Create hyperparams.yaml

hyperparams.yaml

@@ -0,0 +1,133 @@
+# Decoding parameters
+# Be sure that the bos and eos index match with the BPEs ones
+blank_index: 0
+bos_index: 0
+eos_index: 0
+lm_weight: 0.1
+beam_size: 4
+nbest: 1
+state_beam: 2.3
+expand_beam: 2.3
+
+sample_rate: 16000
+n_fft: 512
+n_mels: 80
+win_length: 32
+
+# Transformer
+d_model: 256
+joint_dim: 640
+nhead: 4
+num_encoder_layers: 12
+num_decoder_layers: 0
+d_ffn: 2048
+activation: !name:torch.nn.GELU
+output_neurons: 5000
+dec_dim: 512
+
+normalize: !new:speechbrain.processing.features.InputNormalization
+   norm_type: global
+
+compute_features: !new:speechbrain.lobes.features.Fbank
+   sample_rate: !ref <sample_rate>
+   n_fft: !ref <n_fft>
+   n_mels: !ref <n_mels>
+   win_length: !ref <win_length>
+    
+CNN: !new:speechbrain.lobes.models.convolution.ConvolutionFrontEnd
+   input_shape: (8, 10, 80)
+   num_blocks: 2
+   num_layers_per_block: 1
+   out_channels: (64, 32)
+   kernel_sizes: (3, 3)
+   strides: (2, 2)
+   residuals: (False, False)
+
+Transformer: !new:speechbrain.lobes.models.transformer.TransformerASR.TransformerASR # yamllint disable-line rule:line-length
+   input_size: 640
+   tgt_vocab: !ref <output_neurons>
+   d_model: !ref <d_model>
+   nhead: !ref <nhead>
+   num_encoder_layers: !ref <num_encoder_layers>
+   num_decoder_layers: !ref <num_decoder_layers>
+   d_ffn: !ref <d_ffn>
+   activation: !ref <activation>
+   encoder_module: conformer
+   attention_type: RelPosMHAXL
+   normalize_before: True
+   causal: False
+
+# We must call an encoder wrapper so the decoder isn't run (we don't have any)
+enc: !new:speechbrain.lobes.models.transformer.TransformerASR.EncoderWrapper
+   transformer: !ref <Transformer>
+
+proj_dec: !new:speechbrain.nnet.linear.Linear
+   input_size: !ref <dec_dim>
+   n_neurons: !ref <joint_dim>
+   bias: False
+
+emb: !new:speechbrain.nnet.embedding.Embedding
+   num_embeddings: !ref <output_neurons>
+   consider_as_one_hot: True
+   blank_id: !ref <blank_index>
+
+dec: !new:speechbrain.nnet.RNN.LSTM
+   input_shape: [null, null, !ref <output_neurons> - 1]
+   hidden_size: !ref <dec_dim>
+   num_layers: 1
+   re_init: True
+
+Tjoint: !new:speechbrain.nnet.transducer.transducer_joint.Transducer_joint
+   joint: sum # joint [sum | concat]
+   nonlinearity: !ref <activation>
+
+transducer_lin: !new:speechbrain.nnet.linear.Linear
+   input_size: !ref <joint_dim>
+   n_neurons: !ref <output_neurons>
+   bias: False
+
+asr_model: !new:torch.nn.ModuleList
+   - [!ref <enc>, !ref <emb>, !ref <dec>, !ref <transducer_lin>]
+
+Beamsearcher: !new:speechbrain.decoders.transducer.TransducerBeamSearcher
+   decode_network_lst: [!ref <emb>, !ref <dec>, !ref <proj_dec>]
+   tjoint: !ref <Tjoint>
+   classifier_network: [!ref <transducer_lin>]
+   blank_id: !ref <blank_index>
+   beam_size: !ref <beam_size>
+   nbest: !ref <nbest>
+   lm_module: !ref <lm_model>
+   lm_weight: !ref <lm_weight>
+   state_beam: !ref <state_beam>
+   expand_beam: !ref <expand_beam>
+
+lm_model: !new:speechbrain.lobes.models.RNNLM.RNNLM
+   output_neurons: !ref <output_neurons>
+   embedding_dim: 256
+   activation: !name:torch.nn.LeakyReLU
+   dropout: 0.0
+   rnn_layers: 6
+   rnn_neurons: 512
+   dnn_blocks: 1
+   dnn_neurons: 256
+   return_hidden: True  # For inference
+
+tokenizer: !new:sentencepiece.SentencePieceProcessor
+
+# We compose the inference (encoder) pipeline.
+encoder: !new:speechbrain.nnet.containers.LengthsCapableSequential
+    input_shape: [null, null, !ref <n_mels>]
+    compute_features: !ref <compute_features>
+    normalize: !ref <normalizer>
+    model: !ref <enc>
+
+modules:
+    normalizer: !ref <normalizer>
+    encoder: !ref <encoder>
+    decoder: !ref <decoder>
+
+pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
+    loadables:
+        normalizer: !ref <normalizer>
+        asr: !ref <asr_model>
+        tokenizer: !ref <tokenizer>