ipa_ottc/inference.yaml

# Deployment wrapper for the selected L2-ARCTIC ZipA-IPA monophone OTTC
# checkpoint. Inference uses the content head; OTTC changes training loss and
# posterior geometry, not the public label-space contract.
prefix: "l2arctic_ipa_space_ottc"
save_folder: "/work/gm64/m64000/IF-MDD/exp_l2arctic_ipa_fullft_ottc_clean_20260708/monophone/wavlm_large_PhnMonoSSL_l2arctic_ipa_monoottc_fullft_from10ms/save/CKPT+060_content_PER_24.1661_MDDF1_0.9599.ckpt"
lab_enc_file: "/work/gm64/m64000/IF-MDD/data/l2arctic_zipa_ipa_ctc/label_encoder.txt"
pretrained_models_path: "/work/gm64/m64000/IF-MDD/pretrained_models"
wavlm_source: "/work/gm64/m64000/IF-MDD/pretrained_models/models--microsoft--wavlm-large/snapshots/c1423ed94bb01d80a3f5ce5bc39f6026a0f4828c"
perceived_ssl_model: "wavlm_large"
sample_rate: 16000
ENCODER_DIM: 1024
output_neurons: 127
blank_index: 0
freeze_perceived_ssl: true
freeze_perceived_feature_extractor: true
wavlm_last_conv_stride: 1
wavlm_feature_extractor_sampling_rate: !ref <sample_rate>
dnn_layers: 2
dnn_neurons: 384
activation: !name:torch.nn.LeakyReLU

perceived_ssl: !new:trainer.WavLMFrameRate.WavLMFrameRate
    source: !ref <wavlm_source>
    freeze: !ref <freeze_perceived_ssl>
    freeze_feature_extractor: !ref <freeze_perceived_feature_extractor>
    save_path: !ref <pretrained_models_path>
    output_all_hiddens: False
    last_conv_stride: !ref <wavlm_last_conv_stride>
    feature_extractor_sampling_rate: !ref <wavlm_feature_extractor_sampling_rate>
enc: !new:torch.nn.Sequential
  - !new:speechbrain.lobes.models.VanillaNN.VanillaNN
     input_shape: [null, null, !ref <ENCODER_DIM>]
     activation: !ref <activation>
     dnn_blocks: !ref <dnn_layers>
     dnn_neurons: !ref <dnn_neurons>
  - !new:torch.nn.LayerNorm
     normalized_shape: !ref <dnn_neurons>
ctc_lin: !new:speechbrain.nnet.linear.Linear
    input_size: !ref <dnn_neurons>
    n_neurons: !ref <output_neurons>
lm_weight: !new:speechbrain.nnet.linear.Linear
    input_size: !ref <dnn_neurons>
    n_neurons: 1
log_softmax: !new:speechbrain.nnet.activations.Softmax
    apply_log: True
model: !new:torch.nn.ModuleList
    - [!ref <enc>, !ref <ctc_lin>, !ref <lm_weight>]
pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
    collect_in: !ref <save_folder>/
    loadables:
        model: !ref <model>
        perceived_ssl: !ref <perceived_ssl>
encoder: !new:speechbrain.nnet.containers.LengthsCapableSequential
    perceived_ssl: !ref <perceived_ssl>
    enc: !ref <enc>
    ctc_lin: !ref <ctc_lin>
    log_softmax: !ref <log_softmax>
decoding_function: !name:speechbrain.decoders.ctc_greedy_decode
    blank_id: !ref <blank_index>
tokenizer: !new:speechbrain.dataio.encoder.TextEncoder
    load_from_file: !ref <lab_enc_file>
modules:
    encoder: !ref <encoder>