--- language: - ms --- # Unsupervised malay speakers from youtube videos 10492 unique speakers with at least 75 hours of voice activities. Steps to reproduce at https://github.com/huseinzol05/malaya-speech/blob/master/data/youtube/process-youtube.ipynb ## how-to 1. Download and extract [processed-youtube.tar.gz](processed-youtube.tar.gz), each processed videos saved as pickle, `{video_name}.pkl`. 2. Each pickle file got, ```python [{'wav_data': '/home/husein/ssd2/processed-youtube-v2/"Abam_peluk_saya_lama_atas_pentas_akhir_MLM"-_Ali_Puteh_menangis_imbau_saat_manis_dengan_arwah_abang-_MdgGr7VD7w/0.mp3', 'timestamp': datetime.datetime(2023, 3, 2, 18, 45, 45, 778042), 'asr_model': ('kenapa tak mahu bangun kau abang', [0.5325799628135358], [309, 9, 399, 633, 108, 252]), 'classification_model': (array([ 3.02432757e-03, -3.64390127e-02, 2.93319039e-02, -2.84599233e-02, -5.04244901e-02, 6.03185333e-02, 7.04260264e-03, 7.36895157e-03, 2.41034012e-02, -3.31214964e-02, -1.61228217e-02, -1.92081463e-02, -1.77928973e-02, 1.05488757e-02, 5.11314301e-03, 2.08497643e-02, 2.80407351e-02, -1.34683009e-02, 1.10213496e-02, -5.76948654e-03, 2.11171638e-02, -3.10498872e-03, 1.60899870e-02, -2.22061612e-02, -3.09270490e-02, 1.03673469e-02, 2.29822248e-02, 5.44358939e-02, -9.44061391e-03, 3.24469656e-02, -1.40673192e-02, 6.55731931e-03, 1.94134321e-02, 2.31755860e-02, -8.62774719e-03, -3.72681394e-03, -3.17485556e-02, -1.12474747e-02, 1.65595114e-02, 2.31244415e-02, 3.28784771e-02, 8.52510054e-03, -6.41896739e-04, 3.13562714e-03, -3.15982029e-02, 1.72785181e-03, 1.58039071e-02, -9.93900001e-03, 2.03248486e-02, -2.98949536e-02, 3.53759155e-02, 3.06809470e-02, -3.68881435e-03, -3.98267582e-02, -2.07101982e-02, 2.51877047e-02, -2.51530181e-03, 1.06034977e-02, 1.24978041e-02, 2.35916697e-03, 1.31300613e-02, -1.62451845e-02, -2.09861826e-02, 3.17490734e-02, -1.18532358e-02, 4.25735563e-02, 4.17908467e-02, 1.21251179e-03, -3.85571155e-03, -9.50544327e-03, -7.37808086e-03, 2.63940021e-02, 1.09219365e-02, 3.05683501e-02, -4.08848785e-02, -1.71920974e-02, -1.46033484e-02, -3.29053291e-05, 3.84788848e-02, -7.86552951e-03, 1.01251132e-03, 2.72140447e-02, 2.52339337e-02, 3.39004360e-02, -1.38184745e-02, 2.60320995e-02, -1.01425601e-02, -1.16012329e-02, 4.30319924e-03, -1.01203052e-02, -4.66396799e-03, -2.64480542e-02, 3.44322808e-02, -4.64622118e-03, 1.06053520e-02, 1.37923108e-02, -2.05409434e-03, -1.19995829e-02, 2.10450366e-02, -2.87155900e-03, -1.39515549e-02, -1.51185887e-02, 2.29053162e-02, -1.78178120e-02, 1.95855577e-03, 2.37271357e-02, 2.80657201e-03, -6.08753460e-03, -2.01220363e-02, 3.22612897e-02, 1.82474777e-02, 5.31493872e-02, -7.08705634e-02, 2.76431069e-03, 1.03597697e-02, -3.53837833e-02, 1.38167264e-02, -5.91275143e-03, 1.84398554e-02, 6.05177172e-02, 1.14565976e-02, 1.56977493e-02, -1.82731878e-02, -4.58574407e-02, -1.08330613e-02, -1.16500622e-02, -1.19803764e-04, 6.48374185e-02, -1.21538760e-03, -5.41793741e-02, 1.38867721e-02, 3.52845751e-02, -2.08288375e-02, 1.03750750e-02, -2.17110049e-02, 2.29265504e-02, -1.21381739e-02, -1.47071329e-03, -4.36875001e-02, -2.25690063e-02, -4.16939743e-02, -8.39853752e-03, -2.06098761e-02, 2.30504461e-02, 3.48615423e-02, -4.18495797e-02, -2.41985917e-03, -3.18994140e-03, 1.22078639e-02, -9.50168632e-03, -1.97298196e-03, 1.30731370e-02, 2.07234323e-02, 1.08521534e-02, 2.30542179e-02, -2.54045837e-02, 1.45645533e-02, -1.08493539e-02, -1.30415503e-02, 3.29123251e-02, 3.46204527e-02, 2.58748885e-04, -1.28235819e-03, -1.32823242e-02, 5.47284493e-03, -2.62062326e-02, 2.31803600e-02, -2.04505119e-02, 2.32407395e-02, 2.12946888e-02, -1.28869051e-02, -6.81399694e-03, 5.68802692e-02, 4.31004271e-04, 1.67261921e-02, 2.93559525e-02, 1.32581135e-02, -9.03073605e-03, -9.38207190e-03, 1.74718127e-02, 1.72506981e-02, 5.02267219e-02, -1.32851647e-02, 5.07321544e-02, -1.87530685e-02, 4.18599546e-02, 1.50075918e-02, -2.61102356e-02, -1.59594957e-02, 1.36823149e-03, -9.64679196e-03, 1.71130225e-02], dtype=float32), 'speaker 0')}] ``` - all mp3 files postprocessing using https://malaya-speech.readthedocs.io/en/latest/load-noise-reduction.html and https://malaya-speech.readthedocs.io/en/latest/load-speech-enhancement.html - `wav_data` is directory of the audio, prune the path to proper extracted directory. - `asr_model` is predicted using the best model that we have, `conformer-medium`, returned `(text, probability, subwords)`, https://malaya-speech.readthedocs.io/en/latest/load-stt-transducer-model-pt.html - `classification_model` is predicted using NEMO TITANET Large speaker verification model, https://catalog.ngc.nvidia.com/orgs/nvidia/teams/nemo/models/titanet_large, with streaming speaker similarity, https://malaya-speech.readthedocs.io/en/latest/huggingface-repository.html 3. Group by similar speakers using pagerank method (scipy.sparse.linalg.gmres), - 90% similar, from 10492 unique speakers become 6085 unique speakers, https://github.com/huseinzol05/malaya-speech/blob/master/data/youtube/mapping-youtube-speakers-90.json - 85% similar, from 10492 unique speakers become 4312 unique speakers, https://github.com/huseinzol05/malaya-speech/blob/master/data/youtube/mapping-youtube-speakers-85.json - 80% similar, from 10492 unique speakers become 2912 unique speakers, https://github.com/huseinzol05/malaya-speech/blob/master/data/youtube/mapping-youtube-speakers-80.json Speaker name defined as, ```python import os import pickle pkl = 'filename.pkl' with open(pkl, 'rb') as fopen: data = pickle.load(fopen) filename = os.path.split(pkl)[1].replace('.pkl', '') for result in data: speaker_name = f'{filename}-{speaker}' actual_speaker = unique_speakers[speaker_name] ``` Check example at https://github.com/huseinzol05/malaya-speech/blob/master/data/youtube/calculate-lengths-80.ipynb