Add People's Speech to multidataset

egs/librispeech/ASR/prepare_giga_speech.sh

@@ -149,11 +149,3 @@ if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
     touch data/fbank/.gigaspeech_XL.done
   fi
 fi
-
-if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
-  log "Stage 6: Combine features for XL (may take 15 hours)"
-  if [ ! -f data/fbank/gigaspeech_cuts_XL.jsonl.gz ]; then
-    pieces=$(find data/fbank/gigaspeech_XL_split_${num_splits} -name "gigaspeech_cuts_XL.*.jsonl.gz")
-    lhotse combine $pieces data/fbank/gigaspeech_cuts_XL.jsonl.gz
-  fi
-fi

egs/librispeech/ASR/prepare_multidataset.sh

@@ -37,10 +37,6 @@ stop_stage=100
 #     - noise
 #     - speech
 
-# Split all dataset to this number of pieces and mix each dataset pieces
-# into multidataset pieces with shuffling.
-num_splits=1998
-
 dl_dir=$PWD/download
 
 . shared/parse_options.sh || exit 1
@@ -61,6 +57,7 @@ vocab_sizes=(
 multidataset=(
   "gigaspeech",
   "commonvoice",
+  "peoples_speech",
 )
 
 # All files generated by this script are saved in "data".
@@ -319,7 +316,7 @@ if [ $stage -le 10 ] && [ $stop_stage -ge 10 ]; then
   # GigaSpeech
   if [[ "${multidataset[@]}" =~ "gigaspeech" ]]; then
     log "Dataset: GigaSpeech"
-    ./prepare_giga_speech.sh --stop_stage 5
+    ./prepare_giga_speech.sh
   fi
 
   # CommonVoice
@@ -327,4 +324,10 @@ if [ $stage -le 10 ] && [ $stop_stage -ge 10 ]; then
     log "Dataset: CommonVoice"
     ./prepare_common_voice.sh
   fi
+
+  # People's Speech
+  if [[ "${multidataset[@]}" =~ "peoples_speech" ]]; then
+    log "Dataset: People's Speech"
+    ./prepare_peoples_speech.sh
+  fi
 fi

egs/librispeech/ASR/prepare_peoples_speech.sh

@@ -0,0 +1,127 @@
+#!/usr/bin/env bash
+
+set -eou pipefail
+
+nj=32
+stage=-1
+stop_stage=100
+
+# Split data/set to a number of pieces
+# This is to avoid OOM during feature extraction.
+num_per_split=4000
+
+# We assume dl_dir (download dir) contains the following
+# directories and files. If not, they will be downloaded
+# by this script automatically.
+#
+#  - $dl_dir/peoples_speech
+#      This directory contains the following files downloaded from
+#       https://huggingface.co/datasets/MLCommons/peoples_speech
+# 
+#     - test
+#     - train
+#     - validation
+
+dl_dir=$PWD/download
+
+. shared/parse_options.sh || exit 1
+
+# vocab size for sentence piece models.
+# It will generate data/lang_bpe_xxx,
+# data/lang_bpe_yyy if the array contains xxx, yyy
+vocab_sizes=(
+  # 5000
+  # 2000
+  # 1000
+  500
+)
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Download data"
+
+  # If you have pre-downloaded it to /path/to/peoples_speech,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/peoples_speech $dl_dir/peoples_speech
+  #
+  if [ ! -d $dl_dir/peoples_speech/train ]; then
+    git lfs install
+    git clone https://huggingface.co/datasets/MLCommons/peoples_speech
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare People's Speech manifest"
+  # We assume that you have downloaded the People's Speech corpus
+  # to $dl_dir/peoples_speech
+  mkdir -p data/manifests
+  if [ ! -e data/manifests/.peoples_speech.done ]; then
+    lhotse prepare peoples-speech -j $nj $dl_dir/peoples_speech data/manifests
+    touch data/manifests/.peoples_speech.done
+  fi
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Preprocess People's Speech manifest"
+  mkdir -p data/fbank
+  if [ ! -e data/fbank/.preprocess_complete ]; then
+    ./local/preprocess_peoples_speech.py
+    touch data/fbank/.preprocess_complete
+  fi
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Compute fbank for valid and test subsets of People's Speech"
+  if [ ! -e data/fbank/.peoples_speech_valid_test.done ]; then
+    ./local/compute_fbank_peoples_speech_valid_test.py
+    touch data/fbank/.peoples_speech_valid_test.done
+  fi
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Split train subset into pieces"
+  split_dir=data/fbank/peoples_speech_train_split
+  if [ ! -e $split_dir/.peoples_speech_dirty_split.done ]; then
+    lhotse split-lazy ./data/fbank/peoples_speech_cuts_dirty_raw.jsonl.gz $split_dir $num_per_split
+    touch $split_dir/.peoples_speech_dirty_split.done
+  fi
+
+  if [ ! -e $split_dir/.peoples_speech_dirty_sa_split.done ]; then
+    lhotse split-lazy ./data/fbank/peoples_speech_cuts_dirty_sa_raw.jsonl.gz $split_dir $num_per_split
+    touch $split_dir/.peoples_speech_dirty_sa_split.done
+  fi
+
+  if [ ! -e $split_dir/.peoples_speech_clean_split.done ]; then
+    lhotse split-lazy ./data/fbank/peoples_speech_cuts_clean_raw.jsonl.gz $split_dir $num_per_split
+    touch $split_dir/.peoples_speech_clean_split.done
+  fi
+
+  if [ ! -e $split_dir/.peoples_speech_clean_sa_split.done ]; then
+    lhotse split-lazy ./data/fbank/peoples_speech_cuts_clean_sa_raw.jsonl.gz $split_dir $num_per_split
+    touch $split_dir/.peoples_speech_clean_sa_split.done
+  fi
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Compute features for train subset of People's Speech"
+  if [ ! -e data/fbank/.peoples_speech_train.done ]; then
+    ./local/compute_fbank_peoples_speech_splits.py \
+      --num-workers $nj \
+      --batch-duration 600 \
+      --start 0 \
+      --num-splits 2000
+    touch data/fbank/.peoples_speech_train.done
+  fi
+fi

egs/librispeech/ASR/pruned_transducer_stateless7/multidataset.py

@@ -33,6 +33,10 @@ class MultiDataset:
 
             - librispeech_cuts_train-all-shuf.jsonl.gz
             - gigaspeech_XL_split_2000/gigaspeech_cuts_XL.*.jsonl.gz
+            - peoples_speech_train_split/peoples_speech_cuts_dirty.*.jsonl.gz
+            - peoples_speech_train_split/peoples_speech_cuts_dirty_sa.*.jsonl.gz
+            - peoples_speech_train_split/peoples_speech_cuts_clean.*.jsonl.gz
+            - peoples_speech_train_split/peoples_speech_cuts_clean_sa.*.jsonl.gz
 
           cv_manifest_dir:
             It is expected to contain the following files:
@@ -74,4 +78,34 @@ class MultiDataset:
             self.cv_manifest_dir / f"cv-en_cuts_train.jsonl.gz"
         )
 
-        return CutSet.mux(librispeech_cuts, gigaspeech_cuts, commonvoice_cuts)
+        # People's Speech
+        filenames = glob.glob(
+            f"{self.manifest_dir}/peoples_speech_train_split/peoples_speech_cuts_*.*.jsonl.gz"
+        )
+
+        pattern = re.compile(r"peoples_speech_cuts.([0-9]+).jsonl.gz")
+        idx_filenames = ((int(pattern.search(f).group(1)), f) for f in filenames)
+        idx_filenames = sorted(idx_filenames, key=lambda x: x[0])
+
+        sorted_filenames = [f[1] for f in idx_filenames]
+
+        logging.info(
+            f"Loading People's Speech {len(sorted_filenames)} splits in lazy mode"
+        )
+
+        peoples_speech_cuts = lhotse.combine(
+            lhotse.load_manifest_lazy(p) for p in sorted_filenames
+        )
+
+        return CutSet.mux(
+            librispeech_cuts,
+            gigaspeech_cuts,
+            commonvoice_cuts,
+            peoples_speech_cuts,
+            weights=[
+                len(librispeech_cuts),
+                len(gigaspeech_cuts),
+                len(commonvoice_cuts),
+                len(peoples_speech_cuts),
+            ],
+        )