{{"audio_filepath": "path/to/audio.wav", "duration": 3.45, "text": "sometext"}…{"audio_filepath": "path/to/audio.wav", "duration": 3.45, "text": "sometext"}}

def convertToWav(self, ext):
        if not os.path.exists(self.datadir + '/dataset'):
            tar = tarfile.open(self.an4Path);
            tar.extractall(path=self.datadir);
        sphList = glob.glob(self.datadir + '/dataset/**/*' + ext, recursive=True);
        for sph in sphList:
            wav = sph[:-4] + '.wav';
            cmd = ["sox", sph, wav];
            subprocess.run(cmd);
            print('renamed ' + ext + ' to ' + wav);

def buildManifest(self, transcript_path, manifest_path, wav_path):
        with open(transcript_paths, 'r') as fin:
            with open(manifest_path, 'w') as fout:
                for line in fin:
                    transcript = line[: line.find('(')-1].lower();
                    transcript = transcript.replace('<s>', '').replace('</s>', '');
                    transcript = transcript.strip();
                    file_id = line[line.find('(')+1 : -2];
                    audio_path = os.path.join(self.datadir, wav_paths, file_id[file_id.find('-')+1 : file_id.rfind('-')], file_id +'.wav');
                    duration = librosa.core.get_duration(filename=audio_path);
                    metadata = {
                        "audio_filepath": audio_path,
                        "duration": duration,
                        "text": transcript
                    }
                    print(metadata);
                    json.dump(metadata, fout);
                    fout.write('\n');

config.yaml:
name: &name "Jasper10x5"
model:
  sample_rate: &sample_rate 16000
  labels: &labels [" ", "a", "б", "в", "г", "д", "е", "ё", "ж", "з", "и", "й", "к", "л",
                   "м", "н", "о", "п", "р", "с", "т", "у", "ф", "х", "ц", "ч", "ш", "щ", "ъ", "ь", "э", "ю", "я", "'"]
preprocessor:
    _target_: nemo.collections.asr.modules.AudioToMelSpectrogramPreprocessor
    normalize: "per_feature"
    sample_rate: *sample_rate
    features: &n_mels 64
    n_fft: 512
    frame_splicing: 1
    dither: 0.00001
    stft_conv: false

import nemo;
class NemoASR:
    def __init__(self, dataDir):
        self.datadir = dataDir;
        self.CONF_PATH = './config.yaml';
        yaml = YAML(typ="safe");
        with open(self.CONF_PATH) as f:
            self.CONFIG = yaml.load(f);
    def train(self, transcriptionPATH, manifestPATH, wavPATH, testTranscriptionPATH, testManifestPATH, testWavPATH):
        print("begin train");
        train_transcripts = self.datadir + transcriptionPATH;
        train_manifest = self.datadir + manifestPATH;
        if not os.path.isfile(train_manifest):
            self.buildManifest(train_transcripts, train_manifest, wavPATH);
        test_transcripts = self.datadir + testTranscriptionPATH;
        test_manifest = self.datadir + testManifestPATH;
        if not os.path.isfile(test_manifest):
            self.buildManifest(test_transcripts, test_manifest, testWavPATH);
        # params from ./config.yaml
        self.CONFIG['model']['train_ds']['manifest_filepath'] = train_manifest;
        self.CONFIG['model']['validation_ds']['manifest_filepath'] = test_manifest;
        trainer = pl.Trainer(max_epochs=500, gpus=1);
        self.model = nemo_asr.models.EncDecCTCModel(cfg=DictConfig(self.CONFIG['model']), trainer=trainer);
        trainer.fit(self.model);
        print("end train");
#-------------------------------------------------------------
nemoASR = NemoASR('.');
if (nemoASR.checkExistsDataSet()):
    print('dataset loaded');
    nemoASR.train('./dataset/etc/train.transcription',  './dataset/train_manifest.json','./dataset/wav/an4_clstk', './dataset/etc/test.transcription', './dataset/test_manifest.json', './dataset/wav/an4test_clstk');
    nemoASR.model.save_to('./model.sbc');

files = ['./an4/wav/an4_clstk/mgah/cen2-mgah-b.wav'];
    for fname, transcription in zip(files, nemoASR.model.transcribe(paths2audio_files=files)):
        print(f"Audio in {fname} was recognized as: {transcription}");