hyperparams.yaml

# ############################################################################
# Model: ECAPA big for Speaker verification
# ############################################################################

# Feature parameters
n_mels: 80

# Pretrain folder (HuggingFace)
pretrained_path: poonehmousavi/discrete_wavlm_spk_rec_ecapatdn
# Output parameters
save_folder: tmp

### Configuration for  discrete SSL model
# ssl_model_type: hubert, wavlm, wav2vec2
# ssl_hub: facebook/hubert-large-ll60k, microsoft/wavlm-large,  facebook/wav2vec2-large
ssl_model_type: wavlm # hubert, wavml or wav2vec2
ssl_hub: microsoft/wavlm-large
ssl_folder: !ref <save_folder>/ssl_checkpoint
kmeans_repo_id: speechbrain/SSL_Quantization
kmeans_cache_dir: !ref <save_folder>/kmeans_checkpoint
kmeans_dataset: LibriSpeech-100-360-500
freeze_ssl: True
freeze_feature_extractor: True
num_clusters: 1000

### Config for Tokenizer
# Layer number should be among the supported layers for discrete SSL models(kmenas  model should be available for that layer)
# ssl_layer_num: [3, 7, 12, 23]
# deduplicate: [False, False, False, False]
# bpe_tokenizer_path: [null , null,  null, null]
ssl_layer_num: [1, 3, 7, 12, 18, 23]
ssl_layer_num_selected: [1, 3, 7, 12, 18, 23]
num_codebooks: 6
deduplicate: [False, False, False, False, False, False]
bpe_tokenizer_path: [null, null, null, null, null, null]
sample_rate: 16000

# Feature parameters
encoder_dim: 1024
# Modules
tokenizer_config:
    SSL_layers: !ref <ssl_layer_num>
    deduplicates: !ref <deduplicate>
    bpe_tokenizers: !ref <bpe_tokenizer_path>

discrete_embedding_layer: !new:custom_interface.Discrete_EmbeddingLayer
    num_codebooks: !ref <num_codebooks>
    vocab_size: !ref <num_clusters>
    emb_dim: !ref <encoder_dim>
    available_layers: !ref <ssl_layer_num>
    layers: !ref <ssl_layer_num_selected>

attention_mlp: !new:custom_interface.AttentionMLP
    input_dim: !ref <encoder_dim>
    hidden_dim: !ref <encoder_dim>

embedding_model: !new:speechbrain.lobes.models.ECAPA_TDNN.ECAPA_TDNN
    input_size: !ref <encoder_dim>
    channels: [1024, 1024, 1024, 1024, 3072]
    kernel_sizes: [5, 3, 3, 3, 1]
    dilations: [1, 2, 3, 4, 1]
    groups: [1, 1, 1, 1, 1]
    attention_channels: 128
    lin_neurons: 192

modules:
    embedding_model: !ref <embedding_model>
    attention_mlp: !ref <attention_mlp>
    discrete_embedding_layer: !ref <discrete_embedding_layer>

        
pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
    loadables:
        embedding_model: !ref <embedding_model>
        attention_mlp: !ref <attention_mlp>
        discrete_embedding_layer: !ref <discrete_embedding_layer>

    paths:
        embedding_model: !ref <pretrained_path>/embedding_model.ckpt
        attention_mlp: !ref <pretrained_path>/attention_mlp.ckpt
        discrete_embedding_layer: !ref <pretrained_path>/discrete_embedding_layer.ckpt