minor updates to the scripts

egs/librispeech/ASR/zipformer/attention_decoder.py

@@ -236,7 +236,7 @@ class TransformerDecoder(nn.Module):
         causal_mask = subsequent_mask(x.shape[0], device=x.device)  # (seq_len, seq_len)
         attn_mask = torch.logical_or(
             padding_mask.unsqueeze(1),  # (batch, 1, seq_len)
-            torch.logical_not(causal_mask).unsqueeze(0)  # (1, seq_len, seq_len)
+            torch.logical_not(causal_mask).unsqueeze(0),  # (1, seq_len, seq_len)
         )  # (batch, seq_len, seq_len)
 
         if memory is not None:
@@ -367,7 +367,9 @@ class MultiHeadAttention(nn.Module):
         self.num_heads = num_heads
         self.head_dim = attention_dim // num_heads
         assert self.head_dim * num_heads == attention_dim, (
-            self.head_dim, num_heads, attention_dim
+            self.head_dim,
+            num_heads,
+            attention_dim,
         )
         self.dropout = dropout
         self.name = None  # will be overwritten in training code; for diagnostics.
@@ -437,15 +439,19 @@ class MultiHeadAttention(nn.Module):
         if key_padding_mask is not None:
             assert key_padding_mask.shape == (batch, src_len), key_padding_mask.shape
             attn_weights = attn_weights.masked_fill(
-                key_padding_mask.unsqueeze(1).unsqueeze(2), float("-inf"),
+                key_padding_mask.unsqueeze(1).unsqueeze(2),
+                float("-inf"),
             )
 
         if attn_mask is not None:
-            assert (
-                attn_mask.shape == (batch, 1, src_len)
-                or attn_mask.shape == (batch, tgt_len, src_len)
+            assert attn_mask.shape == (batch, 1, src_len) or attn_mask.shape == (
+                batch,
+                tgt_len,
+                src_len,
             ), attn_mask.shape
-            attn_weights = attn_weights.masked_fill(attn_mask.unsqueeze(1), float("-inf"))
+            attn_weights = attn_weights.masked_fill(
+                attn_mask.unsqueeze(1), float("-inf")
+            )
 
         attn_weights = attn_weights.view(batch * num_heads, tgt_len, src_len)
         attn_weights = nn.functional.softmax(attn_weights, dim=-1)
@@ -456,7 +462,11 @@ class MultiHeadAttention(nn.Module):
 
         # (batch * head, tgt_len, head_dim)
         attn_output = torch.bmm(attn_weights, v)
-        assert attn_output.shape == (batch * num_heads, tgt_len, head_dim), attn_output.shape
+        assert attn_output.shape == (
+            batch * num_heads,
+            tgt_len,
+            head_dim,
+        ), attn_output.shape
 
         attn_output = attn_output.transpose(0, 1).contiguous()
         attn_output = attn_output.view(tgt_len, batch, num_heads * head_dim)

egs/librispeech/ASR/zipformer/export-onnx-streaming.py

@@ -487,6 +487,7 @@ def export_encoder_model_onnx(
 
     add_meta_data(filename=encoder_filename, meta_data=meta_data)
 
+
 def export_decoder_model_onnx(
     decoder_model: OnnxDecoder,
     decoder_filename: str,
@@ -754,30 +755,31 @@ def main():
     )
     logging.info(f"Exported joiner to {joiner_filename}")
 
-    if(params.fp16) :
+    if params.fp16:
         from onnxconverter_common import float16
+
         logging.info("Generate fp16 models")
 
         encoder = onnx.load(encoder_filename)
         encoder_fp16 = float16.convert_float_to_float16(encoder, keep_io_types=True)
         encoder_filename_fp16 = params.exp_dir / f"encoder-{suffix}.fp16.onnx"
-        onnx.save(encoder_fp16,encoder_filename_fp16)
+        onnx.save(encoder_fp16, encoder_filename_fp16)
 
         decoder = onnx.load(decoder_filename)
         decoder_fp16 = float16.convert_float_to_float16(decoder, keep_io_types=True)
         decoder_filename_fp16 = params.exp_dir / f"decoder-{suffix}.fp16.onnx"
-        onnx.save(decoder_fp16,decoder_filename_fp16)
+        onnx.save(decoder_fp16, decoder_filename_fp16)
 
         joiner = onnx.load(joiner_filename)
         joiner_fp16 = float16.convert_float_to_float16(joiner, keep_io_types=True)
         joiner_filename_fp16 = params.exp_dir / f"joiner-{suffix}.fp16.onnx"
-        onnx.save(joiner_fp16,joiner_filename_fp16)
+        onnx.save(joiner_fp16, joiner_filename_fp16)
 
     # Generate int8 quantization models
     # See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection
 
     logging.info("Generate int8 quantization models")
-    
+
     encoder_filename_int8 = params.exp_dir / f"encoder-{suffix}.int8.onnx"
     quantize_dynamic(
         model_input=encoder_filename,

egs/librispeech/ASR/zipformer/export-onnx.py

@@ -592,23 +592,23 @@ def main():
     )
     logging.info(f"Exported joiner to {joiner_filename}")
 
-    if(params.fp16) :
+    if params.fp16:
         logging.info("Generate fp16 models")
 
         encoder = onnx.load(encoder_filename)
         encoder_fp16 = float16.convert_float_to_float16(encoder, keep_io_types=True)
         encoder_filename_fp16 = params.exp_dir / f"encoder-{suffix}.fp16.onnx"
-        onnx.save(encoder_fp16,encoder_filename_fp16)
+        onnx.save(encoder_fp16, encoder_filename_fp16)
 
         decoder = onnx.load(decoder_filename)
         decoder_fp16 = float16.convert_float_to_float16(decoder, keep_io_types=True)
         decoder_filename_fp16 = params.exp_dir / f"decoder-{suffix}.fp16.onnx"
-        onnx.save(decoder_fp16,decoder_filename_fp16)
+        onnx.save(decoder_fp16, decoder_filename_fp16)
 
         joiner = onnx.load(joiner_filename)
         joiner_fp16 = float16.convert_float_to_float16(joiner, keep_io_types=True)
         joiner_filename_fp16 = params.exp_dir / f"joiner-{suffix}.fp16.onnx"
-        onnx.save(joiner_fp16,joiner_filename_fp16)
+        onnx.save(joiner_fp16, joiner_filename_fp16)
 
     # Generate int8 quantization models
     # See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection

egs/libritts/ASR/local/compute_fbank_libritts.py

@@ -124,7 +124,7 @@ def compute_fbank_libritts(
                 supervisions=m["supervisions"],
             )
             if sampling_rate != 24000:
-                logging.info(f"Resampling audio to {sampling_rate}")
+                logging.info(f"Resampling audio to {sampling_rate}Hz")
                 cut_set = cut_set.resample(sampling_rate)
             if "train" in partition:
                 if perturb_speed:

egs/libritts/ASR/local/convert_transcript_words_to_tokens.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/local/convert_transcript_words_to_tokens.py

egs/libritts/ASR/local/download_lm.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/local/download_lm.py

egs/libritts/ASR/local/norm_text.py

@@ -0,0 +1 @@
+../../../libriheavy/ASR/local/norm_text.py

egs/libritts/ASR/local/prepare_lang.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lang.py

egs/libritts/ASR/local/prepare_lang_bpe.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lang_bpe.py

egs/libritts/ASR/local/prepare_lang_fst.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lang_fst.py

egs/libritts/ASR/local/prepare_lm_training_data.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lm_training_data.py

egs/libritts/ASR/local/train_bpe_model.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/local/train_bpe_model.py

egs/libritts/ASR/local/validate_bpe_lexicon.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/local/validate_bpe_lexicon.py

egs/libritts/ASR/prepare.sh

@@ -7,9 +7,15 @@ set -eou pipefail
 
 stage=0
 stop_stage=100
-sampling_rate=24000
+sampling_rate=16000
 nj=32
 perturb_speed=true
+vocab_sizes=(
+  # 5000
+  # 2000
+  # 1000
+  500
+)
 
 dl_dir=$PWD/download
 
@@ -27,6 +33,15 @@ log() {
 
 log "dl_dir: $dl_dir"
 
+if [ $stage -le -1 ] && [ $stop_stage -ge -1 ]; then
+  log "Stage -1: Download LM" # we directly use the librispeech lm here
+  mkdir -p $dl_dir/lm
+  if [ ! -e $dl_dir/lm/.done ]; then
+    ./local/download_lm.py --out-dir=$dl_dir/lm
+    touch $dl_dir/lm/.done
+  fi
+fi
+
 if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
   log "Stage 0: Download data"
 
@@ -107,3 +122,73 @@ if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
     touch data/fbank/.msuan.done
   fi
 fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Train BPE model for normalized text"
+
+  if [ ! -f data/texts ]; then
+    gunzip -c data/manifests/libritts_supervisions_train-clean-100.jsonl.gz \
+      | jq ".text" | sed 's/"//g' \
+      | ./local/norm_text.py > data/texts
+
+    gunzip -c data/manifests/libritts_supervisions_train-clean-360.jsonl.gz \
+      | jq ".text" | sed 's/"//g' \
+      | ./local/norm_text.py >> data/texts
+
+    gunzip -c data/manifests/libritts_supervisions_train-other-500.jsonl.gz \
+      | jq ".text" | sed 's/"//g' \
+      | ./local/norm_text.py >> data/texts
+  fi
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+    mkdir -p $lang_dir
+
+    cp data/texts $lang_dir/text
+
+    if [ ! -f $lang_dir/bpe.model ]; then
+      ./local/train_bpe_model.py \
+        --lang-dir $lang_dir \
+        --vocab-size $vocab_size \
+        --transcript $lang_dir/text
+    fi
+  done
+fi
+
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 6: Prepare phone based lang"
+  lang_dir=data/lang_phone
+  mkdir -p $lang_dir
+
+  if [ ! -f $dl_dir/lm/librispeech-lexicon.txt ]; then
+    log "No lexicon file in $dl_dir/lm, please run :"
+    log "prepare.sh --stage -1 --stop-stage -1"
+    exit -1
+  fi
+
+  if [ ! -f $lang_dir/lexicon.txt ]; then
+    (echo '!SIL SIL'; echo '<SPOKEN_NOISE> SPN'; echo '<UNK> SPN'; ) |
+      cat - $dl_dir/lm/librispeech-lexicon.txt |
+      sort | uniq > $lang_dir/lexicon.txt
+  fi
+
+  if [ ! -f $lang_dir/L_disambig.pt ]; then
+    ./local/prepare_lang.py --lang-dir $lang_dir
+  fi
+
+  if [ ! -f $lang_dir/L.fst ]; then
+    log "Converting L.pt to L.fst"
+    ./shared/convert-k2-to-openfst.py \
+      --olabels aux_labels \
+      $lang_dir/L.pt \
+      $lang_dir/L.fst
+  fi
+
+  if [ ! -f $lang_dir/L_disambig.fst ]; then
+    log "Converting L_disambig.pt to L_disambig.fst"
+    ./shared/convert-k2-to-openfst.py \
+      --olabels aux_labels \
+      $lang_dir/L_disambig.pt \
+      $lang_dir/L_disambig.fst
+  fi
+fi

egs/libritts/ASR/prepare_lm.sh

@@ -0,0 +1,264 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+# This script generate Ngram LM / NNLM and related files that needed by decoding.
+
+# We assume dl_dir (download dir) contains the following
+# directories and files. If not, they will be downloaded
+# by this script automatically.
+#
+#  - $dl_dir/lm
+#      This directory contains the following files downloaded from
+#       http://www.openslr.org/resources/11
+#
+#        - 3-gram.pruned.1e-7.arpa.gz
+#        - 3-gram.pruned.1e-7.arpa
+#        - 4-gram.arpa.gz
+#        - 4-gram.arpa
+#        - librispeech-vocab.txt
+#        - librispeech-lexicon.txt
+#        - librispeech-lm-norm.txt.gz
+#
+
+. prepare.sh --stage -1 --stop-stage 6 || exit 1
+
+log "Running prepare_lm.sh"
+
+stage=0
+stop_stage=100
+
+. shared/parse_options.sh || exit 1
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Prepare BPE based lexicon."
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+    # We reuse words.txt from phone based lexicon
+    # so that the two can share G.pt later.
+    cp data/lang_phone/words.txt $lang_dir
+
+    if [ ! -f $lang_dir/L_disambig.pt ]; then
+      ./local/prepare_lang_bpe.py --lang-dir $lang_dir
+
+      log "Validating $lang_dir/lexicon.txt"
+      ./local/validate_bpe_lexicon.py \
+        --lexicon $lang_dir/lexicon.txt \
+        --bpe-model $lang_dir/bpe.model
+    fi
+
+    if [ ! -f $lang_dir/L.fst ]; then
+      log "Converting L.pt to L.fst"
+      ./shared/convert-k2-to-openfst.py \
+        --olabels aux_labels \
+        $lang_dir/L.pt \
+        $lang_dir/L.fst
+    fi
+
+    if [ ! -f $lang_dir/L_disambig.fst ]; then
+      log "Converting L_disambig.pt to L_disambig.fst"
+      ./shared/convert-k2-to-openfst.py \
+        --olabels aux_labels \
+        $lang_dir/L_disambig.pt \
+        $lang_dir/L_disambig.fst
+    fi
+  done
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare word level G"
+  # We assume you have installed kaldilm, if not, please install
+  # it using: pip install kaldilm
+
+  mkdir -p data/lm
+  if [ ! -f data/lm/G_3_gram.fst.txt ]; then
+    # It is used in building HLG
+    python3 -m kaldilm \
+      --read-symbol-table="data/lang_phone/words.txt" \
+      --disambig-symbol='#0' \
+      --max-order=3 \
+      $dl_dir/lm/3-gram.pruned.1e-7.arpa > data/lm/G_3_gram.fst.txt
+  fi
+
+  if [ ! -f data/lm/G_4_gram.fst.txt ]; then
+    # It is used for LM rescoring
+    python3 -m kaldilm \
+      --read-symbol-table="data/lang_phone/words.txt" \
+      --disambig-symbol='#0' \
+      --max-order=4 \
+      $dl_dir/lm/4-gram.arpa > data/lm/G_4_gram.fst.txt
+  fi
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+
+    if [ ! -f $lang_dir/HL.fst ]; then
+      ./local/prepare_lang_fst.py  \
+        --lang-dir $lang_dir \
+        --ngram-G ./data/lm/G_3_gram.fst.txt
+    fi
+  done
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Compile HLG"
+  ./local/compile_hlg.py --lang-dir data/lang_phone
+
+  # Note If ./local/compile_hlg.py throws OOM,
+  # please switch to the following command
+  #
+  # ./local/compile_hlg_using_openfst.py --lang-dir data/lang_phone
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+    ./local/compile_hlg.py --lang-dir $lang_dir
+
+    # Note If ./local/compile_hlg.py throws OOM,
+    # please switch to the following command
+    #
+    # ./local/compile_hlg_using_openfst.py --lang-dir $lang_dir
+  done
+fi
+
+# Compile LG for RNN-T fast_beam_search decoding
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Compile LG"
+  ./local/compile_lg.py --lang-dir data/lang_phone
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+    ./local/compile_lg.py --lang-dir $lang_dir
+  done
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Prepare token level ngram G"
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+
+    if [ ! -f $lang_dir/transcript_tokens.txt ]; then
+      ./local/convert_transcript_words_to_tokens.py \
+        --lexicon $lang_dir/lexicon.txt \
+        --transcript $lang_dir/transcript_words.txt \
+        --oov "<UNK>" \
+        > $lang_dir/transcript_tokens.txt
+    fi
+
+    for ngram in 2 3 4 5; do
+      if [ ! -f $lang_dir/${ngram}gram.arpa ]; then
+        ./shared/make_kn_lm.py \
+          -ngram-order ${ngram} \
+          -text $lang_dir/transcript_tokens.txt \
+          -lm $lang_dir/${ngram}gram.arpa
+      fi
+    
+      if [ ! -f $lang_dir/${ngram}gram.fst.txt ]; then
+        python3 -m kaldilm \
+          --read-symbol-table="$lang_dir/tokens.txt" \
+          --disambig-symbol='#0' \
+          --max-order=${ngram} \
+          $lang_dir/${ngram}gram.arpa > $lang_dir/${ngram}gram.fst.txt
+      fi
+    done
+  done
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Generate NNLM training data"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    log "Processing vocab_size == ${vocab_size}"
+    lang_dir=data/lang_bpe_${vocab_size}
+    out_dir=data/lm_training_bpe_${vocab_size}
+    mkdir -p $out_dir
+
+    ./local/prepare_lm_training_data.py \
+      --bpe-model $lang_dir/bpe.model \
+      --lm-data $dl_dir/lm/librispeech-lm-norm.txt \
+      --lm-archive $out_dir/lm_data.pt
+  done
+fi
+
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 6: Generate NNLM validation data"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    log "Processing vocab_size == ${vocab_size}"
+    out_dir=data/lm_training_bpe_${vocab_size}
+    mkdir -p $out_dir
+
+    if [ ! -f $out_dir/valid.txt ]; then
+      gunzip -c data/manifests/libritts_supervisions_dev-clean.jsonl.gz \
+        | jq ".text" | sed 's/"//g' \
+        | ./local/norm_text.py > $out_dir/valid.txt
+
+      gunzip -c data/manifests/libritts_supervisions_dev-other.jsonl.gz \
+        | jq ".text" | sed 's/"//g' \
+        | ./local/norm_text.py >> $out_dir/valid.txt
+    fi
+
+    lang_dir=data/lang_bpe_${vocab_size}
+    ./local/prepare_lm_training_data.py \
+      --bpe-model $lang_dir/bpe.model \
+      --lm-data $out_dir/valid.txt \
+      --lm-archive $out_dir/lm_data-valid.pt
+  done
+fi
+
+if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
+  log "Stage 7: Generate NNLM test data"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    log "Processing vocab_size == ${vocab_size}"
+    out_dir=data/lm_training_bpe_${vocab_size}
+    mkdir -p $out_dir
+
+    if [ ! -f $out_dir/test.txt ]; then
+      gunzip -c data/manifests/libritts_supervisions_test-clean.jsonl.gz \
+        | jq ".text" | sed 's/"//g' \
+        | ./local/norm_text.py > $out_dir/test.txt
+
+      gunzip -c data/manifests/libritts_supervisions_test-other.jsonl.gz \
+        | jq ".text" | sed 's/"//g' \
+        | ./local/norm_text.py >> $out_dir/test.txt
+    fi
+
+    lang_dir=data/lang_bpe_${vocab_size}
+    ./local/prepare_lm_training_data.py \
+      --bpe-model $lang_dir/bpe.model \
+      --lm-data $out_dir/test.txt \
+      --lm-archive $out_dir/lm_data-test.pt
+  done
+fi
+
+if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then
+  log "Stage 8: Sort NNLM training data"
+  # Sort LM training data by sentence length in descending order
+  # for ease of training.
+  #
+  # Sentence length equals to the number of BPE tokens
+  # in a sentence.
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    out_dir=data/lm_training_bpe_${vocab_size}
+    mkdir -p $out_dir
+    ./local/sort_lm_training_data.py \
+      --in-lm-data $out_dir/lm_data.pt \
+      --out-lm-data $out_dir/sorted_lm_data.pt \
+      --out-statistics $out_dir/statistics.txt
+
+    ./local/sort_lm_training_data.py \
+      --in-lm-data $out_dir/lm_data-valid.pt \
+      --out-lm-data $out_dir/sorted_lm_data-valid.pt \
+      --out-statistics $out_dir/statistics-valid.txt
+
+    ./local/sort_lm_training_data.py \
+      --in-lm-data $out_dir/lm_data-test.pt \
+      --out-lm-data $out_dir/sorted_lm_data-test.pt \
+      --out-statistics $out_dir/statistics-test.txt
+  done
+fi

egs/libritts/ASR/zipformer/decode.py

@@ -1041,13 +1041,13 @@ def main():
 
     # we need cut ids to display recognition results.
     args.return_cuts = True
-    librispeech = LibriTTSAsrDataModule(args)
+    libritts = LibriTTSAsrDataModule(args)
 
-    test_clean_cuts = librispeech.test_clean_cuts()
-    test_other_cuts = librispeech.test_other_cuts()
+    test_clean_cuts = libritts.test_clean_cuts()
+    test_other_cuts = libritts.test_other_cuts()
 
-    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
-    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+    test_clean_dl = libritts.test_dataloaders(test_clean_cuts)
+    test_other_dl = libritts.test_dataloaders(test_other_cuts)
 
     test_sets = ["test-clean", "test-other"]
     test_dl = [test_clean_dl, test_other_dl]

egs/libritts/ASR/zipformer/streaming_decode.py

@@ -864,10 +864,10 @@ def main():
     num_param = sum([p.numel() for p in model.parameters()])
     logging.info(f"Number of model parameters: {num_param}")
 
-    librispeech = LibriTTSAsrDataModule(args)
+    libritts = LibriTTSAsrDataModule(args)
 
-    test_clean_cuts = librispeech.test_clean_cuts()
-    test_other_cuts = librispeech.test_other_cuts()
+    test_clean_cuts = libritts.test_clean_cuts()
+    test_other_cuts = libritts.test_other_cuts()
 
     test_sets = ["test-clean", "test-other"]
     test_cuts = [test_clean_cuts, test_other_cuts]

egs/libritts/ASR/zipformer/train.py

@@ -603,6 +603,15 @@ def _to_int_tuple(s: str):
     return tuple(map(int, s.split(",")))
 
 
+def remove_punc_to_upper(text: str) -> str:
+    text = text.replace("‘", "'")
+    text = text.replace("’", "'")
+    tokens = set("abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789'")
+    s_list = [x.upper() if x in tokens else " " for x in text]
+    s = " ".join("".join(s_list).split()).strip()
+    return s
+
+
 def get_encoder_embed(params: AttributeDict) -> nn.Module:
     # encoder_embed converts the input of shape (N, T, num_features)
     # to the shape (N, (T - 7) // 2, encoder_dims).
@@ -1284,21 +1293,26 @@ def run(rank, world_size, args):
     if params.inf_check:
         register_inf_check_hooks(model)
 
-    librispeech = LibriTTSAsrDataModule(args)
+    libritts = LibriTTSAsrDataModule(args)
 
     if params.full_libri:
-        train_cuts = librispeech.train_all_shuf_cuts()
+        train_cuts = libritts.train_all_shuf_cuts()
 
         # previously we used the following code to load all training cuts,
         # strictly speaking, shuffled training cuts should be used instead,
         # but we leave the code here to demonstrate that there is an option
         # like this to combine multiple cutsets
 
-        # train_cuts = librispeech.train_clean_100_cuts()
-        # train_cuts += librispeech.train_clean_360_cuts()
-        # train_cuts += librispeech.train_other_500_cuts()
+        # train_cuts = libritts.train_clean_100_cuts()
+        # train_cuts += libritts.train_clean_360_cuts()
+        # train_cuts += libritts.train_other_500_cuts()
     else:
-        train_cuts = librispeech.train_clean_100_cuts()
+        train_cuts = libritts.train_clean_100_cuts()
+
+    def normalize_text(c: Cut):
+        text = remove_punc_to_upper(c.supervisions[0].text)
+        c.supervisions[0].text = text
+        return c
 
     def remove_short_and_long_utt(c: Cut):
         # Keep only utterances with duration between 1 second and 20 seconds
@@ -1338,6 +1352,7 @@ def run(rank, world_size, args):
         return True
 
     train_cuts = train_cuts.filter(remove_short_and_long_utt)
+    train_cuts = train_cuts.map(normalize_text)
 
     if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
         # We only load the sampler's state dict when it loads a checkpoint
@@ -1346,13 +1361,13 @@ def run(rank, world_size, args):
     else:
         sampler_state_dict = None
 
-    train_dl = librispeech.train_dataloaders(
+    train_dl = libritts.train_dataloaders(
         train_cuts, sampler_state_dict=sampler_state_dict
     )
 
-    valid_cuts = librispeech.dev_clean_cuts()
-    valid_cuts += librispeech.dev_other_cuts()
-    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+    valid_cuts = libritts.dev_clean_cuts()
+    valid_cuts += libritts.dev_other_cuts()
+    valid_dl = libritts.valid_dataloaders(valid_cuts)
 
     if not params.print_diagnostics:
         scan_pessimistic_batches_for_oom(

egs/libritts/CODEC/prepare.sh

@@ -37,15 +37,6 @@ if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
   if [ ! -d $dl_dir/LibriTTS ]; then
     lhotse download libritts $dl_dir
   fi
-
-  # If you have pre-downloaded it to /path/to/musan,
-  # you can create a symlink
-  #
-  #   ln -sfv /path/to/musan $dl_dir/musan
-  #
-  if [ ! -d $dl_dir/musan ]; then
-    lhotse download musan $dl_dir
-  fi
 fi
 
 if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then