VITS recipe for LibriTTS corpus

README.md

@@ -333,6 +333,7 @@ We provide a Colab notebook to test the pre-trained model: [![Open In Colab](htt
 
   - [LJSpeech][ljspeech]
   - [VCTK][vctk]
+  - [LibriTTS][libritts_tts]
 
 ### Supported Models
 
@@ -372,6 +373,7 @@ Please see: [![Open In Colab](https://colab.research.google.com/assets/colab-bad
 [commonvoice]: egs/commonvoice/ASR
 [csj]: egs/csj/ASR
 [libricss]: egs/libricss/SURT
+[libritts_asr]: egs/libritts/ASR
 [libriheavy]: egs/libriheavy/ASR
 [mgb2]: egs/mgb2/ASR
 [spgispeech]: egs/spgispeech/ASR
@@ -380,3 +382,4 @@ Please see: [![Open In Colab](https://colab.research.google.com/assets/colab-bad
 
 [vctk]: egs/vctk/TTS
 [ljspeech]: egs/ljspeech/TTS
+[libritts_tts]: egs/libritts/TTS

egs/libritts/CODEC/encodec/train.py

@@ -138,7 +138,7 @@ def get_parser():
     parser.add_argument(
         "--save-every-n",
         type=int,
-        default=1,
+        default=5,
         help="""Save checkpoint after processing this number of epochs"
         periodically. We save checkpoint to exp-dir/ whenever
         params.cur_epoch % save_every_n == 0. The checkpoint filename
@@ -1093,14 +1093,14 @@ def run(rank, world_size, args):
         rank=rank,
     )
 
-    # if not params.print_diagnostics:
-    #     scan_pessimistic_batches_for_oom(
-    #         model=model,
-    #         train_dl=train_dl,
-    #         optimizer_g=optimizer_g,
-    #         optimizer_d=optimizer_d,
-    #         params=params,
-    #     )
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer_g=optimizer_g,
+            optimizer_d=optimizer_d,
+            params=params,
+        )
 
     scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
     if checkpoints and "grad_scaler" in checkpoints:

egs/libritts/CODEC/prepare.sh

@@ -45,12 +45,11 @@ if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
   # to $dl_dir/LibriTTS
   mkdir -p data/manifests
   if [ ! -e data/manifests/.libritts.done ]; then
-    lhotse prepare libritts --num-jobs 32 $dl_dir/LibriTTS data/manifests
+    lhotse prepare libritts --num-jobs ${nj} $dl_dir/LibriTTS data/manifests
     touch data/manifests/.libritts.done
   fi
 fi
 
-
 if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
   log "Stage 2: Compute Spectrogram for LibriTTS"
   mkdir -p data/spectrogram
@@ -64,7 +63,7 @@ if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
   if [ ! -f data/spectrogram/libritts_cuts_train-all-shuf.jsonl.gz ]; then
     cat <(gunzip -c data/spectrogram/libritts_cuts_train-clean-100.jsonl.gz) \
       <(gunzip -c data/spectrogram/libritts_cuts_train-clean-360.jsonl.gz) \
-      <(gunzip -c /data/spectrogramlibritts_cuts_train-other-500.jsonl.gz) | \
+      <(gunzip -c data/spectrogramlibritts_cuts_train-other-500.jsonl.gz) | \
       shuf | gzip -c > data/spectrogram/libritts_cuts_train-all-shuf.jsonl.gz
   fi
 

egs/libritts/TTS/README.md

@@ -0,0 +1,51 @@
+# Introduction
+
+LibriTTS is a multi-speaker English corpus of approximately 585 hours of read English speech at 24kHz sampling rate, prepared by Heiga Zen with the assistance of Google Speech and Google Brain team members. 
+The LibriTTS corpus is designed for TTS research. It is derived from the original materials (mp3 audio files from LibriVox and text files from Project Gutenberg) of the LibriSpeech corpus. 
+The main differences from the LibriSpeech corpus are listed below:
+1. The audio files are at 24kHz sampling rate.
+2. The speech is split at sentence breaks.
+3. Both original and normalized texts are included.
+4. Contextual information (e.g., neighbouring sentences) can be extracted.
+5. Utterances with significant background noise are excluded.
+For more information, refer to the paper "LibriTTS: A Corpus Derived from LibriSpeech for Text-to-Speech", Heiga Zen, Viet Dang, Rob Clark, Yu Zhang, Ron J. Weiss, Ye Jia, Zhifeng Chen, and Yonghui Wu, arXiv, 2019. If you use the LibriTTS corpus in your work, please cite this paper where it was introduced.
+
+> [!CAUTION]
+> The next-gen Kaldi framework provides tools and models for generating high-quality, synthetic speech (Text-to-Speech, TTS). 
+> While these recipes has the potential to advance various fields such as accessibility, language education, and AI-driven solutions, it also carries certain ethical and legal responsibilities.
+> 
+> By using this framework, you agree to the following:
+> 1.	Legal and Ethical Use: You shall not use this framework, or any models derived from it, for any unlawful or unethical purposes. This includes, but is not limited to: Creating voice clones without the explicit, informed consent of the individual whose voice is being cloned. Engaging in any form of identity theft, impersonation, or fraud using cloned voices. Violating any local, national, or international laws regarding privacy, intellectual property, or personal data.
+> 
+> 2.	Responsibility of Use: The users of this framework are solely responsible for ensuring that their use of voice cloning technologies complies with all applicable laws and ethical guidelines. We explicitly disclaim any liability for misuse of the technology.
+> 
+> 3.	Attribution and Use of Open-Source Components: This project is provided under the Apache 2.0 license. Users must adhere to the terms of this license and provide appropriate attribution when required.
+> 
+> 4.	No Warranty: This framework is provided “as-is,” without warranty of any kind, either express or implied. We do not guarantee that the use of this software will comply with legal requirements or that it will not infringe the rights of third parties.
+
+
+# VITS
+
+This recipe provides a VITS model trained on the LibriTTS dataset.
+
+Pretrained model can be found [here](https://huggingface.co/zrjin/icefall-tts-libritts-vits-2024-10-30).
+
+The training command is given below:
+```
+export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+./vits/train.py \
+  --world-size 4 \
+  --num-epochs 400 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir vits/exp \
+  --max-duration 500
+```
+
+To inference, use:
+```
+./vits/infer.py \
+  --exp-dir vits/exp \
+  --epoch 400 \
+  --tokens data/tokens.txt
+```

egs/libritts/TTS/local/compute_spectrogram_libritts.py

@@ -0,0 +1 @@
+../../CODEC/local/compute_spectrogram_libritts.py

egs/libritts/TTS/local/prepare_token_file.py

@@ -0,0 +1 @@
+../../../ljspeech/TTS/local/prepare_token_file.py

egs/libritts/TTS/local/prepare_tokens_libritts.py

@@ -0,0 +1,89 @@
+#!/usr/bin/env python3
+# Copyright         2023  Xiaomi Corp.        (authors: Zengwei Yao,
+#                                                       Zengrui Jin,)
+#                   2024  Tsinghua University (authors: Zengrui Jin,)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file reads the texts in given manifest and save the new cuts with phoneme tokens.
+"""
+
+import logging
+from pathlib import Path
+
+import tacotron_cleaner.cleaners
+from lhotse import CutSet, load_manifest
+from piper_phonemize import phonemize_espeak
+from tqdm.auto import tqdm
+
+
+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 prepare_tokens_libritts():
+    output_dir = Path("data/spectrogram")
+    prefix = "libritts"
+    suffix = "jsonl.gz"
+    partitions = (
+        "dev-clean",
+        "dev-other",
+        "test-clean",
+        "test-other",
+        "train-all-shuf",
+        "train-clean-460",
+        # "train-clean-100",
+        # "train-clean-360",
+        # "train-other-500",
+    )
+
+    for partition in partitions:
+        cut_set = load_manifest(output_dir / f"{prefix}_cuts_{partition}.{suffix}")
+
+        new_cuts = []
+        for cut in tqdm(cut_set):
+            # Each cut only contains one supervision
+            assert len(cut.supervisions) == 1, (len(cut.supervisions), cut)
+            text = cut.supervisions[0].text
+            # Text normalization
+            text = tacotron_cleaner.cleaners.custom_english_cleaners(text)
+            # Convert to phonemes
+            tokens_list = phonemize_espeak(text, "en-us")
+            tokens = []
+            for t in tokens_list:
+                tokens.extend(t)
+            cut.tokens = tokens
+            cut.supervisions[0].normalized_text = remove_punc_to_upper(text)
+
+            new_cuts.append(cut)
+
+        new_cut_set = CutSet.from_cuts(new_cuts)
+        new_cut_set.to_file(
+            output_dir / f"{prefix}_cuts_with_tokens_{partition}.{suffix}"
+        )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    prepare_tokens_libritts()

egs/libritts/TTS/local/validate_manifest.py

@@ -0,0 +1 @@
+../../../ljspeech/TTS/local/validate_manifest.py

egs/libritts/TTS/prepare.sh

@@ -0,0 +1,134 @@
+#!/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
+
+stage=0
+stop_stage=100
+sampling_rate=24000
+nj=32
+
+dl_dir=$PWD/download
+
+. shared/parse_options.sh || exit 1
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+log() {
+  # This function is from espnet
+  local fname=${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 -1 ] && [ $stop_stage -ge -1 ]; then
+  log "Stage -1: build monotonic_align lib"
+  if [ ! -d vits/monotonic_align/build ]; then
+    cd vits/monotonic_align
+    python setup.py build_ext --inplace
+    cd ../../
+  else 
+    log "monotonic_align lib already built"
+  fi
+fi
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Download data"
+
+  # If you have pre-downloaded it to /path/to/LibriTTS,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/LibriTTS $dl_dir/LibriTTS
+  #
+  if [ ! -d $dl_dir/LibriTTS ]; then
+    lhotse download libritts $dl_dir
+  fi
+
+  if [ ! -d $dl_dir/xvector_nnet_1a_libritts_clean_460 ]; then
+    log "Downloading x-vector"
+
+    git clone https://huggingface.co/datasets/zrjin/xvector_nnet_1a_libritts_clean_460 $dl_dir/xvector_nnet_1a_libritts_clean_460
+
+    mkdir -p exp/xvector_nnet_1a/
+    cp -r $dl_dir/xvector_nnet_1a_libritts_clean_460/* exp/xvector_nnet_1a/
+  fi
+
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare LibriTTS manifest"
+  # We assume that you have downloaded the LibriTTS corpus
+  # to $dl_dir/LibriTTS
+  mkdir -p data/manifests
+  if [ ! -e data/manifests/.libritts.done ]; then
+    lhotse prepare libritts --num-jobs ${nj} $dl_dir/LibriTTS data/manifests
+    touch data/manifests/.libritts.done
+  fi
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Compute Spectrogram for LibriTTS"
+  mkdir -p data/spectrogram
+  if [ ! -e data/spectrogram/.libritts.done ]; then
+    ./local/compute_spectrogram_libritts.py --sampling-rate $sampling_rate 
+    touch data/spectrogram/.libritts.done
+  fi
+
+  # Here we shuffle and combine the train-clean-100, train-clean-360 and 
+  # train-other-500 together to form the training set.
+  if [ ! -f data/spectrogram/libritts_cuts_train-all-shuf.jsonl.gz ]; then
+    cat <(gunzip -c data/spectrogram/libritts_cuts_train-clean-100.jsonl.gz) \
+      <(gunzip -c data/spectrogram/libritts_cuts_train-clean-360.jsonl.gz) \
+      <(gunzip -c data/spectrogramlibritts_cuts_train-other-500.jsonl.gz) | \
+      shuf | gzip -c > data/spectrogram/libritts_cuts_train-all-shuf.jsonl.gz
+  fi
+
+  # Here we shuffle and combine the train-clean-100, train-clean-360 
+  # together to form the training set.
+  if [ ! -f data/spectrogram/libritts_cuts_train-clean-460.jsonl.gz ]; then
+    cat <(gunzip -c data/spectrogram/libritts_cuts_train-clean-100.jsonl.gz) \
+      <(gunzip -c data/spectrogram/libritts_cuts_train-clean-360.jsonl.gz) | \
+      shuf | gzip -c > data/spectrogram/libritts_cuts_train-clean-460.jsonl.gz
+  fi
+
+  if [ ! -e data/spectrogram/.libritts-validated.done ]; then
+    log "Validating data/spectrogram for LibriTTS"
+    ./local/validate_manifest.py \
+      data/spectrogram/libritts_cuts_train-all-shuf.jsonl.gz
+    touch data/spectrogram/.libritts-validated.done
+  fi
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Prepare phoneme tokens for LibriTTS"
+  # We assume you have installed piper_phonemize and espnet_tts_frontend.
+  # If not, please install them with:
+  #   - piper_phonemize: 
+  #       refer to https://github.com/rhasspy/piper-phonemize,
+  #       could install the pre-built wheels from https://github.com/csukuangfj/piper-phonemize/releases/tag/2023.12.5
+  #   - espnet_tts_frontend: 
+  #       `pip install espnet_tts_frontend`, refer to https://github.com/espnet/espnet_tts_frontend/
+  if [ ! -e data/spectrogram/.libritts_with_token.done ]; then
+    ./local/prepare_tokens_libritts.py
+    touch data/spectrogram/.libritts_with_token.done
+  fi
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Generate token file"
+  # We assume you have installed piper_phonemize and espnet_tts_frontend.
+  # If not, please install them with:
+  #   - piper_phonemize: 
+  #       refer to https://github.com/rhasspy/piper-phonemize,
+  #       could install the pre-built wheels from https://github.com/csukuangfj/piper-phonemize/releases/tag/2023.12.5
+  #   - espnet_tts_frontend: 
+  #       `pip install espnet_tts_frontend`, refer to https://github.com/espnet/espnet_tts_frontend/
+  if [ ! -e data/tokens.txt ]; then
+    ./local/prepare_token_file.py --tokens data/tokens.txt
+  fi
+fi

egs/libritts/TTS/shared

@@ -0,0 +1 @@
+../../../icefall/shared/

egs/libritts/TTS/vits/duration_predictor.py

@@ -0,0 +1 @@
+../../../ljspeech/TTS/vits/duration_predictor.py

egs/libritts/TTS/vits/flow.py

@@ -0,0 +1 @@
+../../../ljspeech/TTS/vits/flow.py

egs/libritts/TTS/vits/generator.py

@@ -0,0 +1 @@
+../../../ljspeech/TTS/vits/generator.py

egs/libritts/TTS/vits/hifigan.py

@@ -0,0 +1 @@
+../../../ljspeech/TTS/vits/hifigan.py