# ############################################################################ # Model: E2E ASR with attention-based ASR # Encoder: CRDNN model # Decoder: GRU + beamsearch + Transformer # Tokens: BPE with unigram # losses: CTC+ NLL # Training: Librispeech 960h # Authors: Ju-Chieh Chou, Mirco Ravanelli, Abdel Heba, Peter Plantinga, Samuele Cornell # Sung-Lin Yeh, Titouan Parcollet 2020 # ############################################################################ # Feature parameters sample_rate: 16000 n_fft: 400 n_mels: 40 # Model parameters activation: !name:torch.nn.LeakyReLU dropout: 0.15 cnn_blocks: 2 cnn_channels: (64, 128) inter_layer_pooling_size: (2, 2) cnn_kernelsize: (3, 3) time_pooling_size: 4 rnn_class: !name:speechbrain.nnet.RNN.LSTM rnn_layers: 4 rnn_neurons: 1024 rnn_bidirectional: True dnn_blocks: 1 dnn_neurons: 1024 emb_size: 1024 dec_neurons: 1024 output_neurons: 5000 # Number of tokens (same as LM) blank_index: 0 pad_index: -1 bos_index: 1 eos_index: 2 unk_index: 0 # Decoding parameters min_decode_ratio: 0.0 max_decode_ratio: 1.0 beam_size: 40 eos_threshold: 1.5 using_max_attn_shift: True max_attn_shift: 300 lm_weight: 0.80 ctc_weight_decode: 0.40 ctc_window_size: 200 coverage_penalty: 1.5 temperature: 1.0 temperature_lm: 1.0 normalize: !new:speechbrain.processing.features.InputNormalization norm_type: global compute_features: !new:speechbrain.lobes.features.Fbank sample_rate: !ref n_fft: !ref n_mels: !ref enc: !new:speechbrain.lobes.models.CRDNN.CRDNN input_shape: [null, null, !ref ] activation: !ref dropout: !ref cnn_blocks: !ref cnn_channels: !ref cnn_kernelsize: !ref inter_layer_pooling_size: !ref time_pooling: True using_2d_pooling: False time_pooling_size: !ref rnn_class: !ref rnn_layers: !ref rnn_neurons: !ref rnn_bidirectional: !ref rnn_re_init: True dnn_blocks: !ref dnn_neurons: !ref use_rnnp: True emb: !new:speechbrain.nnet.embedding.Embedding num_embeddings: !ref embedding_dim: !ref dec: !new:speechbrain.nnet.RNN.AttentionalRNNDecoder enc_dim: !ref input_size: !ref rnn_type: gru attn_type: location hidden_size: !ref attn_dim: 1024 num_layers: 1 scaling: 1.0 channels: 10 kernel_size: 100 re_init: True dropout: !ref ctc_lin: !new:speechbrain.nnet.linear.Linear input_size: !ref n_neurons: !ref seq_lin: !new:speechbrain.nnet.linear.Linear input_size: !ref n_neurons: !ref log_softmax: !new:speechbrain.nnet.activations.Softmax apply_log: True # This is the Transformer LM that is used according to the Huggingface repository # Visit the HuggingFace model corresponding to the pretrained_lm_tokenizer_path # For more details about the model! # NB: It has to match the pre-trained TransformerLM!! lm: !new:speechbrain.lobes.models.transformer.TransformerLM.TransformerLM vocab: 5000 d_model: 768 nhead: 12 num_encoder_layers: 12 num_decoder_layers: 0 d_ffn: 3072 dropout: 0.0 activation: !name:torch.nn.GELU normalize_before: False tokenizer: !new:sentencepiece.SentencePieceProcessor asr_model: !new:torch.nn.ModuleList - [!ref , !ref , !ref , !ref , !ref ] lm_model: !new:speechbrain.lobes.models.transformer.TransformerLM.TransformerLM vocab: 5000 d_model: 768 nhead: 12 num_encoder_layers: 12 num_decoder_layers: 0 d_ffn: 3072 dropout: 0.0 activation: !name:torch.nn.GELU normalize_before: False beam_searcher: !new:speechbrain.decoders.S2SRNNBeamSearchTransformerLM embedding: !ref decoder: !ref linear: !ref ctc_linear: !ref language_model: !ref bos_index: !ref eos_index: !ref blank_index: !ref min_decode_ratio: !ref max_decode_ratio: !ref beam_size: !ref eos_threshold: !ref using_max_attn_shift: !ref max_attn_shift: !ref coverage_penalty: !ref lm_weight: !ref ctc_weight: !ref ctc_window_size: !ref temperature: !ref temperature_lm: !ref modules: compute_features: !ref asr_enc: !ref asr_dec: !ref ctc_lin: !ref seq_lin: !ref normalize: !ref lm_model: !ref beam_searcher: !ref # The pretrainer allows a mapping between pretrained files and instances that # are declared in the yaml. pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer loadables: asr: !ref lm: !ref tokenizer: !ref