[TTS][ZH] add Chinese TTS recipes based on IPA symbol sets. (#6893)

* [TTS] add Chinese TTS recipe based on IPA.
* add new pinyin and ipa dictionaries with 36 finals.
* add yaml configs for 24-final pinyin and ipa.
* add copyright header
* add a directory level 24finals to discriminate from 36 finals.

Signed-off-by: Xuesong Yang <[email protected]>

* unify configs into a single one and add detailed comments providing supported candidates.

Signed-off-by: Xuesong Yang <[email protected]>

* choose 36-final IPA as default phoneme dict

Signed-off-by: Xuesong Yang <[email protected]>

---------

Signed-off-by: Xuesong Yang <[email protected]>

Author: XuesongYang

examples/tts/conf/zh/fastpitch_align_22050.yaml

@@ -1,5 +1,5 @@
-# This config contains the default values for training FastPitch model with aligner using 22KHz sampling 
-# rate. If you want to train model on other dataset, you can change config values according to your dataset. 
+# This config contains the default values for training FastPitch model with aligner using 22KHz sampling
+# rate. If you want to train model on other dataset, you can change config values according to your dataset.
 # Most dataset-specific arguments are in the head of the config file, see below.
 
 name: FastPitch
@@ -28,7 +28,13 @@ lowfreq: 0
 highfreq: null
 window: hann
 
-phoneme_dict_path: "scripts/tts_dataset_files/zh/pinyin_dict_nv_22.10.txt"
+# There are four candidates of `phoneme_dict_path` provided for Chinese as shown below,
+#     1) 24-final Pinyin: "scripts/tts_dataset_files/zh/24finals/pinyin_dict_nv_22.10.txt",
+#     2) IPA converted from 24-final Pinyin: "scripts/tts_dataset_files/zh/24finals/ipa_dict_nv23.05.txt",
+#     3) 36-final Pinyin: "scripts/tts_dataset_files/zh/36finals/pinyin_dict_nv23.05.txt",
+#     4) (default) IPA converted from 36-final Pinyin: "scripts/tts_dataset_files/zh/36finals/ipa_dict_nv23.05.txt"
+# Suggest to choose IPA symbol set converted from 36-final Pinyin because better audio quality were observed.
+phoneme_dict_path: "scripts/tts_dataset_files/zh/36finals/ipa_dict_nv23.05.txt"
 
 model:
   learn_alignment: true
@@ -73,6 +79,11 @@ model:
       _target_: nemo.collections.tts.g2p.models.zh_cn_pinyin.ChineseG2p
       phoneme_dict: ${phoneme_dict_path}
       word_segmenter: jieba # Only jieba is supported now.
+      phoneme_prefix: ""
+      phoneme_case: lower
+      tone_prefix: "#"
+      ascii_letter_prefix: ""
+      ascii_letter_case: upper
 
   train_ds:
     dataset:

nemo/collections/common/tokenizers/text_to_speech/tts_tokenizers.py

@@ -694,7 +694,10 @@ def __init__(
         pad_with_space=False,
         text_preprocessing_func=chinese_text_preprocessing,
     ):
-        """Chinese phoneme-based tokenizer.
+        """
+        Chinese phoneme-based tokenizer.
+        Note: This tokenizer for now covers Chinese phonemes/tones and English letters because our dataset contains
+              both Chinese and English graphemes.
         Args:
             g2p: Grapheme to phoneme module.
             punct: Whether to reserve grapheme for basic punctuation or not.

scripts/dataset_processing/tts/sfbilingual/ds_conf/ds_for_fastpitch_align.yaml

@@ -3,7 +3,7 @@ name: "ds_for_fastpitch_align"
 manifest_filepath: "train_manifest.json"
 sup_data_path: "sup_data"
 sup_data_types: [ "align_prior_matrix", "pitch" ]
-phoneme_dict_path: "scripts/tts_dataset_files/zh/pinyin_dict_nv_22.10.txt"
+phoneme_dict_path: "scripts/tts_dataset_files/zh/24finals/pinyin_dict_nv_22.10.txt"
 
 dataset:
   _target_: nemo.collections.tts.data.dataset.TTSDataset