hyperparams.yaml

# HyperPyYAML for the WavLM/FocalCodec hybrid codec.
# Checkpoint variant: scratch at 50 Hz.

codec_sample_rate: 16000
codec_frequency_hz: 50

# The trained encoder and decoder are provided by the upstream FocalCodec model.
focalcodec_source: lucadellalib/focalcodec
focalcodec_model: focalcodec
focalcodec_config: lucadellalib/focalcodec_50hz
focalcodec: !apply:torch.hub.load
    repo_or_dir: !ref <focalcodec_source>
    model: !ref <focalcodec_model>
    config: !ref <focalcodec_config>
    trust_repo: True
    pretrained: True

encoder: !new:hybridcodec.model.FocalCodecEncoder
    model: !ref <focalcodec.encoder>

quantizer: !new:hybridcodec.model.SingleLayerResidualQuantizer
    input_dim: 1024
    bottleneck_dim: 32
    codebook_size: 8192
    semantic_hidden_sizes: [1024, 512, 256]
    semantic_downscale_factors: [1, 1, 1]
    residual_hidden_sizes: [512, 256]
    residual_downscale_factors: [1, 1]
    continuous_dropout: 0.2
    use_post_norm: False

vocoder: !ref <focalcodec.decoder>

codec: !new:hybridcodec.model.HybridCodec
    encoder: !ref <encoder>
    quantizer: !ref <quantizer>
    vocoder: !ref <vocoder>

modules:
    codec: !ref <codec>

# Pretrainer uses this key as the relative filename under the source repo.
pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
    loadables:
        "weights/quantizer_scratch_50hz": !ref <quantizer>