add mos / snr prediction model

[felixbur]

to be used for data quality and bias checking

[felixbur]

for examples this one: https://github.com/microsoft/DNS-Challenge/tree/master/DNSMOS

[bagustris]

Torchaudio already has this automatic MOS prediction tool (including PESQ, STOI, and SI-SDR) presented at ICASSP 2023 [1]. SNR estimation is not implemented yet and will be a good tool.

[1] A. Kumar et al., “TorchAudio-Squim: Reference-less Speech Quality and Intelligibility measures in TorchAudio,” pp. 2–6, 2023, [Online]. Available: http://arxiv.org/abs/2304.01448.

[bagustris]

@felixbur

I have a working code to estimate SNR here: https://github.com/bagustris/audio-SNR. The error is about 2 dB (20 dB SNR is estimated as 22 dB). Show me how to incorporate that codes into Nkululeko if you agree.

[felixbur]

super. I haven't though yet of a way how to use that. It is part of a larger idea to enable Nkululeko to add speaker/speech characteristics from publicly available models to the databases, in order to check their bias on the target feature. An example: you have a database that is labeled with depression and speaker id, but nothing else. You would like to use public models for

• age prediction
• sex prediction
• snr prediction

to automatically label your data, and then use the explore module to check on biases between the target (depression) and the three added features (age, sex and SNR). I guess it would be easiest to implement this with a new, dedicated, module, like e.g. segmentation or augmentation. we could name it, e.g. "autopredict"

[felixbur]

I think best is to start with a template/example for a "add_speech_feat" class that takes dataframes and adds the predicted labels

I planned to do that for age/gender, but perhaps you start with snr