README.md

Librosa_embedded_mfcc

procedures from Librosa library for mfcc embeddings calculation of sound files for Raspberry Pi 64bit OS

Problem why this part of Librosa have made

Librosa can't be installed (now 2021-08-10) for RK3399 and Raspberry 64-bit OS (numba and lliblvm problems). But with pytorch model we need to made MFCC embeddings which is used for inference and sound classification.

Dependencies need to be installed

#dependencies for custom librosa procedures
pip3 install soundfile
pip3 install audioread
pip3 install samplerate
pip3 install soxr

Example of use

'''make a class prediction for one row of data'''
def predict(row, model):
    # convert row to data
    row = Tensor([row])
    # make prediction
    yhat = model(row)
    # retrieve numpy array
    yhat = yhat.detach().numpy()
    return yhat

model = torch.load('torch_MFCC.pt')

'''make a single prediction'''
name_arr = ['event', 'hairdryer', 'sinkwater', 'event', 'hairdryer', 'sinkwater', 'event', 'hairdryer', 'sinkwater', 'event', 'hairdryer', 'sinkwater',]
for i in range(10):
    time_start =  time.time()
    y,sr=load('/home/pi/torch_python/' + str(name_arr[i]) + '.wav', sr=None)
    print(sr)
    mfccs = mfcc(y, sr, n_mfcc=40)
    mfccs_mean = np.mean(mfccs.T,axis=0)
    mfccs_mean = np.reshape(mfccs_mean, (1,40,1))
    #print(mfccs_mean)
    yhat = predict(mfccs_mean, model)
    print(time.time()-time_start)
    print('Predicted: '+str(yhat))

time of inference on RPI 4B4

bug fixing

2021-09-08 - for preparing embeddings and for inference need to use good quality rate transform algorithms like sink_best from resample. put it like default for last version of procedures

def load(
    path,
    sr=22050,
    mono=True,
    offset=0.0,
    duration=None,
    dtype=np.float32,
    res_type="sinc_best",
):