first commit

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+*.arpa filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,3 +1,21 @@
----
-license: apache-2.0
----
+# Tunisian Arabic ASR Model with wav2vec2
+
+This repository provides all the necessary tools to perform automatic speech recognition from an end-to-end system pretrained on Tunisian arabic dialect
+
+## Performance
+the performance of the mode is : 
+| Release Version | |WER (%) | CER (%) |
+|-----------------|----|---------|---------|
+| v1.0            | Without LM |11.82    | 6.33   |
+## Dataset
+This ASR model was trained on :
+* TARIC : The corpus, named TARIC (Tunisian Arabic Railway Interaction Corpus) has a collection of audio recordings and transcriptions from dialogues in the Tunisian Railway Transport Network. - [Taric Corpus](https://aclanthology.org/L14-1385/) -
+* STAC :A corpus of spoken Tunisian Arabic - [STAC Corpus](https://www.researchgate.net/publication/307583782_Spoken_Tunisian_Arabic_Corpus_STAC_Transcription_and_Annotation)
+* IWSLT : A Tunisian conversational speech - [IWSLT Corpus](https://iwslt.org/2022/dialect)-
+* Tunspeech : Our custom dataset
+
+## Install
+```python
+pip install speechbrain transformers
+```
+

outdomain.arpa

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:24654c1d236bb1bd367125131c847c4a734e69914eda71a6786964c20440d8fe
+size 324243244

semi_wavlm_large_tunisian_ctc/1234/hyperparams.yaml

@@ -0,0 +1,194 @@
+# Generated 2023-09-08 from:
+# /gpfsdsstore/projects/rech/nou/uzn19yk/switched_code_tunisian/train/tunisian_asr/hparams/train_semi.yaml
+# yamllint disable
+# ################################
+# Model: wav2vec2 + DNN + CTC
+# Augmentation: SpecAugment
+# Authors: Titouan Parcollet 2021
+# ################################
+
+# Seed needs to be set at top of yaml, before objects with parameters are made
+seed: 1234
+__set_seed: !!python/object/apply:torch.manual_seed [1234]
+output_folder: results/semi_wavlm_large_tunisian_ctc/1234
+wer_file: results/semi_wavlm_large_tunisian_ctc/1234/wer.txt
+save_folder: results/semi_wavlm_large_tunisian_ctc/1234/save
+train_log: results/semi_wavlm_large_tunisian_ctc/1234/train_log.txt
+
+# URL for the biggest LeBenchmark wav2vec french.
+wav2vec2_folder: results/semi_wavlm_large_tunisian_ctc/1234/save/wav2vec2_checkpoint
+
+# Data files
+data_folder: /gpfsscratch/rech/nou/uzn19yk/tunisian_junk  # e.g, /localscratch/cv-corpus-5.1-2020-06-22/fr
+train_tsv_file: /gpfsscratch/rech/nou/uzn19yk/tunisian_junk/train.tsv # Standard CommonVoice .tsv files
+dev_tsv_file: /gpfsscratch/rech/nou/uzn19yk/tunisian_junk/dev.tsv # Standard CommonVoice .tsv files
+test_tsv_file: /gpfsscratch/rech/nou/uzn19yk/tunisian_junk/test.tsv # Standard CommonVoice .tsv files
+accented_letters: true
+language: fr # use 'it' for Italian, 'rw' for Kinyarwanda, 'en' for english
+train_csv: /gpfsscratch/rech/nou/uzn19yk/tunisian_csvs/good_final/train_enhanced.csv
+valid_csv: /gpfsscratch/rech/nou/uzn19yk/tunisian_csvs/good_final/dev.csv
+test_csv:
+- /gpfsscratch/rech/nou/uzn19yk/tunisian_csvs/full_annotation_test.csv
+- /gpfsscratch/rech/nou/uzn19yk/tunisian_csvs/good_final/iwslt_test.csv
+- /gpfsscratch/rech/nou/uzn19yk/tunisian_csvs/good_final/taric_test.csv
+
+skip_prep: true # Skip data preparation
+
+use_language_modelling: true
+ngram_lm_path: arpas/outdomain.arpa
+
+# We remove utterance slonger than 10s in the train/dev/test sets as
+# longer sentences certainly correspond to "open microphones".
+avoid_if_longer_than: 10.0
+avoid_if_shorter_than: 1.2
+
+
+# Training parameters
+number_of_epochs: 12
+lr: 1.0
+lr_wav2vec: 0.0001
+sorting: ascending
+auto_mix_prec: false
+sample_rate: 16000
+ckpt_interval_minutes: 30 # save checkpoint every N min
+
+# With data_parallel batch_size is split into N jobs
+# With DDP batch_size is multiplied by N jobs
+# Must be 6 per GPU to fit 16GB of VRAM
+batch_size: 10
+test_batch_size: 4
+
+dataloader_options:
+  batch_size: 10
+  num_workers: 6
+test_dataloader_options:
+  batch_size: 4
+  num_workers: 6
+
+# BPE parameters
+token_type: char  # ["unigram", "bpe", "char"]
+character_coverage: 1.0
+
+# Model parameters
+# activation: !name:torch.nn.LeakyReLU
+wav2vec_output_dim: 1024
+dnn_neurons: 1024
+freeze_wav2vec: false
+freeze_feature_extractor: true
+dropout: 0.15
+warmup_steps: 500 # The wav2vec 2 model isn't updated for this amount of steps
+
+# Outputs
+output_neurons: 40  # BPE size, index(blank/eos/bos) = 0
+
+# Decoding parameters
+# Be sure that the bos and eos index match with the BPEs ones
+blank_index: 0
+unk_index: 1
+
+#
+# Functions and classes
+#
+epoch_counter: &id007 !new:speechbrain.utils.epoch_loop.EpochCounter
+
+  limit: 12
+
+augmentation: !new:speechbrain.lobes.augment.TimeDomainSpecAugment
+  sample_rate: 16000
+  speeds: [95, 100, 105]
+
+enc: &id002 !new:speechbrain.nnet.containers.Sequential
+  input_shape: [null, null, 1024]
+  linear1: !name:speechbrain.nnet.linear.Linear
+    n_neurons: 1024
+    bias: true
+  bn1: !name:speechbrain.nnet.normalization.BatchNorm1d
+  activation: !new:torch.nn.LeakyReLU
+  drop: !new:torch.nn.Dropout
+    p: 0.15
+  linear2: !name:speechbrain.nnet.linear.Linear
+    n_neurons: 1024
+    bias: true
+  bn2: !name:speechbrain.nnet.normalization.BatchNorm1d
+  activation2: !new:torch.nn.LeakyReLU
+  drop2: !new:torch.nn.Dropout
+    p: 0.15
+  linear3: !name:speechbrain.nnet.linear.Linear
+    n_neurons: 1024
+    bias: true
+  bn3: !name:speechbrain.nnet.normalization.BatchNorm1d
+  activation3: !new:torch.nn.LeakyReLU
+
+wav2vec2: &id001 !new:speechbrain.lobes.models.huggingface_wav2vec.HuggingFaceWav2Vec2
+  source: /gpfsstore/rech/nou/uzn19yk/wavlm/
+  output_norm: false
+  freeze: false
+  freeze_feature_extractor: true
+  save_path: results/semi_wavlm_large_tunisian_ctc/1234/save/wav2vec2_checkpoint
+
+#####
+# Uncomment this block if you prefer to use a Fairseq pretrained model instead
+# of a HuggingFace one. Here, we provide an URL that is obtained from the
+# Fairseq github for the multilingual XLSR.
+#
+#wav2vec2_url: https://dl.fbaipublicfiles.com/fairseq/wav2vec/xlsr_53_56k.pt
+#wav2vec2: !new:speechbrain.lobes.models.fairseq_wav2vec.FairseqWav2Vec2
+#    pretrained_path: !ref <wav2vec2_url>
+#    output_norm: True
+#    freeze: False
+#    save_path: !ref <save_folder>/wav2vec2_checkpoint/model.pt
+#####
+
+
+ctc_lin: &id003 !new:speechbrain.nnet.linear.Linear
+
+  input_size: 1024
+  n_neurons: 40
+
+log_softmax: !new:speechbrain.nnet.activations.Softmax
+  apply_log: true
+
+ctc_cost: !name:speechbrain.nnet.losses.ctc_loss
+  blank_index: 0
+
+modules:
+  wav2vec2: *id001
+  enc: *id002
+  ctc_lin: *id003
+model: &id004 !new:torch.nn.ModuleList
+- [*id002, *id003]
+model_opt_class: !name:torch.optim.Adadelta
+  lr: 1.0
+  rho: 0.95
+  eps: 1.e-8
+
+wav2vec_opt_class: !name:torch.optim.Adam
+  lr: 0.0001
+
+lr_annealing_model: &id005 !new:speechbrain.nnet.schedulers.NewBobScheduler
+  initial_value: 1.0
+  improvement_threshold: 0.0025
+  annealing_factor: 0.8
+  patient: 0
+
+lr_annealing_wav2vec: &id006 !new:speechbrain.nnet.schedulers.NewBobScheduler
+  initial_value: 0.0001
+  improvement_threshold: 0.0025
+  annealing_factor: 0.9
+  patient: 0
+
+checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
+  checkpoints_dir: results/semi_wavlm_large_tunisian_ctc/1234/save
+  recoverables:
+    wav2vec2: *id001
+    model: *id004
+    scheduler_model: *id005
+    scheduler_wav2vec: *id006
+    counter: *id007
+train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
+  save_file: results/semi_wavlm_large_tunisian_ctc/1234/train_log.txt
+
+error_rate_computer: !name:speechbrain.utils.metric_stats.ErrorRateStats
+
+cer_computer: !name:speechbrain.utils.metric_stats.ErrorRateStats
+  split_tokens: true

semi_wavlm_large_tunisian_ctc/1234/save/CKPT+2023-09-05+01-09-23+00/CKPT.yaml

@@ -0,0 +1,4 @@
+# yamllint disable
+WER: 27.83210816487267
+end-of-epoch: true
+unixtime: 1693868963.5220973

semi_wavlm_large_tunisian_ctc/1234/save/CKPT+2023-09-05+01-09-23+00/brain.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3947a24e8dff5a14299b9cf2fe66ffb4d738cb88717de7f0cf7e8547a76e9776
+size 51

semi_wavlm_large_tunisian_ctc/1234/save/CKPT+2023-09-05+01-09-23+00/counter.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6b51d431df5d7f141cbececcf79edf3dd861c3b4069f0b11661a3eefacbba918
+size 2

semi_wavlm_large_tunisian_ctc/1234/save/CKPT+2023-09-05+01-09-23+00/dataloader-TRAIN.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b363886c229e536bd3c84e0c3e89312d70e00422578e076a62df1b45c9390793
+size 5

semi_wavlm_large_tunisian_ctc/1234/save/CKPT+2023-09-05+01-09-23+00/model.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bc1dbeca1e1f1340b08d8ebea6e492f474708dddbbe8cabbcdde5ee9660704f2
+size 12814446

semi_wavlm_large_tunisian_ctc/1234/save/CKPT+2023-09-05+01-09-23+00/modelopt.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3af1791eb9a5bfbfc087d2c10b94634df24cad3ac503ce9ba280a3ecc4737781
+size 25575663

semi_wavlm_large_tunisian_ctc/1234/save/CKPT+2023-09-05+01-09-23+00/scheduler_model.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c275ab9245b440d1586f72058d9edaac1a2fb3e7a52712aa9a9ad022b99a1c0d
+size 639

semi_wavlm_large_tunisian_ctc/1234/save/CKPT+2023-09-05+01-09-23+00/scheduler_wav2vec.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a88187f7882dc3e10c108f1b7abfbd819285b34bded4e88e91c4ff699c1bb5d2
+size 643

semi_wavlm_large_tunisian_ctc/1234/save/CKPT+2023-09-05+01-09-23+00/wav2vec2.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:788267bd25ef37623715fa21a975090e5e316fff05971375cd3f62e5160f0743
+size 1262005979

semi_wavlm_large_tunisian_ctc/1234/save/CKPT+2023-09-05+01-09-23+00/wav2vec_opt.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:efa967fdd8067be7d88c18cd197980c9c91f344a3dff2b2518b8381c49f28b1e
+size 2490361859

semi_wavlm_large_tunisian_ctc/1234/save/label_encoder.txt

@@ -0,0 +1,44 @@
+'ب' => 38
+'ا' => 1
+'ه' => 2
+'ي' => 3
+'و' => 4
+'ن' => 5
+'أ' => 6
+' ' => 7
+'م' => 8
+'ش' => 9
+'ل' => 10
+'س' => 11
+'ت' => 12
+'د' => 13
+'ر' => 14
+'ى' => 15
+'ح' => 16
+'ط' => 17
+'ع' => 18
+'ك' => 19
+'ف' => 20
+'ق' => 21
+'آ' => 22
+'ة' => 23
+'ج' => 24
+'ض' => 25
+'ز' => 26
+'ص' => 27
+'إ' => 28
+'ث' => 29
+'خ' => 30
+'ڨ' => 31
+'ذ' => 32
+'ظ' => 33
+'ء' => 34
+'غ' => 35
+'ئ' => 36
+'ؤ' => 37
+'<blank>' => 0
+1 => 39
+================
+'starting_index' => 0
+'unk_label' => 1
+'blank_label' => '<blank>'

train_semi.yaml

@@ -0,0 +1,175 @@
+# ################################
+# Model: wav2vec2 + DNN + CTC
+# Augmentation: SpecAugment
+# Authors: Titouan Parcollet 2021
+# ################################
+
+# Seed needs to be set at top of yaml, before objects with parameters are made
+seed: 1234
+__set_seed: !!python/object/apply:torch.manual_seed [!ref <seed>]
+output_folder: !ref semi_wavlm_large_tunisian_ctc/<seed>
+wer_file: !ref <output_folder>/wer.txt
+save_folder: !ref <output_folder>/save
+train_log: !ref <output_folder>/train_log.txt
+
+# URL for the biggest LeBenchmark wav2vec french.
+wav2vec2_folder: !ref <save_folder>/wav2vec2_checkpoint
+
+# Data files
+data_folder: /path/to/data  # e.g, /localscratch/cv-corpus-5.1-2020-06-22/fr
+train_tsv_file: !ref <data_folder>/train.tsv  # Standard CommonVoice .tsv files
+dev_tsv_file: !ref <data_folder>/dev.tsv  # Standard CommonVoice .tsv files
+test_tsv_file: !ref <data_folder>/test.tsv  # Standard CommonVoice .tsv files
+accented_letters: True
+language: fr # use 'it' for Italian, 'rw' for Kinyarwanda, 'en' for english
+test_csv: 
+    - /path/to/test_data
+
+skip_prep: True # Skip data preparation
+
+use_language_modelling: True
+ngram_lm_path: outdomain.arpa
+
+# We remove utterance slonger than 10s in the train/dev/test sets as
+# longer sentences certainly correspond to "open microphones".
+avoid_if_longer_than: 10.0
+avoid_if_shorter_than: 1.2
+
+
+# Training parameters
+number_of_epochs: 12
+lr: 1.0
+lr_wav2vec: 0.0001
+sorting: ascending
+auto_mix_prec: False
+sample_rate: 16000
+ckpt_interval_minutes: 30 # save checkpoint every N min
+
+# With data_parallel batch_size is split into N jobs
+# With DDP batch_size is multiplied by N jobs
+# Must be 6 per GPU to fit 16GB of VRAM
+batch_size: 10
+test_batch_size: 4
+
+dataloader_options:
+    batch_size: !ref <batch_size>
+    num_workers: 6
+test_dataloader_options:
+    batch_size: !ref <test_batch_size>
+    num_workers: 6
+
+# BPE parameters
+token_type: char  # ["unigram", "bpe", "char"]
+character_coverage: 1.0
+
+# Model parameters
+# activation: !name:torch.nn.LeakyReLU
+wav2vec_output_dim: 1024
+dnn_neurons: 1024
+freeze_wav2vec: False
+freeze_feature_extractor: True
+dropout: 0.15
+warmup_steps: 500 # The wav2vec 2 model isn't updated for this amount of steps
+
+# Outputs
+output_neurons: 40  # BPE size, index(blank/eos/bos) = 0
+
+# Decoding parameters
+# Be sure that the bos and eos index match with the BPEs ones
+blank_index: 0
+unk_index: 1
+
+#
+# Functions and classes
+#
+epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
+    limit: !ref <number_of_epochs>
+
+augmentation: !new:speechbrain.lobes.augment.TimeDomainSpecAugment
+    sample_rate: !ref <sample_rate>
+    speeds: [95, 100, 105]
+
+enc: !new:speechbrain.nnet.containers.Sequential
+    input_shape: [null, null, !ref <wav2vec_output_dim>]
+    linear1: !name:speechbrain.nnet.linear.Linear
+        n_neurons: !ref <dnn_neurons>
+        bias: True
+    bn1: !name:speechbrain.nnet.normalization.BatchNorm1d
+    activation: !new:torch.nn.LeakyReLU
+    drop: !new:torch.nn.Dropout
+        p: !ref <dropout>
+    linear2: !name:speechbrain.nnet.linear.Linear
+        n_neurons: !ref <dnn_neurons>
+        bias: True
+    bn2: !name:speechbrain.nnet.normalization.BatchNorm1d
+    activation2: !new:torch.nn.LeakyReLU
+    drop2: !new:torch.nn.Dropout
+        p: !ref <dropout>
+    linear3: !name:speechbrain.nnet.linear.Linear
+        n_neurons: !ref <dnn_neurons>
+        bias: True
+    bn3: !name:speechbrain.nnet.normalization.BatchNorm1d
+    activation3: !new:torch.nn.LeakyReLU
+
+wav2vec2: !new:speechbrain.lobes.models.huggingface_wav2vec.HuggingFaceWav2Vec2
+    source: /gpfsstore/rech/nou/uzn19yk/wavlm/
+    output_norm: False
+    freeze: !ref <freeze_wav2vec>
+    freeze_feature_extractor: !ref <freeze_feature_extractor>
+    save_path: !ref <wav2vec2_folder>
+
+
+ctc_lin: !new:speechbrain.nnet.linear.Linear
+    input_size: !ref <dnn_neurons>
+    n_neurons: !ref <output_neurons>
+
+log_softmax: !new:speechbrain.nnet.activations.Softmax
+    apply_log: True
+
+ctc_cost: !name:speechbrain.nnet.losses.ctc_loss
+    blank_index: !ref <blank_index>
+
+modules:
+    wav2vec2: !ref <wav2vec2>
+    enc: !ref <enc>
+    ctc_lin: !ref <ctc_lin>
+
+model: !new:torch.nn.ModuleList
+    - [!ref <enc>, !ref <ctc_lin>]
+
+model_opt_class: !name:torch.optim.Adadelta
+    lr: !ref <lr>
+    rho: 0.95
+    eps: 1.e-8
+
+wav2vec_opt_class: !name:torch.optim.Adam
+    lr: !ref <lr_wav2vec>
+
+lr_annealing_model: !new:speechbrain.nnet.schedulers.NewBobScheduler
+    initial_value: !ref <lr>
+    improvement_threshold: 0.0025
+    annealing_factor: 0.8
+    patient: 0
+
+lr_annealing_wav2vec: !new:speechbrain.nnet.schedulers.NewBobScheduler
+    initial_value: !ref <lr_wav2vec>
+    improvement_threshold: 0.0025
+    annealing_factor: 0.9
+    patient: 0
+
+checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
+    checkpoints_dir: !ref <save_folder>
+    recoverables:
+        wav2vec2: !ref <wav2vec2>
+        model: !ref <model>
+        scheduler_model: !ref <lr_annealing_model>
+        scheduler_wav2vec: !ref <lr_annealing_wav2vec>
+        counter: !ref <epoch_counter>
+
+train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
+    save_file: !ref <train_log>
+
+error_rate_computer: !name:speechbrain.utils.metric_stats.ErrorRateStats
+
+cer_computer: !name:speechbrain.utils.metric_stats.ErrorRateStats
+    split_tokens: True

train_with_wavlm.py

@@ -0,0 +1,399 @@
+#!/usr/bin/env python3
+import sys
+import torch
+import logging
+import speechbrain as sb
+from pathlib import Path
+import os
+import torchaudio
+from hyperpyyaml import load_hyperpyyaml
+from speechbrain.tokenizers.SentencePiece import SentencePiece
+from speechbrain.utils.data_utils import undo_padding
+from speechbrain.utils.distributed import run_on_main
+
+"""Recipe for training a sequence-to-sequence ASR system with CommonVoice.
+The system employs a wav2vec2 encoder and a CTC decoder.
+Decoding is performed with greedy decoding (will be extended to beam search).
+
+To run this recipe, do the following:
+> python train_with_wav2vec2.py hparams/train_with_wav2vec2.yaml
+
+With the default hyperparameters, the system employs a pretrained wav2vec2 encoder.
+The wav2vec2 model is pretrained following the model given in the hprams file.
+It may be dependent on the language.
+
+The neural network is trained with CTC on sub-word units estimated with
+Byte Pairwise Encoding (BPE).
+
+The experiment file is flexible enough to support a large variety of
+different systems. By properly changing the parameter files, you can try
+different encoders, decoders, tokens (e.g, characters instead of BPE),
+training languages (all CommonVoice languages), and many
+other possible variations.
+
+Authors
+ * Titouan Parcollet 2021
+"""
+
+logger = logging.getLogger(__name__)
+
+
+# Define training procedure
+class ASR(sb.core.Brain):
+    def compute_forward(self, batch, stage):
+        """Forward computations from the waveform batches to the output probabilities."""
+
+        batch = batch.to(self.device)
+        wavs, wav_lens = batch.sig
+        wavs, wav_lens = wavs.to(self.device), wav_lens.to(self.device)
+        if stage == sb.Stage.TRAIN:
+            if hasattr(self.hparams, "augmentation"):
+                wavs = self.hparams.augmentation(wavs, wav_lens)
+
+        # Forward pass
+        feats = self.modules.wav2vec2(wavs, wav_lens)
+        x = self.modules.enc(feats)
+        logits = self.modules.ctc_lin(x)
+        p_ctc = self.hparams.log_softmax(logits)
+
+        return p_ctc, wav_lens
+
+    def compute_objectives(self, predictions, batch, stage):
+        """Computes the loss (CTC) given predictions and targets."""
+
+        p_ctc, wav_lens = predictions
+
+        ids = batch.id
+        tokens, tokens_lens = batch.tokens
+
+        loss = self.hparams.ctc_cost(p_ctc, tokens, wav_lens, tokens_lens)
+
+        if stage != sb.Stage.TRAIN:
+            predicted_tokens = sb.decoders.ctc_greedy_decode(
+                p_ctc, wav_lens, blank_id=self.hparams.blank_index
+            )
+            # Decode token terms to words
+            if self.hparams.use_language_modelling:
+                predicted_words = []
+                for logs in p_ctc:
+                    text = decoder.decode(logs.detach().cpu().numpy())
+                    predicted_words.append(text.split(" "))
+            else:
+                predicted_words = [
+                    "".join(self.tokenizer.decode_ndim(utt_seq)).split(" ")
+                    for utt_seq in predicted_tokens
+                ]
+            # Convert indices to words
+            target_words = [wrd.split(" ") for wrd in batch.wrd]
+
+            self.wer_metric.append(ids, predicted_words, target_words)
+            self.cer_metric.append(ids, predicted_words, target_words)
+
+        return loss
+
+    def fit_batch(self, batch):
+        """Train the parameters given a single batch in input"""
+        should_step = self.step % self.grad_accumulation_factor == 0
+        # Managing automatic mixed precision
+        # TOFIX: CTC fine-tuning currently is unstable
+        # This is certainly due to CTC being done in fp16 instead of fp32
+        if self.auto_mix_prec:
+            with torch.cuda.amp.autocast():
+                with self.no_sync():
+                    outputs = self.compute_forward(batch, sb.Stage.TRAIN)
+                loss = self.compute_objectives(outputs, batch, sb.Stage.TRAIN)
+            with self.no_sync(not should_step):
+                self.scaler.scale(
+                    loss / self.grad_accumulation_factor
+                ).backward()
+            if should_step:
+
+                if not self.hparams.wav2vec2.freeze:
+                    self.scaler.unscale_(self.wav2vec_optimizer)
+                self.scaler.unscale_(self.model_optimizer)
+                if self.check_gradients(loss):
+                    if not self.hparams.wav2vec2.freeze:
+                        if self.optimizer_step >= self.hparams.warmup_steps:
+                            self.scaler.step(self.wav2vec_optimizer)
+                    self.scaler.step(self.model_optimizer)
+                self.scaler.update()
+                self.zero_grad()
+                self.optimizer_step += 1
+        else:
+            # This is mandatory because HF models have a weird behavior with DDP
+            # on the forward pass
+            with self.no_sync():
+                outputs = self.compute_forward(batch, sb.Stage.TRAIN)
+
+            loss = self.compute_objectives(outputs, batch, sb.Stage.TRAIN)
+
+            with self.no_sync(not should_step):
+                (loss / self.grad_accumulation_factor).backward()
+            if should_step:
+                if self.check_gradients(loss):
+                    if not self.hparams.wav2vec2.freeze:
+                        if self.optimizer_step >= self.hparams.warmup_steps:
+                            self.wav2vec_optimizer.step()
+                    self.model_optimizer.step()
+                self.zero_grad()
+                self.optimizer_step += 1
+
+        self.on_fit_batch_end(batch, outputs, loss, should_step)
+        return loss.detach().cpu()
+
+    def evaluate_batch(self, batch, stage):
+        """Computations needed for validation/test batches"""
+        predictions = self.compute_forward(batch, stage=stage)
+        with torch.no_grad():
+            loss = self.compute_objectives(predictions, batch, stage=stage)
+        return loss.detach()
+
+    def on_stage_start(self, stage, epoch):
+        """Gets called at the beginning of each epoch"""
+        if stage != sb.Stage.TRAIN:
+            self.cer_metric = self.hparams.cer_computer()
+            self.wer_metric = self.hparams.error_rate_computer()
+
+    def on_stage_end(self, stage, stage_loss, epoch):
+        """Gets called at the end of an epoch."""
+        # Compute/store important stats
+        stage_stats = {"loss": stage_loss}
+        if stage == sb.Stage.TRAIN:
+            self.train_stats = stage_stats
+        else:
+            stage_stats["CER"] = self.cer_metric.summarize("error_rate")
+            stage_stats["WER"] = self.wer_metric.summarize("error_rate")
+
+        # Perform end-of-iteration things, like annealing, logging, etc.
+        if stage == sb.Stage.VALID:
+            old_lr_model, new_lr_model = self.hparams.lr_annealing_model(
+                stage_stats["loss"]
+            )
+            old_lr_wav2vec, new_lr_wav2vec = self.hparams.lr_annealing_wav2vec(
+                stage_stats["loss"]
+            )
+            sb.nnet.schedulers.update_learning_rate(
+                self.model_optimizer, new_lr_model
+            )
+            if not self.hparams.wav2vec2.freeze:
+                sb.nnet.schedulers.update_learning_rate(
+                    self.wav2vec_optimizer, new_lr_wav2vec
+                )
+            self.hparams.train_logger.log_stats(
+                stats_meta={
+                    "epoch": epoch,
+                    "lr_model": old_lr_model,
+                    "lr_wav2vec": old_lr_wav2vec,
+                },
+                train_stats=self.train_stats,
+                valid_stats=stage_stats,
+            )
+            self.checkpointer.save_and_keep_only(
+                meta={"WER": stage_stats["WER"]}, min_keys=["WER"],
+            )
+        elif stage == sb.Stage.TEST:
+            self.hparams.train_logger.log_stats(
+                stats_meta={"Epoch loaded": self.hparams.epoch_counter.current},
+                test_stats=stage_stats,
+            )
+            with open(self.hparams.wer_file, "w") as w:
+                self.wer_metric.write_stats(w)
+
+    def init_optimizers(self):
+        "Initializes the wav2vec2 optimizer and model optimizer"
+
+        # If the wav2vec encoder is unfrozen, we create the optimizer
+        if not self.hparams.wav2vec2.freeze:
+            self.wav2vec_optimizer = self.hparams.wav2vec_opt_class(
+                self.modules.wav2vec2.parameters()
+            )
+            if self.checkpointer is not None:
+                self.checkpointer.add_recoverable(
+                    "wav2vec_opt", self.wav2vec_optimizer
+                )
+
+        self.model_optimizer = self.hparams.model_opt_class(
+            self.hparams.model.parameters()
+        )
+
+        if self.checkpointer is not None:
+            self.checkpointer.add_recoverable("modelopt", self.model_optimizer)
+
+    def zero_grad(self, set_to_none=False):
+        if not self.hparams.wav2vec2.freeze:
+            self.wav2vec_optimizer.zero_grad(set_to_none)
+        self.model_optimizer.zero_grad(set_to_none)
+
+
+# Define custom data procedure
+def dataio_prepare(hparams):
+    """This function prepares the datasets to be used in the brain class.
+    It also defines the data processing pipeline through user-defined functions."""
+
+    # 1. Define datasets
+    data_folder = hparams["data_folder"]
+
+    train_data = sb.dataio.dataset.DynamicItemDataset.from_csv(
+        csv_path=hparams["train_csv"], replacements={"data_root": data_folder},
+    )
+
+    if hparams["sorting"] == "ascending":
+        # we sort training data to speed up training and get better results.
+        train_data = train_data.filtered_sorted(
+            sort_key="duration",
+            key_max_value={"duration": hparams["avoid_if_longer_than"]},
+        )
+        # when sorting do not shuffle in dataloader ! otherwise is pointless
+        hparams["dataloader_options"]["shuffle"] = False
+
+    elif hparams["sorting"] == "descending":
+        train_data = train_data.filtered_sorted(
+            sort_key="duration",
+            reverse=True,
+            key_max_value={"duration": hparams["avoid_if_longer_than"]},
+        )
+        # when sorting do not shuffle in dataloader ! otherwise is pointless
+        hparams["dataloader_options"]["shuffle"] = False
+
+    elif hparams["sorting"] == "random":
+        pass
+
+    else:
+        raise NotImplementedError(
+            "sorting must be random, ascending or descending"
+        )
+
+    valid_data = sb.dataio.dataset.DynamicItemDataset.from_csv(
+        csv_path=hparams["valid_csv"], replacements={"data_root": data_folder},
+    )
+    # We also sort the validation data so it is faster to validate
+    valid_data = valid_data.filtered_sorted(sort_key="duration")
+    test_datasets = {}
+    for csv_file in hparams["test_csv"]:
+        name = Path(csv_file).stem
+        test_datasets[name] = sb.dataio.dataset.DynamicItemDataset.from_csv(
+            csv_path=csv_file, replacements={"data_root": data_folder}
+        )
+        test_datasets[name] = test_datasets[name].filtered_sorted(
+            sort_key="duration"
+        )
+
+    datasets = [train_data, valid_data] + [i for k, i in test_datasets.items()]
+
+
+    # 2. Define audio pipeline:
+    @sb.utils.data_pipeline.takes("wav")
+    @sb.utils.data_pipeline.provides("sig")
+    def audio_pipeline(wav):
+        info = torchaudio.info(wav)
+        sig = sb.dataio.dataio.read_audio(wav)
+        resampled = torchaudio.transforms.Resample(
+            info.sample_rate, hparams["sample_rate"],
+        )(sig)
+        return resampled
+
+    sb.dataio.dataset.add_dynamic_item(datasets, audio_pipeline)
+    label_encoder = sb.dataio.encoder.CTCTextEncoder()
+
+    # 3. Define text pipeline:
+    @sb.utils.data_pipeline.takes("wrd")
+    @sb.utils.data_pipeline.provides(
+        "wrd", "char_list", "tokens_list", "tokens"
+    )
+    def text_pipeline(wrd):
+        yield wrd
+        char_list = list(wrd)
+        yield char_list
+        tokens_list = label_encoder.encode_sequence(char_list)
+        yield tokens_list
+        tokens = torch.LongTensor(tokens_list)
+        yield tokens
+
+    sb.dataio.dataset.add_dynamic_item(datasets, text_pipeline)
+    lab_enc_file = os.path.join(hparams["save_folder"], "label_encoder.txt")
+    special_labels = {
+        "blank_label": hparams["blank_index"],
+        "unk_label": hparams["unk_index"]
+    }
+    label_encoder.load_or_create(
+        path=lab_enc_file,
+        from_didatasets=[train_data],
+        output_key="char_list",
+        special_labels=special_labels,
+        sequence_input=True,
+    )
+
+    # 4. Set output:
+    sb.dataio.dataset.set_output_keys(
+        datasets, ["id", "sig", "wrd", "char_list", "tokens"],
+    )
+    return train_data, valid_data,test_datasets, label_encoder
+
+
+if __name__ == "__main__":
+
+    # Load hyperparameters file with command-line overrides
+    hparams_file, run_opts, overrides = sb.parse_arguments(sys.argv[1:])
+    with open(hparams_file) as fin:
+        hparams = load_hyperpyyaml(fin, overrides)
+
+    # If --distributed_launch then
+    # create ddp_group with the right communication protocol
+    sb.utils.distributed.ddp_init_group(run_opts)
+
+
+    # Create experiment directory
+    sb.create_experiment_directory(
+        experiment_directory=hparams["output_folder"],
+        hyperparams_to_save=hparams_file,
+        overrides=overrides,
+    )
+
+    # Due to DDP, we do the preparation ONLY on the main python process
+    # Defining tokenizer and loading it
+    # Create the datasets objects as well as tokenization and encoding :-D
+    train_data, valid_data, test_datasets, label_encoder = dataio_prepare(hparams)
+    if hparams["use_language_modelling"]:
+        print("using langauge_modeeling")
+        from pyctcdecode import build_ctcdecoder
+        ind2lab = label_encoder.ind2lab
+        print(ind2lab)
+        labels = [ind2lab[x] for x in range(len(ind2lab))]
+        labels = [""] + labels[1:-1] + ["1"] 
+        # Replace the <blank> token with a blank character, needed for PyCTCdecode
+        print(labels)
+        decoder = build_ctcdecoder(
+            labels,
+            kenlm_model_path=hparams["ngram_lm_path"],  # .arpa or .bin
+            alpha=0.5,  # Default by KenLM
+            beta=1.0,  # Default by KenLM
+        )
+    # Trainer initialization
+    asr_brain = ASR(
+        modules=hparams["modules"],
+        hparams=hparams,
+        run_opts=run_opts,
+        checkpointer=hparams["checkpointer"],
+    )
+
+    # Adding objects to trainer.
+    asr_brain.tokenizer = label_encoder
+
+    # Training
+    asr_brain.fit(
+        asr_brain.hparams.epoch_counter,
+        train_data,
+        valid_data,
+        train_loader_kwargs=hparams["dataloader_options"],
+        valid_loader_kwargs=hparams["test_dataloader_options"],
+    )
+
+    # Test
+    for k in test_datasets.keys():  # keys are test_clean, test_other etc
+        asr_brain.hparams.wer_file = os.path.join(
+            hparams["output_folder"], "wer_{}.txt".format(k)
+        )
+        asr_brain.evaluate(
+            test_datasets[k], test_loader_kwargs=hparams["test_dataloader_options"]
+        )
+