Merge pull request #5350 from akreal/asr2-librispeech-100

Add discrete-token ASR for LibriSpeech 100h

egs2/librispeech_100/asr2/README.md

@@ -0,0 +1,56 @@
+# ASR2 recipe for LibriSpeech100
+## Related work:
+   Chang, Xuankai, et al. "Exploration of Efficient End-to-End ASR using Discretized Input from Self-Supervised Learning." InterSpeech 2023.
+   <details>
+   <summary>bib info</summary>
+
+   ```
+   @article{chang2023exploration,
+        title={Exploration of Efficient End-to-End ASR using Discretized Input from Self-Supervised Learning},
+        author={Chang, Xuankai and Yan, Brian and Fujita, Yuya and Maekaku, Takashi and Watanabe, Shinji},
+        journal={arXiv preprint arXiv:2305.18108},
+        year={2023}
+   }
+   ```
+   </details>
+
+# E-Branchformer ASR2 Discrete tokens with WavLM_large_Layer21_Kmeans2000_nBPE5000 (7 hours for 70epochs with A6000 x 1)
+
+## Environments
+- date: `Sun Jul 23 11:19:30 CEST 2023`
+- python version: `3.10.8 (main, Nov 14 2022, 00:00:00) [GCC 11.3.1 20220421 (Red Hat 11.3.1-3)]`
+- espnet version: `espnet 202304`
+- pytorch version: `pytorch 1.13.1+cu117`
+- Git hash: `64a1cabc6e7fe4fd22d46b788cec29ba6a37801e`
+  - Commit date: `Sat Jul 22 23:17:44 2023 +0200`
+  - ASR config: [conf/train_discrete_asr_e_branchformer1_1gpu.yaml](conf/train_discrete_asr_e_branchformer1_1gpu.yaml)
+  - Decode config: [conf/decode_ctc0.3.yaml](conf/decode_ctc0.3.yaml)
+  - Pretrained model: [https://huggingface.co/espnet/akreal_ls100_asr2_e_branchformer1_1gpu_raw_wavlm_large_21_km2k_bpe_rm6k_bpe_ts5k_sp](https://huggingface.co/espnet/akreal_ls100_asr2_e_branchformer1_1gpu_raw_wavlm_large_21_km2k_bpe_rm6k_bpe_ts5k_sp)
+
+## exp/asr_train_discrete_asr_e_branchformer1_1gpu_raw_wavlm_large_21_km2000_bpe_rm6000_bpe_ts5000_sp
+### WER
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|decode_ctc0.3_asr_model_valid.acc.ave/dev_clean|2703|54402|96.5|3.3|0.2|0.4|3.8|42.5|
+|decode_ctc0.3_asr_model_valid.acc.ave/dev_other|2864|50948|93.9|5.7|0.4|0.5|6.7|54.4|
+|decode_ctc0.3_asr_model_valid.acc.ave/test_clean|2620|52576|96.5|3.3|0.2|0.4|3.9|43.2|
+|decode_ctc0.3_asr_model_valid.acc.ave/test_other|2939|52343|93.6|5.9|0.5|0.6|7.0|56.2|
+
+### CER
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|decode_ctc0.3_asr_model_valid.acc.ave/dev_clean|2703|288456|99.0|0.6|0.4|0.4|1.4|42.5|
+|decode_ctc0.3_asr_model_valid.acc.ave/dev_other|2864|265951|97.8|1.3|0.9|0.7|2.9|54.4|
+|decode_ctc0.3_asr_model_valid.acc.ave/test_clean|2620|281530|99.0|0.6|0.5|0.4|1.4|43.2|
+|decode_ctc0.3_asr_model_valid.acc.ave/test_other|2939|272758|97.9|1.2|0.9|0.8|2.8|56.2|
+
+### TER
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|decode_ctc0.3_asr_model_valid.acc.ave/dev_clean|2703|69558|94.8|3.2|2.0|0.5|5.6|42.5|
+|decode_ctc0.3_asr_model_valid.acc.ave/dev_other|2864|64524|91.3|5.7|3.0|1.2|9.9|54.4|
+|decode_ctc0.3_asr_model_valid.acc.ave/test_clean|2620|66983|94.8|3.2|2.0|0.5|5.7|43.2|
+|decode_ctc0.3_asr_model_valid.acc.ave/test_other|2939|66650|91.4|5.7|2.9|1.2|9.7|56.2|

egs2/librispeech_100/asr2/asr2.sh

@@ -0,0 +1 @@
+../../TEMPLATE/asr2/asr2.sh

egs2/librispeech_100/asr2/cmd.sh

@@ -0,0 +1,110 @@
+# ====== About run.pl, queue.pl, slurm.pl, and ssh.pl ======
+# Usage: <cmd>.pl [options] JOB=1:<nj> <log> <command...>
+# e.g.
+#   run.pl --mem 4G JOB=1:10 echo.JOB.log echo JOB
+#
+# Options:
+#   --time <time>: Limit the maximum time to execute.
+#   --mem <mem>: Limit the maximum memory usage.
+#   -–max-jobs-run <njob>: Limit the number parallel jobs. This is ignored for non-array jobs.
+#   --num-threads <ngpu>: Specify the number of CPU core.
+#   --gpu <ngpu>: Specify the number of GPU devices.
+#   --config: Change the configuration file from default.
+#
+# "JOB=1:10" is used for "array jobs" and it can control the number of parallel jobs.
+# The left string of "=", i.e. "JOB", is replaced by <N>(Nth job) in the command and the log file name,
+# e.g. "echo JOB" is changed to "echo 3" for the 3rd job and "echo 8" for 8th job respectively.
+# Note that the number must start with a positive number, so you can't use "JOB=0:10" for example.
+#
+# run.pl, queue.pl, slurm.pl, and ssh.pl have unified interface, not depending on its backend.
+# These options are mapping to specific options for each backend and
+# it is configured by "conf/queue.conf" and "conf/slurm.conf" by default.
+# If jobs failed, your configuration might be wrong for your environment.
+#
+#
+# The official documentation for run.pl, queue.pl, slurm.pl, and ssh.pl:
+#   "Parallelization in Kaldi": http://kaldi-asr.org/doc/queue.html
+# =========================================================~
+
+
+# Select the backend used by run.sh from "local", "stdout", "sge", "slurm", or "ssh"
+cmd_backend='local'
+
+# Local machine, without any Job scheduling system
+if [ "${cmd_backend}" = local ]; then
+
+    # The other usage
+    export train_cmd="run.pl"
+    # Used for "*_train.py": "--gpu" is appended optionally by run.sh
+    export cuda_cmd="run.pl"
+    # Used for "*_recog.py"
+    export decode_cmd="run.pl"
+
+# Local machine logging to stdout and log file, without any Job scheduling system
+elif [ "${cmd_backend}" = stdout ]; then
+
+    # The other usage
+    export train_cmd="stdout.pl"
+    # Used for "*_train.py": "--gpu" is appended optionally by run.sh
+    export cuda_cmd="stdout.pl"
+    # Used for "*_recog.py"
+    export decode_cmd="stdout.pl"
+
+
+# "qsub" (Sun Grid Engine, or derivation of it)
+elif [ "${cmd_backend}" = sge ]; then
+    # The default setting is written in conf/queue.conf.
+    # You must change "-q g.q" for the "queue" for your environment.
+    # To know the "queue" names, type "qhost -q"
+    # Note that to use "--gpu *", you have to setup "complex_value" for the system scheduler.
+
+    export train_cmd="queue.pl"
+    export cuda_cmd="queue.pl"
+    export decode_cmd="queue.pl"
+
+
+# "qsub" (Torque/PBS.)
+elif [ "${cmd_backend}" = pbs ]; then
+    # The default setting is written in conf/pbs.conf.
+
+    export train_cmd="pbs.pl"
+    export cuda_cmd="pbs.pl"
+    export decode_cmd="pbs.pl"
+
+
+# "sbatch" (Slurm)
+elif [ "${cmd_backend}" = slurm ]; then
+    # The default setting is written in conf/slurm.conf.
+    # You must change "-p cpu" and "-p gpu" for the "partition" for your environment.
+    # To know the "partion" names, type "sinfo".
+    # You can use "--gpu * " by default for slurm and it is interpreted as "--gres gpu:*"
+    # The devices are allocated exclusively using "${CUDA_VISIBLE_DEVICES}".
+
+    export train_cmd="slurm.pl"
+    export cuda_cmd="slurm.pl"
+    export decode_cmd="slurm.pl"
+
+elif [ "${cmd_backend}" = ssh ]; then
+    # You have to create ".queue/machines" to specify the host to execute jobs.
+    # e.g. .queue/machines
+    #   host1
+    #   host2
+    #   host3
+    # Assuming you can login them without any password, i.e. You have to set ssh keys.
+
+    export train_cmd="ssh.pl"
+    export cuda_cmd="ssh.pl"
+    export decode_cmd="ssh.pl"
+
+# This is an example of specifying several unique options in the JHU CLSP cluster setup.
+# Users can modify/add their own command options according to their cluster environments.
+elif [ "${cmd_backend}" = jhu ]; then
+
+    export train_cmd="queue.pl --mem 2G"
+    export cuda_cmd="queue-freegpu.pl --mem 2G --gpu 1 --config conf/queue.conf"
+    export decode_cmd="queue.pl --mem 4G"
+
+else
+    echo "$0: Error: Unknown cmd_backend=${cmd_backend}" 1>&2
+    return 1
+fi

egs2/librispeech_100/asr2/conf/decode_ctc0.3.yaml

@@ -0,0 +1,6 @@
+beam_size: 20
+ctc_weight: 0.3
+lm_weight: 0.0
+maxlenratio: 1.0
+minlenratio: 0.0
+penalty: 0.0

egs2/librispeech_100/asr2/conf/pbs.conf

@@ -0,0 +1,11 @@
+# Default configuration
+command qsub -V -v PATH -S /bin/bash
+option name=* -N $0
+option mem=* -l mem=$0
+option mem=0          # Do not add anything to qsub_opts
+option num_threads=* -l ncpus=$0
+option num_threads=1  # Do not add anything to qsub_opts
+option num_nodes=* -l nodes=$0:ppn=1
+default gpu=0
+option gpu=0
+option gpu=* -l ngpus=$0

egs2/librispeech_100/asr2/conf/queue.conf

@@ -0,0 +1,12 @@
+# Default configuration
+command qsub -v PATH -cwd -S /bin/bash -j y -l arch=*64*
+option name=* -N $0
+option mem=* -l mem_free=$0,ram_free=$0
+option mem=0          # Do not add anything to qsub_opts
+option num_threads=* -pe smp $0
+option num_threads=1  # Do not add anything to qsub_opts
+option max_jobs_run=* -tc $0
+option num_nodes=* -pe mpi $0  # You must set this PE as allocation_rule=1
+default gpu=0
+option gpu=0
+option gpu=* -l gpu=$0 -q g.q

egs2/librispeech_100/asr2/conf/slurm.conf

@@ -0,0 +1,14 @@
+# Default configuration
+command sbatch --export=PATH
+option name=* --job-name $0
+option time=* --time $0
+option mem=* --mem-per-cpu $0
+option mem=0
+option num_threads=* --cpus-per-task $0
+option num_threads=1 --cpus-per-task 1
+option num_nodes=* --nodes $0
+default gpu=0
+option gpu=0 -p cpu
+option gpu=* -p gpu --gres=gpu:$0 -c $0  # Recommend allocating more CPU than, or equal to the number of GPU
+# note: the --max-jobs-run option is supported as a special case
+# by slurm.pl and you don't have to handle it in the config file.

egs2/librispeech_100/asr2/conf/train_discrete_asr_e_branchformer1_1gpu.yaml

@@ -0,0 +1,93 @@
+# Trained with A6000 (48 GB) x 1 GPUs for Kmeans2K+nbpe5K. It takes 6 minutes per epoch.
+# BPE-Dropout (https://github.com/google/sentencepiece#subword-regularization-and-bpe-dropout)
+src_tokenizer_encode_conf:
+    enable_sampling: true    # If set to true, bpe-dropout is used.
+    alpha: 0.4
+    nbest_size: -1
+
+frontend: embed     # embedding + positional encoding
+frontend_conf:
+    embed_dim: 512
+    positional_dropout_rate: 0.1
+
+specaug: specaug
+specaug_conf:
+    apply_time_warp: false
+    time_warp_window: 5
+    time_warp_mode: bicubic
+    apply_freq_mask: false
+    freq_mask_width_range:
+    - 0
+    - 10
+    num_freq_mask: 0
+    apply_time_mask: true
+    time_mask_width_ratio_range:
+    - 0.
+    - 0.05
+    num_time_mask: 10
+
+encoder: e_branchformer
+encoder_conf:
+    output_size: 256
+    attention_heads: 4
+    attention_layer_type: rel_selfattn
+    pos_enc_layer_type: rel_pos
+    rel_pos_type: latest
+    cgmlp_linear_units: 1024
+    cgmlp_conv_kernel: 31
+    use_linear_after_conv: false
+    gate_activation: identity
+    num_blocks: 12
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.1
+    input_layer: conv1d2
+    layer_drop_rate: 0.0
+    linear_units: 1024
+    positionwise_layer_type: linear
+    use_ffn: true
+    macaron_ffn: true
+    merge_conv_kernel: 31
+
+decoder: transformer
+decoder_conf:
+    attention_heads: 4
+    linear_units: 2048
+    num_blocks: 6
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.1
+    src_attention_dropout_rate: 0.1
+    layer_drop_rate: 0.0
+
+model: discrete_asr
+model_conf:
+    ctc_weight: 0.3
+    lsm_weight: 0.1
+    length_normalized_loss: false
+    share_decoder_input_output_embed: false
+    share_encoder_decoder_input_embed: false
+
+use_amp: true
+num_att_plot: 1
+log_interval: 1000
+num_workers: 4
+batch_type: numel
+batch_bins: 280000000
+accum_grad: 2
+max_epoch: 70
+patience: none
+init: none
+best_model_criterion:
+-   - valid
+    - acc
+    - max
+keep_nbest_models: 10
+
+optim: adam
+optim_conf:
+    lr: 0.006
+    weight_decay: 0.000001
+scheduler: warmuplr
+scheduler_conf:
+    warmup_steps: 15000

egs2/librispeech_100/asr2/db.sh

@@ -0,0 +1 @@
+../../TEMPLATE/asr2/db.sh

egs2/librispeech_100/asr2/local/data.sh

@@ -0,0 +1 @@
+../../asr1/local/data.sh

egs2/librispeech_100/asr2/local/data_prep.sh

@@ -0,0 +1 @@
+../../asr1/local/data_prep.sh

egs2/librispeech_100/asr2/local/download_and_untar.sh

@@ -0,0 +1 @@
+../../asr1/local/download_and_untar.sh

egs2/librispeech_100/asr2/path.sh

@@ -0,0 +1 @@
+../../TEMPLATE/asr2/path.sh

egs2/librispeech_100/asr2/pyscripts

@@ -0,0 +1 @@
+../../TEMPLATE/asr2/pyscripts

egs2/librispeech_100/asr2/run.sh

@@ -0,0 +1,52 @@
+#!/usr/bin/env bash
+# Set bash to 'debug' mode, it will exit on :
+# -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands',
+set -e
+set -u
+set -o pipefail
+
+
+kmeans_feature="wavlm_large/21"  # use model_type/layer_index
+nclusters=2000
+
+src_lang=$(echo "${kmeans_feature}_km${nclusters}" | tr "/" "_")
+tgt_lang=en
+
+train_set="train_clean_100"
+train_dev="dev"
+test_sets="test_clean test_other dev_clean dev_other"
+
+asr_config=conf/train_discrete_asr_e_branchformer1_1gpu.yaml
+inference_config=conf/decode_ctc0.3.yaml
+
+src_nbpe=6000   # I use src_nbpe=6000 for 2000-cluster kmeans.
+tgt_nbpe=5000   # if token_joint is True, then only tgt_nbpe is used
+
+# ts: true sequence
+# rm: deduplicated sequence which removes duplicated tokens
+src_case="rm"
+tgt_case="ts"
+
+./asr2.sh \
+    --kmeans_opts "--batch_bins 4800000 --nj 4" \
+    --kmeans_feature "${kmeans_feature}" \
+    --nclusters "${nclusters}" \
+    --ngpu 1 \
+    --src_lang ${src_lang} \
+    --tgt_lang ${tgt_lang} \
+    --src_token_type "bpe" \
+    --src_nbpe $src_nbpe \
+    --tgt_token_type "bpe" \
+    --tgt_nbpe $tgt_nbpe \
+    --src_case ${src_case} \
+    --tgt_case ${tgt_case} \
+    --speed_perturb_factors "0.9 1.0 1.1" \
+    --use_lm false \
+    --asr_config "${asr_config}" \
+    --inference_config "${inference_config}" \
+    --train_set "${train_set}" \
+    --valid_set "${train_dev}" \
+    --test_sets "${test_sets}" \
+    --src_bpe_train_text "data/${train_set}/text.${src_case}.${src_lang}" \
+    --tgt_bpe_train_text "data/${train_set}/text.${tgt_case}.${tgt_lang}" \
+    --lm_train_text "data/${train_set}/text.${tgt_case}.${tgt_lang}" "$@"

egs2/librispeech_100/asr2/scripts

@@ -0,0 +1 @@
+../../TEMPLATE/asr2/scripts

egs2/librispeech_100/asr2/steps

@@ -0,0 +1 @@
+../../../tools/kaldi/egs/wsj/s5/steps

egs2/librispeech_100/asr2/utils

@@ -0,0 +1 @@
+../../../tools/kaldi/egs/wsj/s5/utils