pncc_asr

An implementation of "Power-Normalized Cepstral Coefficients (PNCC) for Robust Speech Recognition(Chanwoo Kim, 2016)".

This is unofficial code. Original code is implemented by C (please see the paper).

Usage

• This code is based on torchaudio; the audio data have to be loaded by torchaudio.
• All parameters follow the origianl paper.
• You can use GPU computation in this code.

First, Clone this repository and install requirements.

cd ~
git clone https://github.com/LimDoHyeon/pncc_asr.git
cd pncc_asr
pip install -r requirements.txt

And use this function:

def pncc(
    wav: torch.Tensor,
    sr: int,
    n_fft: int = 1024,
    win_length: int = 400,
    hop_length: int = 160,
    n_ch: int = 40,
    f_min: int = 200,
    f_max: int = 8000,
    n_ceps: int = 13,
) -> torch.Tensor:

import torchaudio
import pncc_asr as PNCC

y, sr = torchaudio.load('your_audiofile.wav')  # or you can use dataloader
pncc = PNCC.pncc(y, sr)

Author: Do-Hyeon Lim