The proposal stated 6 effects for anonymization: pitch, bass and treble, distortion, echo, reverb, and wah-wah effect

These are implemented for audacty. Although there exists a scripting API for audacity, the script requires audacity to be running in background. This condition is not favourable to us.

Can these be done using some other python library? No. I have checked a lot of libraries on https://github.com/faroit/awesome-python-scientific-audio however, none of them supported something apart from pitch and noise. The python-audio-effects (https://github.com/carlthome/python-audio-effects)[pysndfx] supported all of them, but it does not work at all.

pysndfx is build on SoX, which has a wrapper https://pysox.readthedocs.io/en/latest/example.html .

Pitch: pitch(n_semitones: float, quick: bool = False)

Treble: treble(gain_db: float, frequency: float = 3000.0, slope: float = 0.5)

Bass: bass(gain_db: float, frequency: float = 100.0, slope: float = 0.5)

Distortion: overdrive(gain_db: float = 20.0, colour: float = 20.0)

Echo: echo(gain_in: float = 0.8, gain_out: float = 0.9, n_echos: int = 1, delays: List[float] = [60], decays: List[float] = [0.4])

echos(gain_in: float = 0.8, gain_out: float = 0.9, n_echos: int = 1, delays: List[float] = [60], decays: List[float] = [0.4])

reverb(reverberance: float = 50, high_freq_damping: float = 50, room_scale: float = 100, stereo_depth: float = 100, pre_delay: float = 0, wet_gain: float = 0, wet_only: bool = False)

Qualitative Analysis of SoX

The anonymization predominantly is effective due to pitch change. Chorus and echo do create an effect, however, they also make audio unnatural

Further complex effects like bass, treble, reverb, flanger, etc. create more unnaturalness as compared to anonymization.

Hence, such effects are better to be used as “Transformers” as compared to “Anonymizer”. Some discussions on the internet emphasise that pitch is the basic attribute to be changed for anonymization, and it will surely work. But, it will be better we have more than one attribute for anonymization.

The question arises: what are differences between audios of same pitch? can we leverage those for further anonymization.

Can randomly increasing/decreasing pitch, lead to anonymization?

Yes, but in some cases. The catch is: male voice is low and female voice is high. Increasing the high voice further, does not affect much. Similarly, for male voice.

We need to compute the pitch of average male and female voices. Given an audio, we need to guess its gender based on the average pitch data. From this, we can estimate the increase/decrease required.

Important pages:

  • https://pyra-handheld.com/boards/threads/help-with-using-sox-as-a-voice-changer.98878/
  • https://askubuntu.com/questions/421947/is-there-a-way-to-modulate-my-voice-on-the-fly
  • https://security.stackexchange.com/questions/227146/is-changing-pitch-enough-for-anonymizing-a-persons-voice