| { |
| "_explicit_embedding_dim": true, |
| "backbone": "conformer", |
| "embedding_dim": 256, |
| "experiment_dict": { |
| "data": { |
| "audio_format": ".flac", |
| "dataset_name": "librispeech", |
| "dataset_type": "librispeech", |
| "domain_temp_alpha": 0.35, |
| "speaker_label_index": -2, |
| "test_csv": "data/speaker_datasets/splits/librispeech/eval.csv", |
| "test_dir": "", |
| "test_librispeech_dir": "", |
| "train_csv": "data/speaker_datasets/splits/librispeech/train.csv", |
| "train_dir": "", |
| "train_librispeech_dir": "", |
| "trials_file": "data/speaker_datasets/splits/librispeech/trials.tsv", |
| "within_domain_prob": 0.75 |
| }, |
| "eval": { |
| "apply_as_norm": false, |
| "as_norm_cohort_size": 100, |
| "as_norm_top_k": 100, |
| "dcf_c_fa": 1.0, |
| "dcf_c_miss": 1.0, |
| "full_sequence_inference": true, |
| "num_enroll_utts": 1, |
| "num_triplets": 1000, |
| "threshold_step": 0.001, |
| "use_as_norm": false |
| }, |
| "export": { |
| "export_tflite": true, |
| "output_dir": "exported_model", |
| "sequence_length": 180 |
| }, |
| "frontend": { |
| "apply_cmvn": true, |
| "cmvn_norm_vars": false, |
| "feature_type": "log_mel", |
| "fmax": 7600.0, |
| "fmin": 20.0, |
| "frame_size_ms": 25.0, |
| "frame_step_ms": 10.0, |
| "frame_stride": 1, |
| "hf_extractor_name": "openai/whisper-tiny", |
| "log_offset": 1e-06, |
| "n_fft": 512, |
| "n_mels": 80, |
| "n_mfcc": 80, |
| "preemph": 0.97, |
| "sample_rate": 16000, |
| "stack_left_context": 0, |
| "stack_right_context": 0 |
| }, |
| "loss": { |
| "ext_softmax_normalize_by_num_speakers": true, |
| "ext_softmax_row_softmax_weight": 0.5, |
| "ge2e_contrast_b_grad_scale": 0.0, |
| "ge2e_contrast_surrogate": false, |
| "ge2e_exact_leave_one_out": true, |
| "ge2e_init_b": -5.0, |
| "ge2e_init_w": 10.0, |
| "ge2e_learn_wb": true, |
| "ge2e_split_batch": false, |
| "loss_type": "ge2e_softmax", |
| "margin": 0.2, |
| "multi_row_window": 0, |
| "num_speakers_per_batch": 8, |
| "num_utts_per_speaker": 4, |
| "pure_extended_set": false, |
| "scale": 30.0, |
| "scoring_type": "cosine", |
| "sphereface_lambda_base": 1000.0, |
| "sphereface_lambda_gamma": 0.12, |
| "sphereface_lambda_min": 5.0, |
| "sphereface_lambda_power": 1.0, |
| "sphereface_m": 1.35, |
| "triplet_alpha": 0.1 |
| }, |
| "model": { |
| "backbone": "conformer", |
| "conformer": { |
| "conv_kernel_size": 15, |
| "dim": 256, |
| "ffn_dim": 512, |
| "layer_order": "mhsa_before_conv", |
| "normalize_output": false, |
| "num_heads": 8, |
| "num_layers": 4, |
| "subsample_after_layer": 3, |
| "subsample_factor": 1, |
| "use_conv_stem": true, |
| "use_group_norm": true, |
| "use_positional_encoding": false |
| }, |
| "dr_vector": { |
| "activation": "relu", |
| "add_cosine_to_network_output": true, |
| "append_cosine_as_feature_to_network_input": true, |
| "append_cosine_to_final_hidden": false, |
| "cosine_dims": 0, |
| "hidden_dims": [ |
| 128, |
| 64 |
| ], |
| "include_cosine_in_input": true, |
| "leaky_relu_slope": 0.2, |
| "num_cosine_terms": 1, |
| "readout": "linear", |
| "residual_weight": 0.25, |
| "symmetric_input": false, |
| "trainable_affine": false, |
| "use_layer_norm": true |
| }, |
| "ecapa_tdnn": { |
| "channels": 384, |
| "dilations": [ |
| 1, |
| 2, |
| 3, |
| 4 |
| ], |
| "kernel_sizes": [ |
| 5, |
| 3, |
| 3, |
| 3 |
| ], |
| "mfa_dim": 768, |
| "norm_after_activation": false, |
| "scale": 4, |
| "se_bottleneck_dim": 64 |
| }, |
| "embedding_dim": 256, |
| "frame_aggregation_mean": true, |
| "full_sequence_inference": true, |
| "lstm": { |
| "bidirectional": false, |
| "frame_aggregation_mean": true, |
| "hidden_size": 384, |
| "num_layers": 4, |
| "proj_size": 0, |
| "use_layer_norm": true |
| }, |
| "mamba": { |
| "bidirectional": false, |
| "d_conv": 4, |
| "d_state": 16, |
| "dim": 256, |
| "dt_rank": 16, |
| "expand_factor": 2, |
| "num_layers": 4 |
| }, |
| "max_eval_seq_len": 500, |
| "max_seq_len": 180, |
| "min_seq_len": 90, |
| "n_mfcc": 80, |
| "normalize_embedding": true, |
| "pfas": { |
| "apply_global_l2_norm_to_concat_form": true, |
| "apply_l2_norm_to_keys": true, |
| "apply_l2_norm_to_values": false, |
| "apply_softmax_per_test_key": false, |
| "atp_ffn_use_layer_norm": true, |
| "attention_dim": 128, |
| "attention_type": "shared_nonlinear", |
| "key_dim": 32, |
| "local_value_scale": 0.25, |
| "normalize_keys_and_values": false, |
| "normalize_overall_output": false, |
| "num_keys": 4, |
| "per_trial_softmax": true, |
| "scale_factor": 16.0, |
| "use_global_anchor": true, |
| "use_trainable_scale_factor": true, |
| "value_dim": 128 |
| }, |
| "pooling": { |
| "attention_dim": 128, |
| "bottleneck_dim": 256, |
| "epsilon": 0.0001, |
| "num_heads": 4, |
| "pooling_type": "atp", |
| "stats_type": "MEAN_STD", |
| "subspace_heads": true, |
| "use_global_context": true, |
| "use_nonlinear_bottleneck": true |
| }, |
| "resnet": { |
| "base_channels": 32, |
| "num_blocks": [ |
| 2, |
| 2, |
| 2, |
| 2 |
| ], |
| "out_dim": 256, |
| "use_se": true |
| }, |
| "saved_model_path": "", |
| "scoring_type": "cosine", |
| "seq_len": 180, |
| "size_variant": "base", |
| "sliding_window_step": 50, |
| "transformer": { |
| "dim": 256, |
| "mlp_dim": 512, |
| "num_encoder_layers": 4, |
| "num_heads": 8, |
| "use_conv_stem": true, |
| "use_positional_encoding": false |
| }, |
| "use_transformer": false, |
| "variable_length_training": true |
| }, |
| "train": { |
| "batch_size": 32, |
| "ge2e_grad_scale": 0.01, |
| "grad_clip_norm": 3.0, |
| "learning_rate": 0.0012, |
| "loss": { |
| "ext_softmax_normalize_by_num_speakers": true, |
| "ext_softmax_row_softmax_weight": 0.5, |
| "ge2e_contrast_b_grad_scale": 0.0, |
| "ge2e_contrast_surrogate": false, |
| "ge2e_exact_leave_one_out": true, |
| "ge2e_init_b": -5.0, |
| "ge2e_init_w": 10.0, |
| "ge2e_learn_wb": true, |
| "ge2e_split_batch": false, |
| "loss_type": "ge2e_softmax", |
| "margin": 0.2, |
| "multi_row_window": 0, |
| "num_speakers_per_batch": 8, |
| "num_utts_per_speaker": 4, |
| "pure_extended_set": false, |
| "scale": 30.0, |
| "scoring_type": "cosine", |
| "sphereface_lambda_base": 1000.0, |
| "sphereface_lambda_gamma": 0.12, |
| "sphereface_lambda_min": 5.0, |
| "sphereface_lambda_power": 1.0, |
| "sphereface_m": 1.35, |
| "triplet_alpha": 0.1 |
| }, |
| "lr_schedule_cosine": true, |
| "min_learning_rate": 1.1999999999999999e-05, |
| "num_processes": 4, |
| "num_spks_per_batch": 8, |
| "num_steps": 400, |
| "num_utts_per_spk": 4, |
| "optimizer": "adamw", |
| "save_model_frequency": 0, |
| "saved_model_dir": "", |
| "seed": 42, |
| "specaug": { |
| "apply_prob": 0.5, |
| "freq_mask_max_width": 8, |
| "freq_mask_prob": 0.5, |
| "gaussian_noise_std": 0.0, |
| "num_freq_masks": 2, |
| "num_time_masks": 2, |
| "time_mask_max_width": 12, |
| "time_mask_prob": 0.5, |
| "use_specaug": true |
| }, |
| "triplet_alpha": 0.1, |
| "warmup_steps": 40, |
| "waveform_aug": { |
| "aug_prob": 0.35, |
| "enabled": true, |
| "reverb_prob": 0.4, |
| "reverb_rt60_max": 0.2, |
| "reverb_rt60_min": 0.05, |
| "snr_max_db": 20.0, |
| "snr_min_db": 12.0, |
| "use_waveform_aug": true |
| }, |
| "weight_decay": 0.0001 |
| }, |
| "vad": { |
| "adaptive_percentile": 10.0, |
| "apply_agc": true, |
| "enabled": true, |
| "energy_db_threshold": 26.0, |
| "energy_threshold_db": -42.0, |
| "frame_size_ms": 25.0, |
| "frame_step_ms": 10.0, |
| "hangover_frames": 8, |
| "min_speech_frames": 3, |
| "mode": "energy", |
| "sample_rate": 16000, |
| "snr_margin_db": 4.5, |
| "spectral_flatness_threshold": 0.65 |
| } |
| }, |
| "feature_type": "log_mel", |
| "loss_type": "ge2e_softmax", |
| "model_type": "flaxspeaker", |
| "n_mels": 80, |
| "n_mfcc": 80, |
| "pooling_type": "atp", |
| "sample_rate": 16000, |
| "scoring_type": "cosine", |
| "seq_len": 180, |
| "size_variant": "base", |
| "transformers_version": "5.17.0" |
| } |
|
|
