init commit

egs/vctk/TTS/local/prepare_tokens_vctk.py

@@ -24,9 +24,9 @@ This file reads the texts in given manifest and save the new cuts with phoneme t
 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
 from tqdm.auto import tqdm
 
 
@@ -37,17 +37,20 @@ def prepare_tokens_vctk():
     partition = "all"
 
     cut_set = load_manifest(output_dir / f"{prefix}_cuts_{partition}.{suffix}")
-    g2p = g2p_en.G2p()
 
     new_cuts = []
     for cut in tqdm(cut_set):
         # Each cut only contains one supervision
-        assert len(cut.supervisions) == 1, len(cut.supervisions)
+        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
-        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/vctk/TTS/prepare.sh

@@ -78,6 +78,13 @@ fi
 
 if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
   log "Stage 3: Prepare phoneme tokens for VCTK"
+  # 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/.vctk_with_token.done ]; then
     ./local/prepare_tokens_vctk.py
     mv data/spectrogram/vctk_cuts_with_tokens_all.jsonl.gz \
@@ -111,14 +118,15 @@ 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
-  #   - espnet_tts_frontend, `pip install espnet_tts_frontend`, refer to https://github.com/espnet/espnet_tts_frontend/
+  #   - 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/vctk_cuts_train.jsonl.gz \
-      --tokens data/tokens.txt
+    ./local/prepare_token_file.py --tokens data/tokens.txt
   fi
 fi
 

egs/vctk/TTS/vits/export-onnx.py

@@ -231,8 +231,7 @@ def main():
     params.update(vars(args))
 
     tokenizer = Tokenizer(params.tokens)
-    params.blank_id = tokenizer.blank_id
-    params.oov_id = tokenizer.oov_id
+    params.blank_id = tokenizer.pad_id
     params.vocab_size = tokenizer.vocab_size
 
     with open(args.speakers) as f:

egs/vctk/TTS/vits/infer.py

@@ -135,14 +135,16 @@ def infer_dataset(
             batch_size = len(batch["tokens"])
 
             tokens = batch["tokens"]
-            tokens = tokenizer.tokens_to_token_ids(tokens)
+            tokens = tokenizer.tokens_to_token_ids(
+                tokens, intersperse_blank=True, add_sos=True, add_eos=True
+            )
             tokens = k2.RaggedTensor(tokens)
             row_splits = tokens.shape.row_splits(1)
             tokens_lens = row_splits[1:] - row_splits[:-1]
             tokens = tokens.to(device)
             tokens_lens = tokens_lens.to(device)
             # tensor of shape (B, T)
-            tokens = tokens.pad(mode="constant", padding_value=tokenizer.blank_id)
+            tokens = tokens.pad(mode="constant", padding_value=tokenizer.pad_id)
             speakers = (
                 torch.Tensor([speaker_map[sid] for sid in batch["speakers"]])
                 .int()
@@ -214,8 +216,7 @@ def main():
         device = torch.device("cuda", 0)
 
     tokenizer = Tokenizer(params.tokens)
-    params.blank_id = tokenizer.blank_id
-    params.oov_id = tokenizer.oov_id
+    params.blank_id = tokenizer.pad_id
     params.vocab_size = tokenizer.vocab_size
 
     # we need cut ids to display recognition results.

egs/vctk/TTS/vits/test_onnx.py

@@ -1,6 +1,7 @@
 #!/usr/bin/env python3
 #
-# Copyright      2023 Xiaomi Corporation     (Author: Zengwei Yao)
+# Copyright   2023-2024   Xiaomi Corporation     (Author: Zengwei Yao,
+#                                                         Zengrui Jin,)
 #
 # See ../../../../LICENSE for clarification regarding multiple authors
 #
@@ -122,7 +123,9 @@ def main():
     model = OnnxModel(args.model_filename)
 
     text = "I went there to see the land, the people and how their system works, end quote."
-    tokens = tokenizer.texts_to_token_ids([text])
+    tokens = tokenizer.texts_to_token_ids(
+        [text], intersperse_blank=True, add_sos=True, add_eos=True
+    )
     tokens = torch.tensor(tokens)  # (1, T)
     tokens_lens = torch.tensor([tokens.shape[1]], dtype=torch.int64)  # (1, T)
     speaker = torch.tensor([1], dtype=torch.int64)  # (1, )

egs/vctk/TTS/vits/train.py

@@ -342,14 +342,16 @@ def prepare_input(
         torch.Tensor([speaker_map[sid] for sid in batch["speakers"]]).int().to(device)
     )
 
-    tokens = tokenizer.tokens_to_token_ids(tokens)
+    tokens = tokenizer.texts_to_token_ids(
+        tokens, intersperse_blank=True, add_sos=True, add_eos=True
+    )
     tokens = k2.RaggedTensor(tokens)
     row_splits = tokens.shape.row_splits(1)
     tokens_lens = row_splits[1:] - row_splits[:-1]
     tokens = tokens.to(device)
     tokens_lens = tokens_lens.to(device)
     # a tensor of shape (B, T)
-    tokens = tokens.pad(mode="constant", padding_value=tokenizer.blank_id)
+    tokens = tokens.pad(mode="constant", padding_value=tokenizer.pad_id)
 
     return audio, audio_lens, features, features_lens, tokens, tokens_lens, speakers
 
@@ -812,8 +814,7 @@ def run(rank, world_size, args):
     logging.info(f"Device: {device}")
 
     tokenizer = Tokenizer(params.tokens)
-    params.blank_id = tokenizer.blank_id
-    params.oov_id = tokenizer.oov_id
+    params.blank_id = tokenizer.pad_id
     params.vocab_size = tokenizer.vocab_size
 
     vctk = VctkTtsDataModule(args)