Updated each time the validation loss is lower than the best one. Name of the checkpoint file to save the current best model version. Note that Xtractors, even trained with torch.nn.DataParallel will be saved in single GPU mode. Name of the checkoint file used to save the model after each iteration Name of a previous checkpoint file to start from tmp_model_name YAML file to describe the model architecture (can be None for default architecture) model_name Learning_rate, default is 0.01 model_yaml Number of epochs to run, default is 100 lr You can preview/listen & extract the Wave samples contained in each file format on your PC, and view their Audio Properties without having to load them on to your keyboard first. The YAML file used to describre the SideSet (training and validation) epochs WAVE Xtractor is a handy audio tool with the ability to extract RAW (.Wav) data from popular File formats from within the music industry.
The number of output classes (speakers) dataset_yaml multi_gpu = "true", clipping = False, num_thread = args. The given parameter is a percentage of chuncks that are modified. temp_augmĪpply temporal augmentation (a temporal band chosen randomlly is spec_augĪpply spectral augmentation (a band of frequency coefficient chosen randomlly is Some parameters refer to transformation appliedįor the case of spectral and temporal augmentation. String that gives the sequence of transformations to apply augmentationĭata augmentation can be applied on-the-fly, the chosen augmentation processes
Section that describes the transformations applied to the audio chuncks pipeline Overlap in percentagre between two possible successive chunks of audio data transformation 1 means selecting all possible segments overlap Maximum number of chunks to select from every speech segment. Train : duration : 4 chunk_per_segment : 1 overlap : 0.0 transformation : pipeline : MFCC,CMVN,FrequencyMask(12-30),TemporalMask(70) augmentation : spec_aug : 0.5 temp_aug : 0.5 durationĭuration of the speech chuncks given in seconds chunk_per_segment