From a623908c6463436e876c4d6aab90025e71463d14 Mon Sep 17 00:00:00 2001
From: Florian Boyer <florian.boyer@airudit.com>
Date: Thu, 6 Oct 2022 14:55:09 +0000
Subject: [PATCH 01/15] add zero pad for convolution (ref: icefall)

---
 espnet2/asr_transducer/encoder/blocks/conformer.py    | 2 +-
 espnet2/asr_transducer/encoder/modules/convolution.py | 4 ++++
 2 files changed, 5 insertions(+), 1 deletion(-)

diff --git a/espnet2/asr_transducer/encoder/blocks/conformer.py b/espnet2/asr_transducer/encoder/blocks/conformer.py
index b95a426e6b5..1d687505d3b 100644
--- a/espnet2/asr_transducer/encoder/blocks/conformer.py
+++ b/espnet2/asr_transducer/encoder/blocks/conformer.py
@@ -123,7 +123,7 @@ def forward(
         residual = x
 
         x = self.norm_conv(x)
-        x, _ = self.conv_mod(x)
+        x, _ = self.conv_mod(x, mask=mask)
         x = residual + self.dropout(x)
 
         residual = x
diff --git a/espnet2/asr_transducer/encoder/modules/convolution.py b/espnet2/asr_transducer/encoder/modules/convolution.py
index 012538a7db9..0e50f33292c 100644
--- a/espnet2/asr_transducer/encoder/modules/convolution.py
+++ b/espnet2/asr_transducer/encoder/modules/convolution.py
@@ -70,6 +70,7 @@ def forward(
         self,
         x: torch.Tensor,
         cache: Optional[torch.Tensor] = None,
+        mask: Optional[torch.Tensor] = None,
         right_context: int = 0,
     ) -> Tuple[torch.Tensor, torch.Tensor]:
         """Compute convolution module.
@@ -87,6 +88,9 @@ def forward(
         x = self.pointwise_conv1(x.transpose(1, 2))
         x = torch.nn.functional.glu(x, dim=1)
 
+        if mask is not None:
+            x.masked_fill(mask.unsqueeze(1).expand_as(x), 0.0)
+
         if self.lorder > 0:
             if cache is None:
                 x = torch.nn.functional.pad(x, (self.lorder, 0), "constant", 0.0)

From c4ef9cf7ccc3093bcdb1f6b5bfadd8ec19fdc596 Mon Sep 17 00:00:00 2001
From: Florian Boyer <florian.boyer@airudit.com>
Date: Thu, 6 Oct 2022 14:55:53 +0000
Subject: [PATCH 02/15] fix aux. lm loss, reporter variable and doc

---
 .../asr_transducer/espnet_transducer_model.py | 48 ++++++++++---------
 1 file changed, 25 insertions(+), 23 deletions(-)

diff --git a/espnet2/asr_transducer/espnet_transducer_model.py b/espnet2/asr_transducer/espnet_transducer_model.py
index 3e66605dfe3..345e5a38a32 100644
--- a/espnet2/asr_transducer/espnet_transducer_model.py
+++ b/espnet2/asr_transducer/espnet_transducer_model.py
@@ -69,7 +69,7 @@ def __init__(
         auxiliary_ctc_weight: float = 0.0,
         auxiliary_ctc_dropout_rate: float = 0.0,
         auxiliary_lm_loss_weight: float = 0.0,
-        auxiliary_lm_loss_smoothing: float = 0.0,
+        auxiliary_lm_loss_smoothing: float = 0.05,
         ignore_id: int = -1,
         sym_space: str = "<space>",
         sym_blank: str = "<blank>",
@@ -82,7 +82,10 @@ def __init__(
 
         assert check_argument_types()
 
-        # The following labels ID are reserved: 0 (blank) and vocab_size - 1 (sos/eos)
+        # The following labels ID are reserved:
+        #    - 0: Blank symbol.
+        #    - 1: Unknown symbol.
+        #    - vocab_size - 1: SOS/EOS symbol.
         self.vocab_size = vocab_size
         self.ignore_id = ignore_id
         self.token_list = token_list.copy()
@@ -110,7 +113,11 @@ def __init__(
 
         if self.use_auxiliary_lm_loss:
             self.lm_lin = torch.nn.Linear(decoder.output_size, vocab_size)
-            self.lm_loss_smoothing = auxiliary_lm_loss_smoothing
+
+            eps = auxiliary_lm_loss_smoothing / (vocab_size - 1)
+
+            self.lm_loss_smooth_neg = eps
+            self.lm_loss_smooth_pos = (1 - auxiliary_lm_loss_smoothing) + eps
 
         self.transducer_weight = transducer_weight
         self.fastemit_lambda = fastemit_lambda
@@ -207,8 +214,8 @@ def forward(
         stats = dict(
             loss=loss.detach(),
             loss_transducer=loss_trans.detach(),
-            aux_ctc_loss=loss_ctc.detach() if loss_ctc > 0.0 else None,
-            aux_lm_loss=loss_lm.detach() if loss_lm > 0.0 else None,
+            loss_aux_ctc=loss_ctc.detach() if loss_ctc > 0.0 else None,
+            loss_aux_lm=loss_lm.detach() if loss_lm > 0.0 else None,
             cer_transducer=cer_trans,
             wer_transducer=wer_trans,
         )
@@ -432,7 +439,7 @@ def _calc_lm_loss(
         decoder_out: torch.Tensor,
         target: torch.Tensor,
     ) -> torch.Tensor:
-        """Compute LM loss.
+        """Compute LM loss (i.e.: Cross-entropy with smoothing).
 
         Args:
             decoder_out: Decoder output sequences. (B, U, D_dec)
@@ -442,26 +449,21 @@ def _calc_lm_loss(
             loss_lm: LM loss value.
 
         """
-        lm_loss_in = self.lm_lin(decoder_out[:, :-1, :]).view(-1, self.vocab_size)
-        lm_target = target.view(-1).type(torch.int64)
+        batch_size = decoder_out.size(0)
 
-        with torch.no_grad():
-            true_dist = lm_loss_in.clone()
-            true_dist.fill_(self.lm_loss_smoothing / (self.vocab_size - 1))
+        logp = torch.log_softmax(
+            self.lm_lin(decoder_out[:, :-1, :]).view(-1, self.vocab_size),
+            dim=1,
+        )
+        target = target.view(-1).type(torch.int64)
+        ignore = (target == 0).unsqueeze(1)
 
-            # Ignore blank ID (0)
-            ignore = lm_target == 0
-            lm_target = lm_target.masked_fill(ignore, 0)
+        with torch.no_grad():
+            true_dist = logp.clone().fill_(self.lm_loss_smooth_neg)
 
-            true_dist.scatter_(1, lm_target.unsqueeze(1), (1 - self.lm_loss_smoothing))
+            true_dist.scatter_(1, target.unsqueeze(1), self.lm_loss_smooth_pos)
 
-        loss_lm = torch.nn.functional.kl_div(
-            torch.log_softmax(lm_loss_in, dim=1),
-            true_dist,
-            reduction="none",
-        )
-        loss_lm = loss_lm.masked_fill(ignore.unsqueeze(1), 0).sum() / decoder_out.size(
-            0
-        )
+        loss_lm = torch.nn.functional.kl_div(logp, true_dist, reduction="none")
+        loss_lm = loss_lm.masked_fill(ignore, 0).sum() / batch_size
 
         return loss_lm

From 4d045dfecc64f9ac14b4975838d8b11f4844d68a Mon Sep 17 00:00:00 2001
From: Florian Boyer <florian.boyer@airudit.com>
Date: Thu, 6 Oct 2022 14:56:46 +0000
Subject: [PATCH 03/15] parser -> group for add_argument

---
 espnet2/tasks/asr_transducer.py | 20 ++++++++++----------
 1 file changed, 10 insertions(+), 10 deletions(-)

diff --git a/espnet2/tasks/asr_transducer.py b/espnet2/tasks/asr_transducer.py
index aa3cbf54e88..8c7786277d2 100644
--- a/espnet2/tasks/asr_transducer.py
+++ b/espnet2/tasks/asr_transducer.py
@@ -135,7 +135,7 @@ def add_task_arguments(cls, parser: argparse.ArgumentParser):
             default={},
             help="The keyword arguments for the joint network class.",
         )
-        group = parser.add_argument_group(description="Preprocess related.")
+        group = group.add_argument_group(description="Preprocess related.")
         group.add_argument(
             "--use_preprocessor",
             type=str2bool,
@@ -155,56 +155,56 @@ def add_task_arguments(cls, parser: argparse.ArgumentParser):
             default=None,
             help="The path of the sentencepiece model.",
         )
-        parser.add_argument(
+        group.add_argument(
             "--non_linguistic_symbols",
             type=str_or_none,
             help="The 'non_linguistic_symbols' file path.",
         )
-        parser.add_argument(
+        group.add_argument(
             "--cleaner",
             type=str_or_none,
             choices=[None, "tacotron", "jaconv", "vietnamese"],
             default=None,
             help="Text cleaner to use.",
         )
-        parser.add_argument(
+        group.add_argument(
             "--g2p",
             type=str_or_none,
             choices=g2p_choices,
             default=None,
             help="g2p method to use if --token_type=phn.",
         )
-        parser.add_argument(
+        group.add_argument(
             "--speech_volume_normalize",
             type=float_or_none,
             default=None,
             help="Normalization value for maximum amplitude scaling.",
         )
-        parser.add_argument(
+        group.add_argument(
             "--rir_scp",
             type=str_or_none,
             default=None,
             help="The RIR SCP file path.",
         )
-        parser.add_argument(
+        group.add_argument(
             "--rir_apply_prob",
             type=float,
             default=1.0,
             help="The probability of the applied RIR convolution.",
         )
-        parser.add_argument(
+        group.add_argument(
             "--noise_scp",
             type=str_or_none,
             default=None,
             help="The path of noise SCP file.",
         )
-        parser.add_argument(
+        group.add_argument(
             "--noise_apply_prob",
             type=float,
             default=1.0,
             help="The probability of the applied noise addition.",
         )
-        parser.add_argument(
+        group.add_argument(
             "--noise_db_range",
             type=str,
             default="13_15",

From df6a0ded3d46b25f1b9e6b62b841a5efa73858d5 Mon Sep 17 00:00:00 2001
From: Florian Boyer <florian.boyer@airudit.com>
Date: Thu, 6 Oct 2022 14:57:17 +0000
Subject: [PATCH 04/15] add libri-100 asr transducer task

---
 egs2/librispeech_100/asr_transducer1/asr.sh   |  1 +
 egs2/librispeech_100/asr_transducer1/cmd.sh   |  1 +
 .../asr_transducer1/conf/decode.yaml          | 14 ++++
 .../conf/decode_streaming.yaml                |  9 ++
 .../conf/train_conformer-rnn_transducer.yaml  | 78 ++++++++++++++++++
 ...in_conformer-rnn_transducer_streaming.yaml | 82 +++++++++++++++++++
 egs2/librispeech_100/asr_transducer1/db.sh    |  1 +
 egs2/librispeech_100/asr_transducer1/local    |  1 +
 egs2/librispeech_100/asr_transducer1/path.sh  |  1 +
 .../librispeech_100/asr_transducer1/pyscripts |  1 +
 egs2/librispeech_100/asr_transducer1/run.sh   | 37 +++++++++
 egs2/librispeech_100/asr_transducer1/scripts  |  1 +
 egs2/librispeech_100/asr_transducer1/steps    |  1 +
 egs2/librispeech_100/asr_transducer1/utils    |  1 +
 14 files changed, 229 insertions(+)
 create mode 120000 egs2/librispeech_100/asr_transducer1/asr.sh
 create mode 120000 egs2/librispeech_100/asr_transducer1/cmd.sh
 create mode 100644 egs2/librispeech_100/asr_transducer1/conf/decode.yaml
 create mode 100644 egs2/librispeech_100/asr_transducer1/conf/decode_streaming.yaml
 create mode 100644 egs2/librispeech_100/asr_transducer1/conf/train_conformer-rnn_transducer.yaml
 create mode 100644 egs2/librispeech_100/asr_transducer1/conf/train_conformer-rnn_transducer_streaming.yaml
 create mode 120000 egs2/librispeech_100/asr_transducer1/db.sh
 create mode 120000 egs2/librispeech_100/asr_transducer1/local
 create mode 120000 egs2/librispeech_100/asr_transducer1/path.sh
 create mode 120000 egs2/librispeech_100/asr_transducer1/pyscripts
 create mode 100755 egs2/librispeech_100/asr_transducer1/run.sh
 create mode 120000 egs2/librispeech_100/asr_transducer1/scripts
 create mode 120000 egs2/librispeech_100/asr_transducer1/steps
 create mode 120000 egs2/librispeech_100/asr_transducer1/utils

diff --git a/egs2/librispeech_100/asr_transducer1/asr.sh b/egs2/librispeech_100/asr_transducer1/asr.sh
new file mode 120000
index 00000000000..60b05122cfd
--- /dev/null
+++ b/egs2/librispeech_100/asr_transducer1/asr.sh
@@ -0,0 +1 @@
+../../TEMPLATE/asr1/asr.sh
\ No newline at end of file
diff --git a/egs2/librispeech_100/asr_transducer1/cmd.sh b/egs2/librispeech_100/asr_transducer1/cmd.sh
new file mode 120000
index 00000000000..297411a5ecc
--- /dev/null
+++ b/egs2/librispeech_100/asr_transducer1/cmd.sh
@@ -0,0 +1 @@
+../asr1/cmd.sh
\ No newline at end of file
diff --git a/egs2/librispeech_100/asr_transducer1/conf/decode.yaml b/egs2/librispeech_100/asr_transducer1/conf/decode.yaml
new file mode 100644
index 00000000000..72bff4d70c8
--- /dev/null
+++ b/egs2/librispeech_100/asr_transducer1/conf/decode.yaml
@@ -0,0 +1,14 @@
+beam_size: 5 # 10 produces slightly better results.
+beam_search_config:
+    search_type: default
+
+    # ALSD (search-type: alsd)
+    u_max: 50
+
+    # TSD (search-type: tsd)
+    max_sym_exp: 2
+
+    # mAES (search-type: maes)
+    nstep: 1
+    expansion_gamma: 1.5
+    expansion_beta: 1
diff --git a/egs2/librispeech_100/asr_transducer1/conf/decode_streaming.yaml b/egs2/librispeech_100/asr_transducer1/conf/decode_streaming.yaml
new file mode 100644
index 00000000000..a064a80d581
--- /dev/null
+++ b/egs2/librispeech_100/asr_transducer1/conf/decode_streaming.yaml
@@ -0,0 +1,9 @@
+beam_size: 5 # 10 produces slightly better results.
+beam_search_config:
+    search_type: maes
+    nstep: 1
+    expansion_gamma: 2.3
+    expansion_beta: 2
+streaming: True
+chunk_size: 64
+left_context: 256
\ No newline at end of file
diff --git a/egs2/librispeech_100/asr_transducer1/conf/train_conformer-rnn_transducer.yaml b/egs2/librispeech_100/asr_transducer1/conf/train_conformer-rnn_transducer.yaml
new file mode 100644
index 00000000000..f9e954c4762
--- /dev/null
+++ b/egs2/librispeech_100/asr_transducer1/conf/train_conformer-rnn_transducer.yaml
@@ -0,0 +1,78 @@
+# general
+batch_type: numel
+batch_bins: 4000000
+accum_grad: 8
+max_epoch: 60 # 100 produces better results.
+patience: none
+init: none
+num_att_plot: 0
+
+# optimizer
+optim: adam
+optim_conf:
+    lr: 0.002
+    weight_decay: 0.000001
+scheduler: warmuplr
+scheduler_conf:
+    warmup_steps: 15000
+
+# criterion
+val_scheduler_criterion:
+    - valid
+    - loss
+best_model_criterion:
+-   - valid
+    - loss
+    - min
+keep_nbest_models: 10 # 20 produces slightly better results.
+
+model_conf:
+    transducer_weight: 1.0
+    auxiliary_ctc_weight: 0.3
+    report_cer: True
+    report_wer: True
+
+# specaug conf
+specaug: specaug
+specaug_conf:
+    apply_time_warp: true
+    time_warp_window: 5
+    time_warp_mode: bicubic
+    apply_freq_mask: true
+    freq_mask_width_range:
+    - 0
+    - 27
+    num_freq_mask: 2
+    apply_time_mask: true
+    time_mask_width_ratio_range:
+    - 0.
+    - 0.05
+    num_time_mask: 5
+
+encoder_conf:
+    main_conf:
+      pos_wise_act_type: swish
+      conv_mod_act_type: swish
+      pos_enc_dropout_rate: 0.2
+    input_conf:
+      vgg_like: True
+    body_conf:
+    - block_type: conformer
+      linear_size: 1024
+      hidden_size: 256
+      heads: 4
+      dropout_rate: 0.1
+      pos_wise_dropout_rate: 0.1
+      att_dropout_rate: 0.1
+      conv_mod_kernel_size: 31
+      num_blocks: 18
+decoder: rnn
+decoder_conf:
+    rnn_type: lstm
+    num_layers: 1
+    embed_size: 256
+    hidden_size: 256
+    dropout_rate: 0.1
+    embed_dropout_rate: 0.2
+joint_network_conf:
+    joint_space_size: 256
diff --git a/egs2/librispeech_100/asr_transducer1/conf/train_conformer-rnn_transducer_streaming.yaml b/egs2/librispeech_100/asr_transducer1/conf/train_conformer-rnn_transducer_streaming.yaml
new file mode 100644
index 00000000000..4036367a5d8
--- /dev/null
+++ b/egs2/librispeech_100/asr_transducer1/conf/train_conformer-rnn_transducer_streaming.yaml
@@ -0,0 +1,82 @@
+# general
+batch_type: numel
+batch_bins: 4000000
+accum_grad: 8
+max_epoch: 60 # 100 produces better results.
+patience: none
+init: none
+num_att_plot: 0
+
+# optimizer
+optim: adam
+optim_conf:
+    lr: 0.002
+    weight_decay: 0.000001
+scheduler: warmuplr
+scheduler_conf:
+    warmup_steps: 15000
+
+# criterion
+val_scheduler_criterion:
+    - valid
+    - loss
+best_model_criterion:
+-   - valid
+    - loss
+    - min
+keep_nbest_models: 10 # 20 produces slightly better results.
+
+model_conf:
+    transducer_weight: 1.0
+    auxiliary_ctc_weight: 0.3
+    report_cer: True
+    report_wer: True
+
+# specaug conf
+specaug: specaug
+specaug_conf:
+    apply_time_warp: true
+    time_warp_window: 5
+    time_warp_mode: bicubic
+    apply_freq_mask: true
+    freq_mask_width_range:
+    - 0
+    - 27
+    num_freq_mask: 2
+    apply_time_mask: true
+    time_mask_width_ratio_range:
+    - 0.
+    - 0.05
+    num_time_mask: 5
+
+encoder_conf:
+    main_conf:
+      pos_wise_act_type: swish
+      conv_mod_act_type: swish
+      pos_enc_dropout_rate: 0.2
+      dynamic_chunk_training: True
+      short_chunk_size: 25
+      short_chunk_threshold: 0.75
+      left_chunk_size: 4
+    input_conf:
+      vgg_like: True
+    body_conf:
+    - block_type: conformer
+      linear_size: 1024
+      hidden_size: 256
+      heads: 4
+      dropout_rate: 0.1
+      pos_wise_dropout_rate: 0.1
+      att_dropout_rate: 0.1
+      conv_mod_kernel_size: 31
+      num_blocks: 18
+decoder: rnn
+decoder_conf:
+    rnn_type: lstm
+    num_layers: 1
+    embed_size: 256
+    hidden_size: 256
+    dropout_rate: 0.1
+    embed_dropout_rate: 0.2
+joint_network_conf:
+    joint_space_size: 256
diff --git a/egs2/librispeech_100/asr_transducer1/db.sh b/egs2/librispeech_100/asr_transducer1/db.sh
new file mode 120000
index 00000000000..50d86130898
--- /dev/null
+++ b/egs2/librispeech_100/asr_transducer1/db.sh
@@ -0,0 +1 @@
+../../TEMPLATE/asr1/db.sh
\ No newline at end of file
diff --git a/egs2/librispeech_100/asr_transducer1/local b/egs2/librispeech_100/asr_transducer1/local
new file mode 120000
index 00000000000..23830fb51b8
--- /dev/null
+++ b/egs2/librispeech_100/asr_transducer1/local
@@ -0,0 +1 @@
+../asr1/local
\ No newline at end of file
diff --git a/egs2/librispeech_100/asr_transducer1/path.sh b/egs2/librispeech_100/asr_transducer1/path.sh
new file mode 120000
index 00000000000..c9ac0a75bc6
--- /dev/null
+++ b/egs2/librispeech_100/asr_transducer1/path.sh
@@ -0,0 +1 @@
+../../TEMPLATE/asr1/path.sh
\ No newline at end of file
diff --git a/egs2/librispeech_100/asr_transducer1/pyscripts b/egs2/librispeech_100/asr_transducer1/pyscripts
new file mode 120000
index 00000000000..ac68ad75b60
--- /dev/null
+++ b/egs2/librispeech_100/asr_transducer1/pyscripts
@@ -0,0 +1 @@
+../../TEMPLATE/asr1/pyscripts
\ No newline at end of file
diff --git a/egs2/librispeech_100/asr_transducer1/run.sh b/egs2/librispeech_100/asr_transducer1/run.sh
new file mode 100755
index 00000000000..1821c0746f9
--- /dev/null
+++ b/egs2/librispeech_100/asr_transducer1/run.sh
@@ -0,0 +1,37 @@
+#!/usr/bin/env bash
+
+set -e
+set -u
+set -o pipefail
+
+train_set="train_clean_100"
+valid_set="dev"
+test_sets="test_clean test_other dev_clean dev_other"
+
+asr_config=conf/train_conformer-rnn_transducer.yaml
+inference_config=conf/decode.yaml
+inference_model=valid.loss.ave_10best.pth
+
+./asr.sh \
+    --asr_task asr_transducer \
+    --skip_data_prep false \
+    --skip_train false \
+    --skip_eval false \
+    --lang en \
+    --ngpu 1 \
+    --nj 32 \
+    --inference_nj 32 \
+    --nbpe 500 \
+    --max_wav_duration 30 \
+    --speed_perturb_factors "0.9 1.0 1.1" \
+    --audio_format "flac.ark" \
+    --feats_type raw \
+    --use_lm false \
+    --asr_config "${asr_config}" \
+    --inference_config "${inference_config}" \
+    --inference_asr_model "${inference_model}" \
+    --train_set "${train_set}" \
+    --valid_set "${valid_set}" \
+    --test_sets "${test_sets}" \
+    --lm_train_text "data/${train_set}/text" \
+    --bpe_train_text "data/${train_set}/text" "$@"
diff --git a/egs2/librispeech_100/asr_transducer1/scripts b/egs2/librispeech_100/asr_transducer1/scripts
new file mode 120000
index 00000000000..b25829705dc
--- /dev/null
+++ b/egs2/librispeech_100/asr_transducer1/scripts
@@ -0,0 +1 @@
+../../TEMPLATE/asr1/scripts
\ No newline at end of file
diff --git a/egs2/librispeech_100/asr_transducer1/steps b/egs2/librispeech_100/asr_transducer1/steps
new file mode 120000
index 00000000000..91f2d234e20
--- /dev/null
+++ b/egs2/librispeech_100/asr_transducer1/steps
@@ -0,0 +1 @@
+../../../tools/kaldi/egs/wsj/s5/steps
\ No newline at end of file
diff --git a/egs2/librispeech_100/asr_transducer1/utils b/egs2/librispeech_100/asr_transducer1/utils
new file mode 120000
index 00000000000..f49247da827
--- /dev/null
+++ b/egs2/librispeech_100/asr_transducer1/utils
@@ -0,0 +1 @@
+../../../tools/kaldi/egs/wsj/s5/utils
\ No newline at end of file

From 2db74a9587a32b659cf4e1abb6b611d9f9551e09 Mon Sep 17 00:00:00 2001
From: Florian Boyer <florian.boyer@airudit.com>
Date: Thu, 6 Oct 2022 15:01:23 +0000
Subject: [PATCH 05/15] typo

---
 espnet2/tasks/asr_transducer.py | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/espnet2/tasks/asr_transducer.py b/espnet2/tasks/asr_transducer.py
index 8c7786277d2..c464f3020e9 100644
--- a/espnet2/tasks/asr_transducer.py
+++ b/espnet2/tasks/asr_transducer.py
@@ -135,7 +135,9 @@ def add_task_arguments(cls, parser: argparse.ArgumentParser):
             default={},
             help="The keyword arguments for the joint network class.",
         )
-        group = group.add_argument_group(description="Preprocess related.")
+
+        group = parser.add_argument_group(description="Preprocess related.")
+
         group.add_argument(
             "--use_preprocessor",
             type=str2bool,

From 8a4c5ea41d6472204644a9b53e729d6b41b6bf23 Mon Sep 17 00:00:00 2001
From: Florian Boyer <florian.boyer@airudit.com>
Date: Fri, 7 Oct 2022 12:41:36 +0000
Subject: [PATCH 06/15] add streaming transducer recipe

---
 .../librispeech_100/asr_transducer1/README.md | 46 +++++++++++++++++++
 1 file changed, 46 insertions(+)
 create mode 100644 egs2/librispeech_100/asr_transducer1/README.md

diff --git a/egs2/librispeech_100/asr_transducer1/README.md b/egs2/librispeech_100/asr_transducer1/README.md
new file mode 100644
index 00000000000..51b9a42eb80
--- /dev/null
+++ b/egs2/librispeech_100/asr_transducer1/README.md
@@ -0,0 +1,46 @@
+# Streaming Conformer-RNN Transducer
+
+- General information
+  - Pretrained model: N.A
+  - Training config: conf/train_conformer-rnn_transducer.streaming.yaml
+  - Decoding config: conf/decode.yaml
+  - GPU: Nvidia A100 40Gb
+  - CPU: AMD EPYC 7502P 32c
+  - Peak VRAM usage during training: 36.7Gb
+  - Training time: ~ 26 hours
+  - Decoding time (32 jobs, 1 thread): ~9,1 minutes
+
+- Environments
+  - date: `Fri Oct  7 12:02:29 UTC 2022`
+  - python version: `3.8.10 (default, Jun 22 2022, 20:18:18)  [GCC 9.4.0]`
+  - espnet version: `espnet 202209`
+  - pytorch version: `pytorch 1.8.1+cu111`
+  - Git hash: `2db74a9587a32b659cf4e1abb6b611d9f9551e09`
+  - Commit date: `Thu Oct 6 15:01:23 2022 +0000`
+
+## WER
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|decode_asr_model_valid.loss.ave_10best/dev_clean|2703|54402|94.3|5.2|0.5|0.7|6.4|56.9|
+|decode_asr_model_valid.loss.ave_10best/dev_other|2864|50948|83.4|14.8|1.8|1.9|18.5|82.1|
+|decode_asr_model_valid.loss.ave_10best/test_clean|2620|52576|93.8|5.6|0.7|0.8|7.0|58.9|
+|decode_asr_model_valid.loss.ave_10best/test_other|2939|52343|82.9|15.0|2.0|1.8|18.9|83.5|
+
+## CER
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|decode_asr_model_valid.loss.ave_10best/dev_clean|2703|288456|98.2|1.0|0.8|0.6|2.4|56.9|
+|decode_asr_model_valid.loss.ave_10best/dev_other|2864|265951|93.1|4.1|2.9|1.9|8.9|82.1|
+|decode_asr_model_valid.loss.ave_10best/test_clean|2620|281530|98.0|1.1|0.9|0.6|2.6|58.9|
+|decode_asr_model_valid.loss.ave_10best/test_other|2939|272758|93.0|4.0|3.0|1.8|8.9|83.5|
+
+## TER
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|decode_asr_model_valid.loss.ave_10best/dev_clean|2703|107929|95.0|3.6|1.4|0.6|5.5|56.9|
+|decode_asr_model_valid.loss.ave_10best/dev_other|2864|98610|84.7|11.6|3.6|2.2|17.4|82.1|
+|decode_asr_model_valid.loss.ave_10best/test_clean|2620|105724|94.7|3.7|1.6|0.6|6.0|58.9|
+|decode_asr_model_valid.loss.ave_10best/test_other|2939|101026|84.3|11.6|4.1|2.0|17.7|83.5|

From 508f33986e3344919b690478798925721a24e72e Mon Sep 17 00:00:00 2001
From: Florian Boyer <florian.boyer@airudit.com>
Date: Sat, 8 Oct 2022 18:12:42 +0000
Subject: [PATCH 07/15] add offline model results

---
 .../librispeech_100/asr_transducer1/README.md | 55 ++++++++++++++++++-
 .../conf/decode_streaming.yaml                |  2 +-
 2 files changed, 53 insertions(+), 4 deletions(-)

diff --git a/egs2/librispeech_100/asr_transducer1/README.md b/egs2/librispeech_100/asr_transducer1/README.md
index 51b9a42eb80..19cad9158c0 100644
--- a/egs2/librispeech_100/asr_transducer1/README.md
+++ b/egs2/librispeech_100/asr_transducer1/README.md
@@ -1,14 +1,15 @@
 # Streaming Conformer-RNN Transducer
+# asr_train_conformer-rnn_transducer_streaming_raw_en_bpe500_sp
 
 - General information
   - Pretrained model: N.A
-  - Training config: conf/train_conformer-rnn_transducer.streaming.yaml
-  - Decoding config: conf/decode.yaml
+  - Training config: conf/train_conformer-rnn_transducer_streaming.yaml
+  - Decoding config: conf/decode.yaml (or conf/decode_streaming.yaml)
   - GPU: Nvidia A100 40Gb
   - CPU: AMD EPYC 7502P 32c
   - Peak VRAM usage during training: 36.7Gb
   - Training time: ~ 26 hours
-  - Decoding time (32 jobs, 1 thread): ~9,1 minutes
+  - Decoding time (32 jobs, 1 thread): ~9,1 minutes (full context)
 
 - Environments
   - date: `Fri Oct  7 12:02:29 UTC 2022`
@@ -44,3 +45,51 @@
 |decode_asr_model_valid.loss.ave_10best/dev_other|2864|98610|84.7|11.6|3.6|2.2|17.4|82.1|
 |decode_asr_model_valid.loss.ave_10best/test_clean|2620|105724|94.7|3.7|1.6|0.6|6.0|58.9|
 |decode_asr_model_valid.loss.ave_10best/test_other|2939|101026|84.3|11.6|4.1|2.0|17.7|83.5|
+
+# Conformer-RNN Transducer
+# asr_train_conformer-rnn_transducer_raw_en_bpe500_sp
+
+- General information
+  - Pretrained model: N.A
+  - Training config: conf/train_conformer-rnn_transducer.yaml
+  - Decoding config: conf/decode.yaml
+  - GPU: Nvidia A100 40Gb
+  - CPU: AMD EPYC 7502P 32c
+  - Peak VRAM usage during training: 36.4Gb
+  - Training time: ~ 26 hours
+  - Decoding time (32 jobs, 1 thread): ~9 minutes
+
+- Environments
+  - date: `Fri Oct  7 12:02:29 UTC 2022`
+  - python version: `3.8.10 (default, Jun 22 2022, 20:18:18)  [GCC 9.4.0]`
+  - espnet version: `espnet 202209`
+  - pytorch version: `pytorch 1.8.1+cu111`
+  - Git hash: `2db74a9587a32b659cf4e1abb6b611d9f9551e09`
+  - Commit date: `Thu Oct 6 15:01:23 2022 +0000`
+
+## WER
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|decode_asr_model_valid.loss.ave_10best/dev_clean|2703|54402|94.7|4.8|0.4|0.6|5.9|55.1|
+|decode_asr_model_valid.loss.ave_10best/dev_other|2864|50948|84.2|14.1|1.7|1.8|17.6|80.2|
+|decode_asr_model_valid.loss.ave_10best/test_clean|2620|52576|94.3|5.2|0.6|0.7|6.4|56.8|
+|decode_asr_model_valid.loss.ave_10best/test_other|2939|52343|83.9|14.2|1.9|1.8|17.9|81.5|
+
+## CER
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|decode_asr_model_valid.loss.ave_10best/dev_clean|2703|288456|98.4|0.9|0.7|0.6|2.2|55.1|
+|decode_asr_model_valid.loss.ave_10best/dev_other|2864|265951|93.4|3.9|2.7|1.9|8.5|80.2|
+|decode_asr_model_valid.loss.ave_10best/test_clean|2620|281530|98.2|1.0|0.8|0.6|2.3|56.8|
+|decode_asr_model_valid.loss.ave_10best/test_other|2939|272758|93.5|3.8|2.7|1.8|8.3|81.5|
+
+## TER
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|decode_asr_model_valid.loss.ave_10best/dev_clean|2703|107929|95.4|3.4|1.2|0.6|5.2|55.1|
+|decode_asr_model_valid.loss.ave_10best/dev_other|2864|98610|85.5|11.0|3.5|2.1|16.6|80.2|
+|decode_asr_model_valid.loss.ave_10best/test_clean|2620|105724|95.1|3.4|1.4|0.6|5.5|56.8|
+|decode_asr_model_valid.loss.ave_10best/test_other|2939|101026|85.3|10.9|3.9|2.0|16.7|81.5|
diff --git a/egs2/librispeech_100/asr_transducer1/conf/decode_streaming.yaml b/egs2/librispeech_100/asr_transducer1/conf/decode_streaming.yaml
index a064a80d581..f7a3da0a1b5 100644
--- a/egs2/librispeech_100/asr_transducer1/conf/decode_streaming.yaml
+++ b/egs2/librispeech_100/asr_transducer1/conf/decode_streaming.yaml
@@ -6,4 +6,4 @@ beam_search_config:
     expansion_beta: 2
 streaming: True
 chunk_size: 64
-left_context: 256
\ No newline at end of file
+left_context: 256

From 9f4f12da4610cae21d0db0283eb28ff610edf838 Mon Sep 17 00:00:00 2001
From: Florian Boyer <florian.boyer@airudit.com>
Date: Fri, 21 Oct 2022 11:18:43 +0000
Subject: [PATCH 08/15] fix conflict

---
 espnet2/asr_transducer/encoder/modules/convolution.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/espnet2/asr_transducer/encoder/modules/convolution.py b/espnet2/asr_transducer/encoder/modules/convolution.py
index 0e50f33292c..df3e7dcc558 100644
--- a/espnet2/asr_transducer/encoder/modules/convolution.py
+++ b/espnet2/asr_transducer/encoder/modules/convolution.py
@@ -89,7 +89,7 @@ def forward(
         x = torch.nn.functional.glu(x, dim=1)
 
         if mask is not None:
-            x.masked_fill(mask.unsqueeze(1).expand_as(x), 0.0)
+            x.masked_fill_(mask.unsqueeze(1).expand_as(x), 0.0)
 
         if self.lorder > 0:
             if cache is None:

From 359ac2de185dd6b89be31f6343f842c65e3e8795 Mon Sep 17 00:00:00 2001
From: Florian Boyer <florian.boyer@airudit.com>
Date: Fri, 21 Oct 2022 11:22:14 +0000
Subject: [PATCH 09/15] add sub factor param for new version

---
 .../asr_transducer1/conf/train_conformer-rnn_transducer.yaml     | 1 +
 .../conf/train_conformer-rnn_transducer_streaming.yaml           | 1 +
 2 files changed, 2 insertions(+)

diff --git a/egs2/librispeech_100/asr_transducer1/conf/train_conformer-rnn_transducer.yaml b/egs2/librispeech_100/asr_transducer1/conf/train_conformer-rnn_transducer.yaml
index f9e954c4762..f265019c581 100644
--- a/egs2/librispeech_100/asr_transducer1/conf/train_conformer-rnn_transducer.yaml
+++ b/egs2/librispeech_100/asr_transducer1/conf/train_conformer-rnn_transducer.yaml
@@ -56,6 +56,7 @@ encoder_conf:
       pos_enc_dropout_rate: 0.2
     input_conf:
       vgg_like: True
+      subsampling_factor: 6
     body_conf:
     - block_type: conformer
       linear_size: 1024
diff --git a/egs2/librispeech_100/asr_transducer1/conf/train_conformer-rnn_transducer_streaming.yaml b/egs2/librispeech_100/asr_transducer1/conf/train_conformer-rnn_transducer_streaming.yaml
index 4036367a5d8..5e93a7308f7 100644
--- a/egs2/librispeech_100/asr_transducer1/conf/train_conformer-rnn_transducer_streaming.yaml
+++ b/egs2/librispeech_100/asr_transducer1/conf/train_conformer-rnn_transducer_streaming.yaml
@@ -60,6 +60,7 @@ encoder_conf:
       left_chunk_size: 4
     input_conf:
       vgg_like: True
+      subsampling_factor: 6
     body_conf:
     - block_type: conformer
       linear_size: 1024

From 9e5005d1b85ab8561dabd007fc2b6e605f713061 Mon Sep 17 00:00:00 2001
From: Florian Boyer <florian.boyer@airudit.com>
Date: Fri, 10 Feb 2023 14:24:27 +0000
Subject: [PATCH 10/15] update offline model and run script

---
 .../librispeech_100/asr_transducer1/README.md | 40 +++++++++----------
 .../asr_transducer1/conf/decode.yaml          |  8 ++--
 .../conf/train_conformer-rnn_transducer.yaml  |  6 +--
 egs2/librispeech_100/asr_transducer1/run.sh   |  3 --
 4 files changed, 27 insertions(+), 30 deletions(-)

diff --git a/egs2/librispeech_100/asr_transducer1/README.md b/egs2/librispeech_100/asr_transducer1/README.md
index 19cad9158c0..b1c429e41bc 100644
--- a/egs2/librispeech_100/asr_transducer1/README.md
+++ b/egs2/librispeech_100/asr_transducer1/README.md
@@ -55,41 +55,41 @@
   - Decoding config: conf/decode.yaml
   - GPU: Nvidia A100 40Gb
   - CPU: AMD EPYC 7502P 32c
-  - Peak VRAM usage during training: 36.4Gb
-  - Training time: ~ 26 hours
-  - Decoding time (32 jobs, 1 thread): ~9 minutes
+  - Peak VRAM usage during training: 37.09 Gb
+  - Training time: ~ 35 hours
+  - Decoding time (32 jobs, 1 thread): ~15,6 minutes w/ default beam search.
 
 - Environments
-  - date: `Fri Oct  7 12:02:29 UTC 2022`
-  - python version: `3.8.10 (default, Jun 22 2022, 20:18:18)  [GCC 9.4.0]`
-  - espnet version: `espnet 202209`
+  - date: `Fri Feb 10 09:27:45 UTC 2023`
+  - python version: `3.8.10 (default, Nov 14 2022, 12:59:47)  [GCC 9.4.0]`
+  - espnet version: `espnet 202301`
   - pytorch version: `pytorch 1.8.1+cu111`
-  - Git hash: `2db74a9587a32b659cf4e1abb6b611d9f9551e09`
-  - Commit date: `Thu Oct 6 15:01:23 2022 +0000`
+  - Git hash: `01893f855ca1a3a3645547ee4d3eaf461f7601bf`
+  - Commit date: `Thu Feb 9 10:04:57 2023 +0000`
 
 ## WER
 
 |dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
 |---|---|---|---|---|---|---|---|---|
-|decode_asr_model_valid.loss.ave_10best/dev_clean|2703|54402|94.7|4.8|0.4|0.6|5.9|55.1|
-|decode_asr_model_valid.loss.ave_10best/dev_other|2864|50948|84.2|14.1|1.7|1.8|17.6|80.2|
-|decode_asr_model_valid.loss.ave_10best/test_clean|2620|52576|94.3|5.2|0.6|0.7|6.4|56.8|
-|decode_asr_model_valid.loss.ave_10best/test_other|2939|52343|83.9|14.2|1.9|1.8|17.9|81.5|
+|decode_asr_model_valid.loss.ave_10best/dev_clean|2703|54402|94.9|4.7|0.5|0.6|5.8|53.6|
+|decode_asr_model_valid.loss.ave_10best/dev_other|2864|50948|84.9|13.4|1.6|1.8|16.9|78.9|
+|decode_asr_model_valid.loss.ave_10best/test_clean|2620|52576|94.6|4.8|0.6|0.6|6.0|54.8|
+|decode_asr_model_valid.loss.ave_10best/test_other|2939|52343|84.7|13.6|1.8|1.6|17.0|80.1|
 
 ## CER
 
 |dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
 |---|---|---|---|---|---|---|---|---|
-|decode_asr_model_valid.loss.ave_10best/dev_clean|2703|288456|98.4|0.9|0.7|0.6|2.2|55.1|
-|decode_asr_model_valid.loss.ave_10best/dev_other|2864|265951|93.4|3.9|2.7|1.9|8.5|80.2|
-|decode_asr_model_valid.loss.ave_10best/test_clean|2620|281530|98.2|1.0|0.8|0.6|2.3|56.8|
-|decode_asr_model_valid.loss.ave_10best/test_other|2939|272758|93.5|3.8|2.7|1.8|8.3|81.5|
+|decode_asr_model_valid.loss.ave_10best/dev_clean|2703|288456|98.4|0.9|0.6|0.6|2.1|53.6|
+|decode_asr_model_valid.loss.ave_10best/dev_other|2864|265951|93.8|3.7|2.4|1.8|8.0|78.9|
+|decode_asr_model_valid.loss.ave_10best/test_clean|2620|281530|98.4|0.9|0.7|0.5|2.2|54.8|
+|decode_asr_model_valid.loss.ave_10best/test_other|2939|272758|93.9|3.6|2.5|1.7|7.9|80.1|
 
 ## TER
 
 |dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
 |---|---|---|---|---|---|---|---|---|
-|decode_asr_model_valid.loss.ave_10best/dev_clean|2703|107929|95.4|3.4|1.2|0.6|5.2|55.1|
-|decode_asr_model_valid.loss.ave_10best/dev_other|2864|98610|85.5|11.0|3.5|2.1|16.6|80.2|
-|decode_asr_model_valid.loss.ave_10best/test_clean|2620|105724|95.1|3.4|1.4|0.6|5.5|56.8|
-|decode_asr_model_valid.loss.ave_10best/test_other|2939|101026|85.3|10.9|3.9|2.0|16.7|81.5|
+|decode_asr_model_valid.loss.ave_10best/dev_clean|2703|107929|95.5|3.4|1.1|0.6|5.0|53.6|
+|decode_asr_model_valid.loss.ave_10best/dev_other|2864|98610|86.2|10.6|3.2|2.0|15.9|78.9|
+|decode_asr_model_valid.loss.ave_10best/test_clean|2620|105724|95.5|3.2|1.3|0.6|5.1|54.8|
+|decode_asr_model_valid.loss.ave_10best/test_other|2939|101026|86.0|10.4|3.5|1.9|15.9|80.1|
diff --git a/egs2/librispeech_100/asr_transducer1/conf/decode.yaml b/egs2/librispeech_100/asr_transducer1/conf/decode.yaml
index 72bff4d70c8..dd2bc435aa8 100644
--- a/egs2/librispeech_100/asr_transducer1/conf/decode.yaml
+++ b/egs2/librispeech_100/asr_transducer1/conf/decode.yaml
@@ -1,14 +1,14 @@
-beam_size: 5 # 10 produces slightly better results.
+beam_size: 10 # 5 is almost equivalent.
 beam_search_config:
     search_type: default
 
     # ALSD (search-type: alsd)
-    u_max: 50
+    u_max: 150
 
     # TSD (search-type: tsd)
-    max_sym_exp: 2
+    max_sym_exp: 3
 
     # mAES (search-type: maes)
-    nstep: 1
+    nstep: 2
     expansion_gamma: 1.5
     expansion_beta: 1
diff --git a/egs2/librispeech_100/asr_transducer1/conf/train_conformer-rnn_transducer.yaml b/egs2/librispeech_100/asr_transducer1/conf/train_conformer-rnn_transducer.yaml
index f265019c581..3eb4f52aa63 100644
--- a/egs2/librispeech_100/asr_transducer1/conf/train_conformer-rnn_transducer.yaml
+++ b/egs2/librispeech_100/asr_transducer1/conf/train_conformer-rnn_transducer.yaml
@@ -2,7 +2,7 @@
 batch_type: numel
 batch_bins: 4000000
 accum_grad: 8
-max_epoch: 60 # 100 produces better results.
+max_epoch: 60 # 100 produces slightly better results.
 patience: none
 init: none
 num_att_plot: 0
@@ -24,7 +24,7 @@ best_model_criterion:
 -   - valid
     - loss
     - min
-keep_nbest_models: 10 # 20 produces slightly better results.
+keep_nbest_models: 10
 
 model_conf:
     transducer_weight: 1.0
@@ -56,7 +56,7 @@ encoder_conf:
       pos_enc_dropout_rate: 0.2
     input_conf:
       vgg_like: True
-      subsampling_factor: 6
+      subsampling_factor: 4
     body_conf:
     - block_type: conformer
       linear_size: 1024
diff --git a/egs2/librispeech_100/asr_transducer1/run.sh b/egs2/librispeech_100/asr_transducer1/run.sh
index 1821c0746f9..ade3a73e1b0 100755
--- a/egs2/librispeech_100/asr_transducer1/run.sh
+++ b/egs2/librispeech_100/asr_transducer1/run.sh
@@ -14,9 +14,6 @@ inference_model=valid.loss.ave_10best.pth
 
 ./asr.sh \
     --asr_task asr_transducer \
-    --skip_data_prep false \
-    --skip_train false \
-    --skip_eval false \
     --lang en \
     --ngpu 1 \
     --nj 32 \

From 17af213961e275a853e617ea0dd24f4ec8b3afa6 Mon Sep 17 00:00:00 2001
From: Florian Boyer <florian.boyer@airudit.com>
Date: Tue, 14 Feb 2023 09:46:51 +0000
Subject: [PATCH 11/15] add ebranchformer config and results

---
 .../librispeech_100/asr_transducer1/README.md | 48 +++++++++++
 .../train_ebranchformer-rnn_transducer.yaml   | 79 +++++++++++++++++++
 2 files changed, 127 insertions(+)
 create mode 100644 egs2/librispeech_100/asr_transducer1/conf/train_ebranchformer-rnn_transducer.yaml

diff --git a/egs2/librispeech_100/asr_transducer1/README.md b/egs2/librispeech_100/asr_transducer1/README.md
index b1c429e41bc..2ff98780397 100644
--- a/egs2/librispeech_100/asr_transducer1/README.md
+++ b/egs2/librispeech_100/asr_transducer1/README.md
@@ -93,3 +93,51 @@
 |decode_asr_model_valid.loss.ave_10best/dev_other|2864|98610|86.2|10.6|3.2|2.0|15.9|78.9|
 |decode_asr_model_valid.loss.ave_10best/test_clean|2620|105724|95.5|3.2|1.3|0.6|5.1|54.8|
 |decode_asr_model_valid.loss.ave_10best/test_other|2939|101026|86.0|10.4|3.5|1.9|15.9|80.1|
+
+# E-Branchformer-RNN Transducer
+# asr_train_ebranchformer-rnn_transducer_raw_en_bpe500_sp
+
+- General information
+  - Pretrained model: N.A
+  - Training config: conf/train_ebranchformer-rnn_transducer.yaml
+  - Decoding config: conf/decode.yaml
+  - GPU: Nvidia A100 40Gb
+  - CPU: AMD EPYC 7502P 32c
+  - Peak VRAM usage during training: 37.39 Gb
+  - Training time: ~ 33,8 hours
+  - Decoding time (32 jobs, 1 thread): ~15,7 minutes w/ default beam search.
+
+- Environments
+  - date: Tue Feb 14 07:41:14 UTC 2023`
+  - python version: `3.8.10 (default, Nov 14 2022, 12:59:47)  [GCC 9.4.0]`
+  - espnet version: `espnet 202301`
+  - pytorch version: `pytorch 1.8.1+cu111`
+  - Git hash: `01893f855ca1a3a3645547ee4d3eaf461f7601bf`
+  - Commit date: `Thu Feb 9 10:04:57 2023 +0000`
+
+## WER
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|decode_asr_model_valid.loss.ave_10best/dev_clean|2703|54402|94.9|4.7|0.4|0.6|5.7|53.0|
+|decode_asr_model_valid.loss.ave_10best/dev_other|2864|50948|85.0|13.4|1.6|1.8|16.8|77.9|
+|decode_asr_model_valid.loss.ave_10best/test_clean|2620|52576|94.6|4.9|0.5|0.6|6.0|55.5|
+|decode_asr_model_valid.loss.ave_10best/test_other|2939|52343|84.7|13.6|1.8|1.7|17.1|80.6|
+
+## CER
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|decode_asr_model_valid.loss.ave_10best/dev_clean|2703|288456|98.5|0.9|0.6|0.6|2.1|53.0|
+|decode_asr_model_valid.loss.ave_10best/dev_other|2864|265951|93.9|3.7|2.4|1.8|7.9|77.9|
+|decode_asr_model_valid.loss.ave_10best/test_clean|2620|281530|98.4|0.9|0.7|0.5|2.1|55.5|
+|decode_asr_model_valid.loss.ave_10best/test_other|2939|272758|93.9|3.6|2.5|1.8|7.9|80.6|
+
+## TER
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|decode_asr_model_valid.loss.ave_10best/dev_clean|2703|107929|95.6|3.3|1.1|0.6|5.0|53.0|
+|decode_asr_model_valid.loss.ave_10best/dev_other|2864|98610|86.2|10.6|3.2|2.0|15.8|77.9|
+|decode_asr_model_valid.loss.ave_10best/test_clean|2620|105724|95.4|3.2|1.3|0.6|5.1|55.5|
+|decode_asr_model_valid.loss.ave_10best/test_other|2939|101026|85.9|10.5|3.6|2.0|16.0|80.6|
diff --git a/egs2/librispeech_100/asr_transducer1/conf/train_ebranchformer-rnn_transducer.yaml b/egs2/librispeech_100/asr_transducer1/conf/train_ebranchformer-rnn_transducer.yaml
new file mode 100644
index 00000000000..fd2fb1804f1
--- /dev/null
+++ b/egs2/librispeech_100/asr_transducer1/conf/train_ebranchformer-rnn_transducer.yaml
@@ -0,0 +1,79 @@
+# general
+batch_type: numel
+batch_bins: 4000000
+accum_grad: 8
+max_epoch: 60 # 100 produces slightly better results.
+patience: none
+init: none
+num_att_plot: 0
+
+# optimizer
+optim: adam
+optim_conf:
+    lr: 0.002
+    weight_decay: 0.000001
+scheduler: warmuplr
+scheduler_conf:
+    warmup_steps: 15000
+
+# criterion
+val_scheduler_criterion:
+    - valid
+    - loss
+best_model_criterion:
+-   - valid
+    - loss
+    - min
+keep_nbest_models: 10
+
+model_conf:
+    transducer_weight: 1.0
+    auxiliary_ctc_weight: 0.3
+    report_cer: True
+    report_wer: True
+
+# specaug conf
+specaug: specaug
+specaug_conf:
+    apply_time_warp: true
+    time_warp_window: 5
+    time_warp_mode: bicubic
+    apply_freq_mask: true
+    freq_mask_width_range:
+    - 0
+    - 27
+    num_freq_mask: 2
+    apply_time_mask: true
+    time_mask_width_ratio_range:
+    - 0.
+    - 0.05
+    num_time_mask: 5
+
+encoder_conf:
+    main_conf:
+      pos_wise_act_type: swish
+      conv_mod_act_type: swish
+      pos_enc_dropout_rate: 0.2
+    input_conf:
+      vgg_like: True
+      subsampling_factor: 4
+    body_conf:
+    - block_type: ebranchformer
+      linear_size: 1024
+      hidden_size: 256
+      heads: 4
+      dropout_rate: 0.1
+      pos_wise_dropout_rate: 0.1
+      att_dropout_rate: 0.1
+      conv_mod_kernel_size: 31
+      num_blocks: 14
+decoder: rnn
+decoder_conf:
+    rnn_type: lstm
+    num_layers: 1
+    embed_size: 256
+    hidden_size: 256
+    dropout_rate: 0.1
+    embed_dropout_rate: 0.2
+joint_network_conf:
+    joint_space_size: 256

From aa033d1bb3040bfd03a5c73d66df30cc467d6020 Mon Sep 17 00:00:00 2001
From: Florian Boyer <florian.boyer@airudit.com>
Date: Tue, 14 Feb 2023 09:53:15 +0000
Subject: [PATCH 12/15] rework README

---
 .../librispeech_100/asr_transducer1/README.md | 118 +++++++++---------
 1 file changed, 61 insertions(+), 57 deletions(-)

diff --git a/egs2/librispeech_100/asr_transducer1/README.md b/egs2/librispeech_100/asr_transducer1/README.md
index 2ff98780397..60238f147b3 100644
--- a/egs2/librispeech_100/asr_transducer1/README.md
+++ b/egs2/librispeech_100/asr_transducer1/README.md
@@ -1,53 +1,7 @@
-# Streaming Conformer-RNN Transducer
-# asr_train_conformer-rnn_transducer_streaming_raw_en_bpe500_sp
+# OFFLINE SYSTEMS
 
-- General information
-  - Pretrained model: N.A
-  - Training config: conf/train_conformer-rnn_transducer_streaming.yaml
-  - Decoding config: conf/decode.yaml (or conf/decode_streaming.yaml)
-  - GPU: Nvidia A100 40Gb
-  - CPU: AMD EPYC 7502P 32c
-  - Peak VRAM usage during training: 36.7Gb
-  - Training time: ~ 26 hours
-  - Decoding time (32 jobs, 1 thread): ~9,1 minutes (full context)
-
-- Environments
-  - date: `Fri Oct  7 12:02:29 UTC 2022`
-  - python version: `3.8.10 (default, Jun 22 2022, 20:18:18)  [GCC 9.4.0]`
-  - espnet version: `espnet 202209`
-  - pytorch version: `pytorch 1.8.1+cu111`
-  - Git hash: `2db74a9587a32b659cf4e1abb6b611d9f9551e09`
-  - Commit date: `Thu Oct 6 15:01:23 2022 +0000`
-
-## WER
-
-|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
-|---|---|---|---|---|---|---|---|---|
-|decode_asr_model_valid.loss.ave_10best/dev_clean|2703|54402|94.3|5.2|0.5|0.7|6.4|56.9|
-|decode_asr_model_valid.loss.ave_10best/dev_other|2864|50948|83.4|14.8|1.8|1.9|18.5|82.1|
-|decode_asr_model_valid.loss.ave_10best/test_clean|2620|52576|93.8|5.6|0.7|0.8|7.0|58.9|
-|decode_asr_model_valid.loss.ave_10best/test_other|2939|52343|82.9|15.0|2.0|1.8|18.9|83.5|
-
-## CER
-
-|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
-|---|---|---|---|---|---|---|---|---|
-|decode_asr_model_valid.loss.ave_10best/dev_clean|2703|288456|98.2|1.0|0.8|0.6|2.4|56.9|
-|decode_asr_model_valid.loss.ave_10best/dev_other|2864|265951|93.1|4.1|2.9|1.9|8.9|82.1|
-|decode_asr_model_valid.loss.ave_10best/test_clean|2620|281530|98.0|1.1|0.9|0.6|2.6|58.9|
-|decode_asr_model_valid.loss.ave_10best/test_other|2939|272758|93.0|4.0|3.0|1.8|8.9|83.5|
-
-## TER
-
-|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
-|---|---|---|---|---|---|---|---|---|
-|decode_asr_model_valid.loss.ave_10best/dev_clean|2703|107929|95.0|3.6|1.4|0.6|5.5|56.9|
-|decode_asr_model_valid.loss.ave_10best/dev_other|2864|98610|84.7|11.6|3.6|2.2|17.4|82.1|
-|decode_asr_model_valid.loss.ave_10best/test_clean|2620|105724|94.7|3.7|1.6|0.6|6.0|58.9|
-|decode_asr_model_valid.loss.ave_10best/test_other|2939|101026|84.3|11.6|4.1|2.0|17.7|83.5|
-
-# Conformer-RNN Transducer
-# asr_train_conformer-rnn_transducer_raw_en_bpe500_sp
+## Conformer/RNN Transducer
+## asr_train_conformer-rnn_transducer_raw_en_bpe500_sp
 
 - General information
   - Pretrained model: N.A
@@ -67,7 +21,7 @@
   - Git hash: `01893f855ca1a3a3645547ee4d3eaf461f7601bf`
   - Commit date: `Thu Feb 9 10:04:57 2023 +0000`
 
-## WER
+### WER
 
 |dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
 |---|---|---|---|---|---|---|---|---|
@@ -76,7 +30,7 @@
 |decode_asr_model_valid.loss.ave_10best/test_clean|2620|52576|94.6|4.8|0.6|0.6|6.0|54.8|
 |decode_asr_model_valid.loss.ave_10best/test_other|2939|52343|84.7|13.6|1.8|1.6|17.0|80.1|
 
-## CER
+### CER
 
 |dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
 |---|---|---|---|---|---|---|---|---|
@@ -85,7 +39,7 @@
 |decode_asr_model_valid.loss.ave_10best/test_clean|2620|281530|98.4|0.9|0.7|0.5|2.2|54.8|
 |decode_asr_model_valid.loss.ave_10best/test_other|2939|272758|93.9|3.6|2.5|1.7|7.9|80.1|
 
-## TER
+### TER
 
 |dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
 |---|---|---|---|---|---|---|---|---|
@@ -94,8 +48,8 @@
 |decode_asr_model_valid.loss.ave_10best/test_clean|2620|105724|95.5|3.2|1.3|0.6|5.1|54.8|
 |decode_asr_model_valid.loss.ave_10best/test_other|2939|101026|86.0|10.4|3.5|1.9|15.9|80.1|
 
-# E-Branchformer-RNN Transducer
-# asr_train_ebranchformer-rnn_transducer_raw_en_bpe500_sp
+## E-Branchformer/RNN Transducer
+## asr_train_ebranchformer-rnn_transducer_raw_en_bpe500_sp
 
 - General information
   - Pretrained model: N.A
@@ -115,7 +69,7 @@
   - Git hash: `01893f855ca1a3a3645547ee4d3eaf461f7601bf`
   - Commit date: `Thu Feb 9 10:04:57 2023 +0000`
 
-## WER
+### WER
 
 |dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
 |---|---|---|---|---|---|---|---|---|
@@ -124,7 +78,7 @@
 |decode_asr_model_valid.loss.ave_10best/test_clean|2620|52576|94.6|4.9|0.5|0.6|6.0|55.5|
 |decode_asr_model_valid.loss.ave_10best/test_other|2939|52343|84.7|13.6|1.8|1.7|17.1|80.6|
 
-## CER
+### CER
 
 |dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
 |---|---|---|---|---|---|---|---|---|
@@ -133,7 +87,7 @@
 |decode_asr_model_valid.loss.ave_10best/test_clean|2620|281530|98.4|0.9|0.7|0.5|2.1|55.5|
 |decode_asr_model_valid.loss.ave_10best/test_other|2939|272758|93.9|3.6|2.5|1.8|7.9|80.6|
 
-## TER
+### TER
 
 |dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
 |---|---|---|---|---|---|---|---|---|
@@ -141,3 +95,53 @@
 |decode_asr_model_valid.loss.ave_10best/dev_other|2864|98610|86.2|10.6|3.2|2.0|15.8|77.9|
 |decode_asr_model_valid.loss.ave_10best/test_clean|2620|105724|95.4|3.2|1.3|0.6|5.1|55.5|
 |decode_asr_model_valid.loss.ave_10best/test_other|2939|101026|85.9|10.5|3.6|2.0|16.0|80.6|
+
+# STREAMING SYSTEMS
+
+## Conformer/RNN Transducer
+## asr_train_conformer-rnn_transducer_streaming_raw_en_bpe500_sp
+
+- General information
+  - Pretrained model: N.A
+  - Training config: conf/train_conformer-rnn_transducer_streaming.yaml
+  - Decoding config: conf/decode.yaml (or conf/decode_streaming.yaml)
+  - GPU: Nvidia A100 40Gb
+  - CPU: AMD EPYC 7502P 32c
+  - Peak VRAM usage during training: 36.7Gb
+  - Training time: ~ 26 hours
+  - Decoding time (32 jobs, 1 thread): ~9,1 minutes (full context)
+
+- Environments
+  - date: `Fri Oct  7 12:02:29 UTC 2022`
+  - python version: `3.8.10 (default, Jun 22 2022, 20:18:18)  [GCC 9.4.0]`
+  - espnet version: `espnet 202209`
+  - pytorch version: `pytorch 1.8.1+cu111`
+  - Git hash: `2db74a9587a32b659cf4e1abb6b611d9f9551e09`
+  - Commit date: `Thu Oct 6 15:01:23 2022 +0000`
+
+### WER
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|decode_asr_model_valid.loss.ave_10best/dev_clean|2703|54402|94.3|5.2|0.5|0.7|6.4|56.9|
+|decode_asr_model_valid.loss.ave_10best/dev_other|2864|50948|83.4|14.8|1.8|1.9|18.5|82.1|
+|decode_asr_model_valid.loss.ave_10best/test_clean|2620|52576|93.8|5.6|0.7|0.8|7.0|58.9|
+|decode_asr_model_valid.loss.ave_10best/test_other|2939|52343|82.9|15.0|2.0|1.8|18.9|83.5|
+
+### CER
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|decode_asr_model_valid.loss.ave_10best/dev_clean|2703|288456|98.2|1.0|0.8|0.6|2.4|56.9|
+|decode_asr_model_valid.loss.ave_10best/dev_other|2864|265951|93.1|4.1|2.9|1.9|8.9|82.1|
+|decode_asr_model_valid.loss.ave_10best/test_clean|2620|281530|98.0|1.1|0.9|0.6|2.6|58.9|
+|decode_asr_model_valid.loss.ave_10best/test_other|2939|272758|93.0|4.0|3.0|1.8|8.9|83.5|
+
+### TER
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|decode_asr_model_valid.loss.ave_10best/dev_clean|2703|107929|95.0|3.6|1.4|0.6|5.5|56.9|
+|decode_asr_model_valid.loss.ave_10best/dev_other|2864|98610|84.7|11.6|3.6|2.2|17.4|82.1|
+|decode_asr_model_valid.loss.ave_10best/test_clean|2620|105724|94.7|3.7|1.6|0.6|6.0|58.9|
+|decode_asr_model_valid.loss.ave_10best/test_other|2939|101026|84.3|11.6|4.1|2.0|17.7|83.5|
\ No newline at end of file

From 29c54874631ea3a2f60ac1b45a04cf8106483ae4 Mon Sep 17 00:00:00 2001
From: b-flo <florian.boyer@airudit.com>
Date: Thu, 22 Jun 2023 08:24:24 +0000
Subject: [PATCH 13/15] add MEGA config and results

---
 .../librispeech_100/asr_transducer1/README.md | 47 ++++++++++
 .../train_ebranchformer-mega_transducer.yaml  | 85 +++++++++++++++++++
 2 files changed, 132 insertions(+)
 create mode 100644 egs2/librispeech_100/asr_transducer1/conf/train_ebranchformer-mega_transducer.yaml

diff --git a/egs2/librispeech_100/asr_transducer1/README.md b/egs2/librispeech_100/asr_transducer1/README.md
index 60238f147b3..a17a686ac06 100644
--- a/egs2/librispeech_100/asr_transducer1/README.md
+++ b/egs2/librispeech_100/asr_transducer1/README.md
@@ -96,6 +96,53 @@
 |decode_asr_model_valid.loss.ave_10best/test_clean|2620|105724|95.4|3.2|1.3|0.6|5.1|55.5|
 |decode_asr_model_valid.loss.ave_10best/test_other|2939|101026|85.9|10.5|3.6|2.0|16.0|80.6|
 
+## E-Branchformer/MEGA Transducer
+## asr_train_ebranchformer-mega_transducer_raw_en_bpe500_sp
+
+- General information
+  - Pretrained model: N.A
+  - Training config: conf/train_ebranchformer-mega_transducer.yaml
+  - Decoding config: conf/decode.yaml
+  - GPU: Nvidia A100 40Gb
+  - CPU: AMD EPYC 7502P 32c
+  - Peak VRAM usage during training: 37.39 Gb
+  - Training time: ~ 48,9 hours
+  - Decoding time (32 jobs, 1 thread): N.A
+
+- Environments
+  - date: Tue Jun 06 05:30:22 UTC 2023`
+  - python version: `3.8.10 (default, Nov 14 2022, 12:59:47)  [GCC 9.4.0]`
+  - espnet version: `espnet 202301`
+  - pytorch version: `pytorch 1.8.1+cu111`
+  - Git hash: `6048cbb8c93019c3931070c7ab0298a2f626945d`
+  - Commit date: `Thu Feb 9 10:04:57 2023 +0000`
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|decode_asr_model_valid.loss.ave_10best/dev_clean|2703|54402|94.9|4.6|0.4|0.6|5.6|53.0|
+|decode_asr_model_valid.loss.ave_10best/dev_other|2864|50948|85.2|13.2|1.6|1.7|16.5|77.3|
+|decode_asr_model_valid.loss.ave_10best/test_clean|2620|52576|94.6|4.8|0.6|0.7|6.1|55.2|
+|decode_asr_model_valid.loss.ave_10best/test_other|2939|52343|84.9|13.3|1.7|1.7|16.7|79.4|
+
+### CER
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|decode_asr_model_valid.loss.ave_10best/dev_clean|2703|288456|98.4|0.9|0.7|0.6|2.2|53.9|
+|decode_asr_model_valid.loss.ave_10best/dev_other|2864|265951|93.9|3.7|2.4|1.9|8.0|77.3|
+|decode_asr_model_valid.loss.ave_10best/test_clean|2620|281530|98.4|0.9|0.7|0.6|2.2|55.2|
+|decode_asr_model_valid.loss.ave_10best/test_other|2939|272758|93.9|3.6|2.5|1.8|7.9|79.4|
+
+### TER
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|decode_asr_model_valid.loss.ave_10best/dev_clean|2703|107929|95.4|3.4|1.2|0.6|5.2|53.9|
+|decode_asr_model_valid.loss.ave_10best/dev_other|2864|98610|86.2|10.6|3.2|2.1|15.9|77.3|
+|decode_asr_model_valid.loss.ave_10best/test_clean|2620|105724|95.4|3.3|1.3|0.6|5.3|55.2|
+|decode_asr_model_valid.loss.ave_10best/test_other|2939|101026|86.0|10.4|3.6|2.0|16.0|79.4|
+
+
 # STREAMING SYSTEMS
 
 ## Conformer/RNN Transducer
diff --git a/egs2/librispeech_100/asr_transducer1/conf/train_ebranchformer-mega_transducer.yaml b/egs2/librispeech_100/asr_transducer1/conf/train_ebranchformer-mega_transducer.yaml
new file mode 100644
index 00000000000..9ff93a77897
--- /dev/null
+++ b/egs2/librispeech_100/asr_transducer1/conf/train_ebranchformer-mega_transducer.yaml
@@ -0,0 +1,85 @@
+# general
+batch_type: numel
+batch_bins: 4000000
+accum_grad: 8
+max_epoch: 60
+patience: none
+init: none
+num_att_plot: 0
+
+# optimizer
+optim: adam
+optim_conf:
+    lr: 0.002
+    weight_decay: 0.000001
+scheduler: warmuplr
+scheduler_conf:
+    warmup_steps: 15000
+
+# criterion
+val_scheduler_criterion:
+    - valid
+    - loss
+best_model_criterion:
+-   - valid
+    - loss
+    - min
+keep_nbest_models: 10
+
+model_conf:
+    transducer_weight: 1.0
+    auxiliary_ctc_weight: 0.3
+    report_cer: True
+    report_wer: True
+
+# specaug conf
+specaug: specaug
+specaug_conf:
+    apply_time_warp: true
+    time_warp_window: 5
+    time_warp_mode: bicubic
+    apply_freq_mask: true
+    freq_mask_width_range:
+    - 0
+    - 27
+    num_freq_mask: 2
+    apply_time_mask: true
+    time_mask_width_ratio_range:
+    - 0.
+    - 0.05
+    num_time_mask: 5
+
+encoder_conf:
+    main_conf:
+      pos_wise_act_type: swish
+      conv_mod_act_type: swish
+      pos_enc_dropout_rate: 0.2
+    input_conf:
+      vgg_like: True
+      subsampling_factor: 4
+    body_conf:
+    - block_type: ebranchformer
+      linear_size: 1024
+      hidden_size: 256
+      heads: 4
+      dropout_rate: 0.1
+      pos_wise_dropout_rate: 0.1
+      att_dropout_rate: 0.1
+      conv_mod_kernel_size: 31
+      num_blocks: 14
+decoder: mega
+decoder_conf:
+    block_size: 256
+    linear_size: 2048
+    qk_size: 128
+    v_size: 1024
+    max_positions: 1024
+    num_heads: 4
+    rel_pos_bias_type: "rotary"
+    num_blocks: 6
+    dropout_rate: 0.1
+    ffn_dropout_rate: 0.1
+    att_dropout_rate: 0.1
+    embed_dropout_rate: 0.1
+joint_network_conf:
+    joint_space_size: 256

From 9890efe0172a92be53d66ba4081816e0c40cadff Mon Sep 17 00:00:00 2001
From: "pre-commit-ci[bot]"
 <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Date: Thu, 22 Jun 2023 08:25:35 +0000
Subject: [PATCH 14/15] [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci
---
 egs2/librispeech_100/asr_transducer1/README.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/egs2/librispeech_100/asr_transducer1/README.md b/egs2/librispeech_100/asr_transducer1/README.md
index a17a686ac06..21a8ada695a 100644
--- a/egs2/librispeech_100/asr_transducer1/README.md
+++ b/egs2/librispeech_100/asr_transducer1/README.md
@@ -191,4 +191,4 @@
 |decode_asr_model_valid.loss.ave_10best/dev_clean|2703|107929|95.0|3.6|1.4|0.6|5.5|56.9|
 |decode_asr_model_valid.loss.ave_10best/dev_other|2864|98610|84.7|11.6|3.6|2.2|17.4|82.1|
 |decode_asr_model_valid.loss.ave_10best/test_clean|2620|105724|94.7|3.7|1.6|0.6|6.0|58.9|
-|decode_asr_model_valid.loss.ave_10best/test_other|2939|101026|84.3|11.6|4.1|2.0|17.7|83.5|
\ No newline at end of file
+|decode_asr_model_valid.loss.ave_10best/test_other|2939|101026|84.3|11.6|4.1|2.0|17.7|83.5|

From b541bd1c7251e1eaf961631eedf09b09c63fec08 Mon Sep 17 00:00:00 2001
From: b-flo <florian.boyer@airudit.com>
Date: Thu, 22 Jun 2023 08:28:32 +0000
Subject: [PATCH 15/15] merge some lines for readibility

---
 egs2/librispeech_100/asr_transducer1/README.md | 14 +++++---------
 1 file changed, 5 insertions(+), 9 deletions(-)

diff --git a/egs2/librispeech_100/asr_transducer1/README.md b/egs2/librispeech_100/asr_transducer1/README.md
index a17a686ac06..b3144229f14 100644
--- a/egs2/librispeech_100/asr_transducer1/README.md
+++ b/egs2/librispeech_100/asr_transducer1/README.md
@@ -1,7 +1,6 @@
 # OFFLINE SYSTEMS
 
-## Conformer/RNN Transducer
-## asr_train_conformer-rnn_transducer_raw_en_bpe500_sp
+## Conformer/RNN Transducer (asr_train_conformer-rnn_transducer_raw_en_bpe500_sp)
 
 - General information
   - Pretrained model: N.A
@@ -48,8 +47,7 @@
 |decode_asr_model_valid.loss.ave_10best/test_clean|2620|105724|95.5|3.2|1.3|0.6|5.1|54.8|
 |decode_asr_model_valid.loss.ave_10best/test_other|2939|101026|86.0|10.4|3.5|1.9|15.9|80.1|
 
-## E-Branchformer/RNN Transducer
-## asr_train_ebranchformer-rnn_transducer_raw_en_bpe500_sp
+## E-Branchformer/RNN Transducer (asr_train_ebranchformer-rnn_transducer_raw_en_bpe500_sp)
 
 - General information
   - Pretrained model: N.A
@@ -96,8 +94,7 @@
 |decode_asr_model_valid.loss.ave_10best/test_clean|2620|105724|95.4|3.2|1.3|0.6|5.1|55.5|
 |decode_asr_model_valid.loss.ave_10best/test_other|2939|101026|85.9|10.5|3.6|2.0|16.0|80.6|
 
-## E-Branchformer/MEGA Transducer
-## asr_train_ebranchformer-mega_transducer_raw_en_bpe500_sp
+## E-Branchformer/MEGA Transducer (asr_train_ebranchformer-mega_transducer_raw_en_bpe500_sp)
 
 - General information
   - Pretrained model: N.A
@@ -145,8 +142,7 @@
 
 # STREAMING SYSTEMS
 
-## Conformer/RNN Transducer
-## asr_train_conformer-rnn_transducer_streaming_raw_en_bpe500_sp
+## Conformer/RNN Transducer (asr_train_conformer-rnn_transducer_streaming_raw_en_bpe500_sp)
 
 - General information
   - Pretrained model: N.A
@@ -191,4 +187,4 @@
 |decode_asr_model_valid.loss.ave_10best/dev_clean|2703|107929|95.0|3.6|1.4|0.6|5.5|56.9|
 |decode_asr_model_valid.loss.ave_10best/dev_other|2864|98610|84.7|11.6|3.6|2.2|17.4|82.1|
 |decode_asr_model_valid.loss.ave_10best/test_clean|2620|105724|94.7|3.7|1.6|0.6|6.0|58.9|
-|decode_asr_model_valid.loss.ave_10best/test_other|2939|101026|84.3|11.6|4.1|2.0|17.7|83.5|
\ No newline at end of file
+|decode_asr_model_valid.loss.ave_10best/test_other|2939|101026|84.3|11.6|4.1|2.0|17.7|83.5|