Merge pull request #4642 from massabaali7/master

qasr update

egs2/TEMPLATE/asr1/db.sh

@@ -68,6 +68,7 @@ JSUT=downloads
 JTUBESPEECH=downloads
 JVS=downloads
 KSS=
+QASR_TTS=downloads
 SNIPS= # smart-light-en-closed-field data path
 SPGISPEECH=
 SWBD=

egs2/qasr_tts/tts1/README.MD

@@ -0,0 +1,19 @@
+# QASR-TTS RECIPE
+
+- Our goal is to build a TTS character based system using semi-supervised data selection in a low-resource scenario by proposing two different methodologies:
+the first one by training non-autoregressive (non-AR) model from scratch on very small amount of data; and the second one by training
+autoregressive (AR) model finetuned on top of a pre-trained model.
+
+- Step 1: Prepare the data 
+
+- Step 2: Download a pretrained model 
+
+- Step 3: Replace token list with the pretrained model's one
+
+- Step 4: Finetune the pre-trained model on our 1 hour dataset and we excluded the embedding layer since we are finetuning on a different language
+
+- Step 5: Use the finetuned model as a teacher model to train the Non-AR model FastSpeech2 
+
+- Step 6: Train Parallel Wav GAN model to produce better wav samples 
+
+Go to this link to download the dataset: https://arabicspeech.org/qasr_tts

egs2/qasr_tts/tts1/RESULTS.MD

@@ -0,0 +1,32 @@
+- Environments
+
+  - python version: `3.8.12`
+
+  - espnet version: `espnet 0.10.7a1`
+
+  - chainer version: `chainer 6.0.0`
+
+  - pytorch version: `pytorch 1.10.0`
+
+
+
+- Model files 
+
+    - training config file (teacher model): `./conf/tuning/finetune_transformer.yaml`
+
+    - training config file: `./conf/tuning/train_conformer_fastspeech2.yaml`
+
+
+
+- Results: {using our pretrained ASR model on MGB2 data} here is the recipe "egs/mgb2"
+
+  - FastSpeech2 (Trans.) w/ PWG: (R = 1) 
+
+      - CER: 3.9
+
+      - WER: 9.13
+
+      - MOS
+
+        - intelligibility: 4.4 ± 0.06
+        - naturalness: 4.2 ± 0.06

egs2/qasr_tts/tts1/adapt.sh

@@ -0,0 +1,32 @@
+# Copyright 2021 Massa Baali
+
+pretrained_model=$1
+# From data preparation to statistics calculation
+./run.sh --stage 1 --stop_stage 5
+
+# download pretrained model kan-bayashi/ljspeech_tts_train_transformer_raw_char_tacotron_train.loss.ave
+wget ${pretrained}
+
+# Replace token list with pretrained model's one
+pyscripts/utils/make_token_list_from_config.py pretrained_model_path/exp/ljspeech_tts_train_transformer_raw_char_tacotron/config.yaml
+# tokens.txt is created in model directory
+mv dump/token_list/ljspeech_tts_train_transformer_raw_char_tacotron/tokens.{txt,txt.bak}
+ln -s pretrained_dir/exp/ljspeech_tts_train_transformer_raw_char_tacotron/tokens.txt dump/token_list/new_model
+
+# Train the model 
+./run.sh --stage 6 --train_config conf/tuning/finetune_transformer.yaml --train_args \ 
+"--init_param pretrained_model/exp/tts_train_transformer_raw_char_tacotron/train.loss.ave_5best.pth":::tts.enc.embed \
+--tag finetune_pretrained_transformers
+
+# Now the trained model above will be used as a teacher model for the Non-AR model FastSpeech2 
+# Prepare durations file
+./run.sh --stage 7  --tts_exp exp/tts_finetune_pretrained_transformers \
+    --inference_args "--use_teacher_forcing true" \
+    --test_sets "tr_no_dev dev eval1"
+
+# Since fastspeech2 requires extra feature calculation, run from stage 5.
+./run.sh --stage 5 \
+    --train_config conf/tuning/train_conformer_fastspeech2.yaml \
+    --teacher_dumpdir exp/tts_finetune_pretrained_transformers/decode_use_teacher_forcingtrue_train.loss.ave \
+    --tts_stats_dir exp/tts_finetune_pretrained_transformers/decode_use_teacher_forcingtrue_train.loss.ave/stats \
+    --write_collected_feats true

egs2/qasr_tts/tts1/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/qasr_tts/tts1/conf/decode.yaml

@@ -0,0 +1 @@
+tuning/decode_tacotron2.yaml

egs2/qasr_tts/tts1/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/qasr_tts/tts1/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/qasr_tts/tts1/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/qasr_tts/tts1/conf/train.yaml

@@ -0,0 +1 @@
+tuning/finetune_transformer.yaml

egs2/qasr_tts/tts1/conf/tuning/decode_fastspeech.yaml

@@ -0,0 +1,10 @@
+# This configuration is the decoding setting for FastSpeech or FastSpeech2.
+
+##########################################################
+#                    DECODING SETTING                    #
+##########################################################
+speed_control_alpha: 1     # alpha to control the speed of generated speech
+                           # 1 < alpha makes slower and 1 > alpha makes faster
+use_teacher_forcing: false # whether to use teacher forcing
+                           # if true, we use groundtruth of durations
+                           # (+ pitch & energy for FastSpeech2)

egs2/qasr_tts/tts1/conf/tuning/decode_tacotron2.yaml

@@ -0,0 +1,16 @@
+# This configuration is the basic decoding setting for Tacotron 2.
+# It can be also applied to Transformer. If you met some problems
+# such as deletions or repetitions, it is worthwhile to try
+# `use_att_constraint: true` to make the generation more stable.
+# Note that attention constraint is not supported in Transformer.
+
+##########################################################
+#                    DECODING SETTING                    #
+##########################################################
+threshold: 0.5             # threshold to stop the generation
+maxlenratio: 10.0          # maximum length of generated samples = input length * maxlenratio
+minlenratio: 0.0           # minimum length of generated samples = input length * minlenratio
+use_att_constraint: true   # whether to use attention constraint, which is introduced in deep voice 3
+backward_window: 1         # backward window size in the attention constraint
+forward_window: 3          # forward window size in the attention constraint
+use_teacher_forcing: false # whether to use teacher forcing

egs2/qasr_tts/tts1/conf/tuning/finetune_tacotron2.yaml

@@ -0,0 +1,78 @@
+# This configuration is for ESPnet2 to finetune Tacotron 2. Compared to the
+# original paper, this configuration additionally use the guided attention
+# loss to accelerate the learning of the diagonal attention. It requires
+# only a single GPU with 12 GB memory and it takes ~1 days to finish the
+# training on Titan V.
+
+##########################################################
+#                  TTS MODEL SETTING                     #
+##########################################################
+tts: tacotron2                   # model architecture
+tts_conf:                        # keyword arguments for the selected model
+    embed_dim: 512               # char or phn embedding dimension
+    elayers: 1                   # number of blstm layers in encoder
+    eunits: 512                  # number of blstm units
+    econv_layers: 3              # number of convolutional layers in encoder
+    econv_chans: 512             # number of channels in convolutional layer
+    econv_filts: 5               # filter size of convolutional layer
+    atype: location              # attention function type
+    adim: 512                    # attention dimension
+    aconv_chans: 32              # number of channels in convolutional layer of attention
+    aconv_filts: 15              # filter size of convolutional layer of attention
+    cumulate_att_w: true         # whether to cumulate attention weight
+    dlayers: 2                   # number of lstm layers in decoder
+    dunits: 1024                 # number of lstm units in decoder
+    prenet_layers: 2             # number of layers in prenet
+    prenet_units: 256            # number of units in prenet
+    postnet_layers: 5            # number of layers in postnet
+    postnet_chans: 512           # number of channels in postnet
+    postnet_filts: 5             # filter size of postnet layer
+    output_activation: null      # activation function for the final output
+    use_batch_norm: true         # whether to use batch normalization in encoder
+    use_concate: true            # whether to concatenate encoder embedding with decoder outputs
+    use_residual: false          # whether to use residual connection in encoder
+    dropout_rate: 0.5            # dropout rate
+    zoneout_rate: 0.1            # zoneout rate
+    reduction_factor: 1          # reduction factor
+    spk_embed_dim: null          # speaker embedding dimension
+    use_masking: true            # whether to apply masking for padded part in loss calculation
+    bce_pos_weight: 5.0          # weight of positive sample in binary cross entropy calculation
+    use_guided_attn_loss: true   # whether to use guided attention loss
+    guided_attn_loss_sigma: 0.4  # sigma of guided attention loss
+    guided_attn_loss_lambda: 1.0 # strength of guided attention loss
+
+##########################################################
+#                  OPTIMIZER SETTING                     #
+##########################################################
+optim: adam           # optimizer type
+optim_conf:           # keyword arguments for selected optimizer
+    lr: 1.0e-04       # learning rate
+    eps: 1.0e-06      # epsilon
+    weight_decay: 0.0 # weight decay coefficient
+
+##########################################################
+#                OTHER TRAINING SETTING                  #
+##########################################################
+num_iters_per_epoch: 200    # number of iters per epoch
+max_epoch: 100              # number of epochs
+grad_clip: 1.0              # gradient clipping norm
+grad_noise: false           # whether to use gradient noise injection
+accum_grad: 1               # gradient accumulation
+# batch_bins: 1000000       # batch bins (for feats_type=fbank)
+batch_bins: 3750000         # batch bins (for feats_type=raw, *= n_shift / n_mels)
+batch_type: numel           # how to make batch
+sort_in_batch: descending   # how to sort data in making batch
+sort_batch: descending      # how to sort created batches
+num_workers: 1              # number of workers of data loader
+train_dtype: float32        # dtype in training
+log_interval: null          # log interval in iterations
+keep_nbest_models: 5        # number of models to keep
+num_att_plot: 3             # number of attention figures to be saved in every check
+seed: 0                     # random seed number
+best_model_criterion:
+-   - valid
+    - loss
+    - min
+-   - train
+    - loss
+    - min