained detector and regressor on TIMIT (WSJ, LibriVox, Youtube)
   d. http://www.apsipa.org/proceedings/2019/pdfs/68.pdf [VAE with speaker individuality block]
   e. https://www.isca-speech.org/archive/Interspeech_2019/pdfs/1778.pdf [accent-changer PPG approach]
   f. https://github.com/SerialLain3170/VoiceConversion [Cycle-GAN]

Current results

First results can be found in test folder in this repo. It contains female voice (recorded from macbook Pro microphone) and male voice (recorded as Telegram audiomessage + applied noise reduction filter). Then morphed files passed through age_regressor pipeline trying to guess age of fake-aged speaker.
Live-demo is available in google-colaboratory demo.ipynb.

Articulation feature (AF)	Male	Female
original	22	24
old	46	51
young	18	17