Use piper_phonemize as text tokenizer in ljspeech recipe

egs/ljspeech/TTS/local/prepare_token_file.py

@@ -17,88 +17,42 @@
 
 
 """
-This file reads the texts in given manifest and generates the file that maps tokens to IDs.
+This file generates the file that maps tokens to IDs.
 """
 
 import argparse
 import logging
 from pathlib import Path
 from typing import Dict
 
-from lhotse import load_manifest
+from piper_phonemize import get_espeak_map
 
 
 def get_args():
  parser = argparse.ArgumentParser()
 
- parser.add_argument(
- "--manifest-file",
- type=Path,
- default=Path("data/spectrogram/ljspeech_cuts_train.jsonl.gz"),
- help="Path to the manifest file",
- )
-
  parser.add_argument(
  "--tokens",
  type=Path,
  default=Path("data/tokens.txt"),
- help="Path to the tokens",
+ help="Path to the dict that maps the text tokens to IDs",
  )
 
  return parser.parse_args()
 
 
-def write_mapping(filename: str, sym2id: Dict[str, int]) -> None:
- """Write a symbol to ID mapping to a file.
-
- Note:
- No need to implement `read_mapping` as it can be done
- through :func:`k2.SymbolTable.from_file`.
-
- Args:
- filename:
- Filename to save the mapping.
- sym2id:
- A dict mapping symbols to IDs.
- Returns:
- Return None.
- """
+def get_token2id(filename: Path) -> Dict[str, int]:
+ """Get a dict that maps token to IDs, and save it to the given filename."""
+ all_tokens = get_espeak_map()
  with open(filename, "w", encoding="utf-8") as f:
- for sym, i in sym2id.items():
- f.write(f"{sym} {i}\n")
-
-
-def get_token2id(manifest_file: Path) -> Dict[str, int]:
- """Return a dict that maps token to IDs."""
- extra_tokens = [
- "<blk>", # 0 for blank
- "<sos/eos>", # 1 for sos and eos symbols.
- "<unk>", # 2 for OOV
- ]
- all_tokens = set()
-
- cut_set = load_manifest(manifest_file)
-
- for cut in cut_set:
- # Each cut only contain one supervision
- assert len(cut.supervisions) == 1, len(cut.supervisions)
- for t in cut.tokens:
- all_tokens.add(t)
-
- all_tokens = extra_tokens + list(all_tokens)
-
- token2id: Dict[str, int] = {token: i for i, token in enumerate(all_tokens)}
- return token2id
+ for token, token_id in all_tokens.items():
+ f.write(f"{token} {token_id[0]}\n")
 
 
 if __name__ == "__main__":
  formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
-
  logging.basicConfig(format=formatter, level=logging.INFO)
 
  args = get_args()
- manifest_file = Path(args.manifest_file)
  out_file = Path(args.tokens)
-
- token2id = get_token2id(manifest_file)
- write_mapping(out_file, token2id)
+ get_token2id(out_file)

egs/ljspeech/TTS/local/prepare_tokens_ljspeech.py

@@ -23,9 +23,9 @@
 import logging
 from pathlib import Path
 
-import g2p_en
 import tacotron_cleaner.cleaners
 from lhotse import CutSet, load_manifest
+from piper_phonemize import phonemize_espeak
 
 
 def prepare_tokens_ljspeech():
@@ -35,7 +35,6 @@ def prepare_tokens_ljspeech():
  partition = "all"
 
  cut_set = load_manifest(output_dir / f"{prefix}_cuts_{partition}.{suffix}")
- g2p = g2p_en.G2p()
 
  new_cuts = []
  for cut in cut_set:
@@ -45,7 +44,11 @@ def prepare_tokens_ljspeech():
  # Text normalization
  text = tacotron_cleaner.cleaners.custom_english_cleaners(text)
  # Convert to phonemes
- cut.tokens = g2p(text)
+ tokens_list = phonemize_espeak(text, "en-us")
+ tokens = []
+ for t in tokens_list:
+ tokens.extend(t)
+ cut.tokens = tokens
  new_cuts.append(cut)
 
  new_cut_set = CutSet.from_cuts(new_cuts)

egs/ljspeech/TTS/prepare.sh

@@ -30,7 +30,7 @@ if [ $stage -le -1 ] && [ $stop_stage -ge -1 ]; then
  cd vits/monotonic_align
  python setup.py build_ext --inplace
  cd ../../
- else 
+ else
  log "monotonic_align lib already built"
  fi
 fi
@@ -80,6 +80,11 @@ fi
 
 if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
  log "Stage 3: Prepare phoneme tokens for LJSpeech"
+ # 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/.ljspeech_with_token.done ]; then
  ./local/prepare_tokens_ljspeech.py
  mv data/spectrogram/ljspeech_cuts_with_tokens_all.jsonl.gz \
@@ -113,13 +118,12 @@ fi
 
 if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
  log "Stage 5: Generate token file"
- # We assume you have installed g2p_en and espnet_tts_frontend.
+ # We assume you have installed piper_phonemize and espnet_tts_frontend.
  # If not, please install them with:
- # - g2p_en: `pip install g2p_en`, refer to https://github.com/Kyubyong/g2p
+ # - 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 \
- --manifest-file data/spectrogram/ljspeech_cuts_train.jsonl.gz \
- --tokens data/tokens.txt
+ ./local/prepare_token_file.py --tokens data/tokens.txt
  fi
 fi

egs/ljspeech/TTS/vits/tokenizer.py

@@ -16,8 +16,8 @@
 
 from typing import Dict, List
 
-import g2p_en
 import tacotron_cleaner.cleaners
+from piper_phonemize import phonemize_espeak
 from utils import intersperse
 
 
@@ -38,21 +38,34 @@ def __init__(self, tokens: str):
  id = int(info[0])
  else:
  token, id = info[0], int(info[1])
+ assert token not in self.token2id, token
  self.token2id[token] = id
 
- self.blank_id = self.token2id["<blk>"]
- self.oov_id = self.token2id["<unk>"]
- self.vocab_size = len(self.token2id)
+ # Refer to https://github.com/rhasspy/piper/blob/master/TRAINING.md
+ self.pad_id = self.token2id["_"] # padding
+ self.sos_id = self.token2id["^"] # beginning of an utterance (bos)
+ self.eos_id = self.token2id["$"] # end of an utterance (eos)
+ self.space_id = self.token2id[" "] # word separator (whitespace)
 
- self.g2p = g2p_en.G2p()
+ self.vocab_size = len(self.token2id)
 
- def texts_to_token_ids(self, texts: List[str], intersperse_blank: bool = True):
+ def texts_to_token_ids(
+ self,
+ texts: List[str],
+ intersperse_blank: bool = True,
+ add_sos: bool = False,
+ add_eos: bool = False,
+ ):
  """
  Args:
  texts:
  A list of transcripts.
  intersperse_blank:
  Whether to intersperse blanks in the token sequence.
+ add_sos:
+ Whether to add sos token at the start.
+ add_eos:
+ Whether to add eos token at the end.
 
  Returns:
  Return a list of token id list [utterance][token_id]
@@ -63,30 +76,44 @@ def texts_to_token_ids(self, texts: List[str], intersperse_blank: bool = True):
  # Text normalization
  text = tacotron_cleaner.cleaners.custom_english_cleaners(text)
  # Convert to phonemes
- tokens = self.g2p(text)
+ tokens_list = phonemize_espeak(text, "en-us")
+ tokens = []
+ for t in tokens_list:
+ tokens.extend(t)
+
  token_ids = []
  for t in tokens:
- if t in self.token2id:
- token_ids.append(self.token2id[t])
- else:
- token_ids.append(self.oov_id)
+ assert t in self.token2id, t
+ token_ids.append(self.token2id[t])
 
  if intersperse_blank:
- token_ids = intersperse(token_ids, self.blank_id)
+ token_ids = intersperse(token_ids, self.pad_id)
+ if add_sos:
+ token_ids = [self.sos_id] + token_ids
+ if add_eos:
+ token_ids = token_ids + [self.eos_id]
 
  token_ids_list.append(token_ids)
 
  return token_ids_list
 
  def tokens_to_token_ids(
- self, tokens_list: List[str], intersperse_blank: bool = True
+ self,
+ tokens_list: List[str],
+ intersperse_blank: bool = True,
+ add_sos: bool = False,
+ add_eos: bool = False,
  ):
  """
  Args:
  tokens_list:
  A list of token list, each corresponding to one utterance.
  intersperse_blank:
  Whether to intersperse blanks in the token sequence.
+ add_sos:
+ Whether to add sos token at the start.
+ add_eos:
+ Whether to add eos token at the end.
 
  Returns:
  Return a list of token id list [utterance][token_id]
@@ -96,13 +123,15 @@ def tokens_to_token_ids(
  for tokens in tokens_list:
  token_ids = []
  for t in tokens:
- if t in self.token2id:
- token_ids.append(self.token2id[t])
- else:
- token_ids.append(self.oov_id)
+ assert t in self.token2id, t
+ token_ids.append(self.token2id[t])
 
  if intersperse_blank:
- token_ids = intersperse(token_ids, self.blank_id)
+ token_ids = intersperse(token_ids, self.pad_id)
+ if add_sos:
+ token_ids = [self.sos_id] + token_ids
+ if add_eos:
+ token_ids = token_ids + [self.eos_id]
 
  token_ids_list.append(token_ids)