Upload 4 files

hyperparams.yaml

@@ -0,0 +1,189 @@
+# ############################################################################
+# Model: E2E ASR with transformer and transducer
+# Encoder: Conformer
+# Decoder: LSTM + beamsearch + RNNLM
+# Tokens: BPE with unigram
+# losses: Transducer + CTC (optional) + CE (optional)
+# Training: GigaSpeech
+# Authors:  Titouan Parcollet 2024
+# ############################################################################
+
+# Feature parameters
+sample_rate: 16000
+n_fft: 512
+n_mels: 80
+win_length: 32
+
+# BPE parameters
+token_type: unigram  # ["unigram", "bpe", "char"]
+character_coverage: 1.0
+
+####################### Model Parameters #######################################
+
+# Transformer
+d_model: 768
+joint_dim: 512
+nhead: 8
+num_encoder_layers: 12
+num_decoder_layers: 0
+d_ffn: 2048
+transformer_dropout: 0.1
+activation: !name:torch.nn.GELU
+output_neurons: 1024
+dec_dim: 512
+dec_emb_dropout: 0.2
+dec_dropout: 0.1
+
+# Decoding parameters
+blank_index: 0
+bos_index: 1
+eos_index: 2
+pad_index: 0
+beam_size: 10
+nbest: 1
+# by default {state,expand}_beam = 2.3 as mention in paper
+# https://arxiv.org/abs/1904.02619
+state_beam: 2.3
+expand_beam: 2.3
+
+normalize: !new:speechbrain.processing.features.InputNormalization
+   norm_type: global
+   update_until_epoch: 4
+
+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>
+
+############################## Models ##########################################
+
+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>
+   dropout: !ref <transformer_dropout>
+   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>
+
+# For MTL CTC over the encoder
+proj_ctc: !new:speechbrain.nnet.linear.Linear
+   input_size: !ref <joint_dim>
+   n_neurons: !ref <output_neurons>
+
+# Define some projection layers to make sure that enc and dec
+# output dim are the same before joining
+proj_enc: !new:speechbrain.nnet.linear.Linear
+   input_size: !ref <d_model>
+   n_neurons: !ref <joint_dim>
+   bias: False
+
+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
+
+# for MTL
+# update model if any HEAD module is added
+modules:
+   CNN: !ref <CNN>
+   enc: !ref <enc>
+   emb: !ref <emb>
+   dec: !ref <dec>
+   Tjoint: !ref <Tjoint>
+   transducer_lin: !ref <transducer_lin>
+   normalize: !ref <normalize>
+   proj_ctc: !ref <proj_ctc>
+   proj_dec: !ref <proj_dec>
+   proj_enc: !ref <proj_enc>
+
+
+# update model if any HEAD module is added
+model: !new:torch.nn.ModuleList
+   - [!ref <CNN>, !ref <enc>, !ref <emb>, !ref <dec>, !ref <proj_enc>, !ref <proj_dec>, !ref <proj_ctc>, !ref <transducer_lin>]
+
+############################## Decoding & optimiser ############################
+
+Greedysearcher: !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: 1
+   nbest: 1
+
+#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>
+#   state_beam: !ref <state_beam>
+#   expand_beam: !ref <expand_beam>
+
+tokenizer: !new:sentencepiece.SentencePieceProcessor
+
+pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
+   loadables:
+      model: !ref <model>
+      normalizer: !ref <normalize>
+      tokenizer: !ref <tokenizer>
+
+make_tokenizer_streaming_context: !name:speechbrain.tokenizers.SentencePiece.SentencePieceDecoderStreamingContext
+tokenizer_decode_streaming: !name:speechbrain.tokenizers.SentencePiece.spm_decode_preserve_leading_space
+
+make_decoder_streaming_context: !name:speechbrain.decoders.transducer.TransducerGreedySearcherStreamingContext # default constructor
+decoding_function: !name:speechbrain.decoders.transducer.TransducerBeamSearcher.transducer_greedy_decode_streaming
+   - !ref <Greedysearcher>  # self
+
+fea_streaming_extractor: !new:speechbrain.lobes.features.StreamingFeatureWrapper
+   module: !new:speechbrain.nnet.containers.LengthsCapableSequential
+      - !ref <compute_features>
+      - !ref <normalize>
+      - !ref <CNN>
+   # don't consider normalization as part of the input filter chain.
+   # normalization will operate at chunk level, which mismatches training
+   # somewhat, but does not appear to result in noticeable degradation.
+   properties: !apply:speechbrain.utils.filter_analysis.stack_filter_properties
+      - [!ref <compute_features>, !ref <CNN>]

model.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cbabf39e84aaf378006ced3db8d3d0128f38fb5e77aed2ff663a6ccfc1603edb
+size 564108886

normalizer.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9635569f61529346ab582d03890dbcd426b840db410102cf7f69ca74418c6f48
+size 2218

tokenizer.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:210d07217677a925cde1900017fde1741d6b746862b249e5f93e2596e5571c3c
+size 253722