Add asr.ckpt

.gitattributes

@@ -34,3 +34,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
 lm.ckpt filter=lfs diff=lfs merge=lfs -text
+asr.ckpt filter=lfs diff=lfs merge=lfs -text

asr.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:59a397ccda6a54e04e47695313779f987fd0bc1487f5764d7b190915d9d94c44
+size 302867350

hyperparams.yaml

@@ -1,13 +1,19 @@
+transducer_beam_search: true
+
 # Decoding parameters
 # Be sure that the bos and eos index match with the BPEs ones
+# Decoding parameters
 blank_index: 0
 bos_index: 0
 eos_index: 0
-lm_weight: 0.1
-beam_size: 4
+pad_index: 0
+beam_size: 20
 nbest: 1
-state_beam: 2.3
-expand_beam: 2.3
+# by default {state,expand}_beam = 2.3 as mention in paper
+# https://arxiv.org/abs/1904.02619
+state_beam: 2.0
+expand_beam: 2.0
+lm_weight: 0.1
 
 sample_rate: 16000
 n_fft: 512
@@ -25,7 +31,7 @@ activation: !name:torch.nn.GELU
 output_neurons: 5000
 dec_dim: 512
 
-normalize: !new:speechbrain.processing.features.InputNormalization
+normalizer: !new:speechbrain.processing.features.InputNormalization
    norm_type: global
 
 compute_features: !new:speechbrain.lobes.features.Fbank
@@ -43,7 +49,7 @@ CNN: !new:speechbrain.lobes.models.convolution.ConvolutionFrontEnd
    strides: (2, 2)
    residuals: (False, False)
 
-Transformer: !new:speechbrain.lobes.models.transformer.TransformerASR.TransformerASR # yamllint disable-line rule:line-length
+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>
@@ -59,7 +65,19 @@ Transformer: !new:speechbrain.lobes.models.transformer.TransformerASR.Transforme
 
 # 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>
+   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>
@@ -87,19 +105,15 @@ transducer_lin: !new:speechbrain.nnet.linear.Linear
    bias: False
 
 asr_model: !new:torch.nn.ModuleList
-   - [!ref <enc>, !ref <emb>, !ref <dec>, !ref <transducer_lin>]
+   - [!ref <CNN>, !ref <enc>, !ref <emb>, !ref <dec>, !ref <proj_enc>, !ref <proj_dec>, !ref <proj_ctc>, !ref <transducer_lin>]
 
-Beamsearcher: !new:speechbrain.decoders.transducer.TransducerBeamSearcher
-   decode_network_lst: [!ref <emb>, !ref <dec>, !ref <proj_dec>]
+decoder: !new:speechbrain.decoders.transducer.TransducerBeamSearcher
+   decode_network_lst: !new:torch.nn.ModuleList 
+      - [!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>
+   classifier_network: !new:torch.nn.ModuleList 
+      -  [!ref <transducer_lin>]
+   blank_id: !ref <blank_index>   
 
 lm_model: !new:speechbrain.lobes.models.RNNLM.RNNLM
    output_neurons: !ref <output_neurons>
@@ -116,18 +130,25 @@ tokenizer: !new:sentencepiece.SentencePieceProcessor
 
 # We compose the inference (encoder) pipeline.
 encoder: !new:speechbrain.nnet.containers.LengthsCapableSequential
-    input_shape: [null, null, !ref <n_mels>]
+    input_shape: [null, null]
     compute_features: !ref <compute_features>
     normalize: !ref <normalizer>
-    model: !ref <enc>
+    cnn: !ref <CNN>
+    transformer_encoder: !ref <enc>
+    proj_enc: !ref <proj_enc>
 
 modules:
-    normalizer: !ref <normalizer>
-    encoder: !ref <encoder>
-    decoder: !ref <decoder>
+   compute_features: !ref <compute_features>
+   normalizer: !ref <normalizer>
+   pre_transformer: !ref <CNN>
+   transformer: !ref <transformer>
+   asr_model: !ref <asr_model>
+   lm_model: !ref <lm_model>
+   encoder: !ref <encoder>
+   decoder: !ref <decoder>
 
 pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
     loadables:
-        normalizer: !ref <normalizer>
         asr: !ref <asr_model>
+        lm: !ref <lm_model>
         tokenizer: !ref <tokenizer>

normalizer.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:61e5e95123865b4283aad2a880af44f8f574f3f170019e26b9ed2be45b00bfd7
+size 2218