add pretrained

.gitignore

@@ -14,11 +14,13 @@
 *.bcf
 *.toc
 *.sh
-data/ckpt
-!data/ckpt/vocoder/pretrained/**
+data/ckpt/*/*
 !data/ckpt/encoder/pretrained.pt
+!data/ckpt/vocoder/pretrained/
 wavs
 log
 !/docker-entrypoint.sh
 !/datasets_download/*.sh
 /datasets
+monotonic_align/build
+monotonic_align/monotonic_align

.vscode/launch.json

@@ -53,7 +53,7 @@
       "request": "launch",
       "program": "train.py",
       "console": "integratedTerminal",
-      "args": ["--type", "synth", "..\\audiodata\\SV2TTS\\synthesizer"]
+      "args": ["--type", "vits"]
     },
     {
       "name": "Python: PPG Convert",

data/ckpt/vocoder/pretrained/config_16k.json

@@ -0,0 +1,31 @@
+{
+    "resblock": "1",
+    "num_gpus": 0,
+    "batch_size": 16,
+    "learning_rate": 0.0002,
+    "adam_b1": 0.8,
+    "adam_b2": 0.99,
+    "lr_decay": 0.999,
+    "seed": 1234,
+
+    "upsample_rates": [5,5,4,2],
+    "upsample_kernel_sizes": [10,10,8,4],
+    "upsample_initial_channel": 512,
+    "resblock_kernel_sizes": [3,7,11],
+    "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
+
+    "segment_size": 6400,
+    "num_mels": 80,
+    "num_freq": 1025,
+    "n_fft": 1024,
+    "hop_size": 200,
+    "win_size": 800,
+
+    "sampling_rate": 16000,
+
+    "fmin": 0,
+    "fmax": 7600,
+    "fmax_for_loss": null,
+
+    "num_workers": 4
+}

models/synthesizer/preprocess.py

@@ -6,7 +6,7 @@ from pathlib import Path
 from tqdm import tqdm
 import numpy as np
 from models.encoder import inference as encoder
-from models.synthesizer.preprocess_audio import preprocess_general
+from models.synthesizer.preprocess_audio import preprocess_general, extract_emo
 from models.synthesizer.preprocess_transcript import preprocess_transcript_aishell3, preprocess_transcript_magicdata
 
 data_info = {
@@ -41,7 +41,7 @@ data_info = {
 
 def preprocess_dataset(datasets_root: Path, out_dir: Path, n_processes: int,
                            skip_existing: bool, hparams, no_alignments: bool, 
-                           dataset: str, emotion_extract = False):
+                           dataset: str, emotion_extract = False, encoder_model_fpath=None):
     dataset_info = data_info[dataset]
     # Gather the input directories
     dataset_root = datasets_root.joinpath(dataset)
@@ -77,7 +77,7 @@ def preprocess_dataset(datasets_root: Path, out_dir: Path, n_processes: int,
     speaker_dirs = list(chain.from_iterable(input_dir.glob("*") for input_dir in input_dirs))
     
     func = partial(dataset_info["speak_func"], out_dir=out_dir, skip_existing=skip_existing, 
-                   hparams=hparams, dict_info=dict_info, no_alignments=no_alignments, emotion_extract=emotion_extract)
+                   hparams=hparams, dict_info=dict_info, no_alignments=no_alignments, encoder_model_fpath=encoder_model_fpath)
     job = Pool(n_processes).imap(func, speaker_dirs)
     
     for speaker_metadata in tqdm(job, dataset, len(speaker_dirs), unit="speakers"):
@@ -110,6 +110,13 @@ def embed_utterance(fpaths, encoder_model_fpath):
     embed = encoder.embed_utterance(wav)
     np.save(embed_fpath, embed, allow_pickle=False)
     
+def _emo_extract_from_utterance(fpaths, hparams, skip_existing=False):
+    if skip_existing and fpaths.exists():
+        return
+    wav_fpath, emo_fpath = fpaths
+    wav = np.load(wav_fpath)
+    emo = extract_emo(np.expand_dims(wav, 0), hparams.sample_rate, True)
+    np.save(emo_fpath, emo.squeeze(0), allow_pickle=False)
  
 def create_embeddings(synthesizer_root: Path, encoder_model_fpath: Path, n_processes: int):
     wav_dir = synthesizer_root.joinpath("audio")
@@ -128,3 +135,21 @@ def create_embeddings(synthesizer_root: Path, encoder_model_fpath: Path, n_proce
     func = partial(embed_utterance, encoder_model_fpath=encoder_model_fpath)
     job = Pool(n_processes).imap(func, fpaths)
     list(tqdm(job, "Embedding", len(fpaths), unit="utterances"))
+
+def create_emo(synthesizer_root: Path, n_processes: int, skip_existing: bool, hparams):
+    wav_dir = synthesizer_root.joinpath("audio")
+    metadata_fpath = synthesizer_root.joinpath("train.txt")
+    assert wav_dir.exists() and metadata_fpath.exists()
+    emo_dir = synthesizer_root.joinpath("emo")
+    emo_dir.mkdir(exist_ok=True)
+    
+    # Gather the input wave filepath and the target output embed filepath
+    with metadata_fpath.open("r", encoding="utf-8") as metadata_file:
+        metadata = [line.split("|") for line in metadata_file]
+        fpaths = [(wav_dir.joinpath(m[0]), emo_dir.joinpath(m[0].replace("audio-", "emo-"))) for m in metadata]
+        
+    # TODO: improve on the multiprocessing, it's terrible. Disk I/O is the bottleneck here.
+    # Embed the utterances in separate threads
+    func = partial(_emo_extract_from_utterance, hparams=hparams, skip_existing=skip_existing)
+    job = Pool(n_processes).imap(func, fpaths)
+    list(tqdm(job, "Emo", len(fpaths), unit="utterances"))

models/synthesizer/preprocess_audio.py

@@ -13,7 +13,11 @@ import torch
 from transformers import Wav2Vec2Processor
 from .models.wav2emo import EmotionExtractorModel
 
-SAMPLE_RATE = 16000
+class PinyinConverter(NeutralToneWith5Mixin, DefaultConverter):
+    pass
+
+pinyin = Pinyin(PinyinConverter()).pinyin
+
 
 # load model from hub 
 device = 'cuda' if torch.cuda.is_available() else "cpu"
@@ -40,14 +44,8 @@ def extract_emo(
 
     return y
 
-class PinyinConverter(NeutralToneWith5Mixin, DefaultConverter):
-    pass
-
-pinyin = Pinyin(PinyinConverter()).pinyin
-
-
 def _process_utterance(wav: np.ndarray, text: str, out_dir: Path, basename: str, 
-                      skip_existing: bool, hparams, emotion_extract: bool):
+                      skip_existing: bool, hparams, encoder_model_fpath):
     ## FOR REFERENCE:
     # For you not to lose your head if you ever wish to change things here or implement your own
     # synthesizer.
@@ -69,6 +67,8 @@ def _process_utterance(wav: np.ndarray, text: str, out_dir: Path, basename: str,
 
     # Trim silence
     if hparams.trim_silence:
+        if not encoder.is_loaded():
+            encoder.load_model(encoder_model_fpath)
         wav = encoder.preprocess_wav(wav, normalize=False, trim_silence=True)
     
     # Skip utterances that are too short
@@ -109,7 +109,7 @@ def _split_on_silences(wav_fpath, words, hparams):
 
     return wav, res
 
-def preprocess_general(speaker_dir, out_dir: Path, skip_existing: bool, hparams, dict_info, no_alignments: bool, emotion_extract: bool):
+def preprocess_general(speaker_dir, out_dir: Path, skip_existing: bool, hparams, dict_info, no_alignments: bool, encoder_model_fpath: Path):
     metadata = []
     extensions = ["*.wav", "*.flac", "*.mp3"]
     for extension in extensions:
@@ -124,14 +124,9 @@ def preprocess_general(speaker_dir, out_dir: Path, skip_existing: bool, hparams,
             sub_basename = "%s_%02d" % (wav_fpath.name, 0)
             wav, text = _split_on_silences(wav_fpath, words, hparams)
             result = _process_utterance(wav, text, out_dir, sub_basename, 
-                                                skip_existing, hparams, emotion_extract)
+                                                skip_existing, hparams, encoder_model_fpath)
             if result is None:
                 continue
             wav_fpath_name, mel_fpath_name, embed_fpath_name, wav, mel_frames, text = result
-            emo_fpath = out_dir.joinpath("emo", "emo-%s.npy" % sub_basename)
-            skip_emo_extract = not emotion_extract or (skip_existing and emo_fpath.exists())
-            if not skip_emo_extract and wav is not None:
-                emo = extract_emo(np.expand_dims(wav, 0), hparams.sample_rate, True)
-                np.save(emo_fpath, emo.squeeze(0), allow_pickle=False)
             metadata.append([wav_fpath_name, mel_fpath_name, embed_fpath_name, len(wav), mel_frames, text])
     return [m for m in metadata if m is not None]

models/synthesizer/vits_dataset.py

@@ -1,10 +1,11 @@
 import os
 import random
 import numpy as np
+import torch.nn.functional as F
 import torch
 import torch.utils.data
 
-from utils.audio_utils import spectrogram1, load_wav_to_torch, spectrogram
+from utils.audio_utils import load_wav_to_torch, spectrogram
 from utils.util import intersperse
 from models.synthesizer.utils.text import text_to_sequence
 
@@ -51,21 +52,10 @@ class VitsDataset(torch.utils.data.Dataset):
         lengths = []
         
         # for audiopath, sid, text in self.audio_metadata:
-        sid = 0
-        spk_to_sid = {}
-        for wav_fpath, mel_fpath, embed_path, wav_length, mel_frames, text in self.audio_metadata:
+        for wav_fpath, mel_fpath, embed_path, wav_length, mel_frames, text, spkid in self.audio_metadata:
             if self.min_text_len <= len(text) and len(text) <= self.max_text_len:
-                # TODO: for magic data only
-                speaker_name = wav_fpath.split("_")[1]
-                # # TODO: for ai data only
-                # speaker_name = wav_fpath.split("-")[1][6:9]
-                if speaker_name not in spk_to_sid:
-                    sid += 1
-                    spk_to_sid[speaker_name] = sid
-                
-                audio_metadata_new.append([wav_fpath, mel_fpath, embed_path, wav_length, mel_frames, text, spk_to_sid[speaker_name]])
+                audio_metadata_new.append([wav_fpath, mel_fpath, embed_path, wav_length, mel_frames, text, spkid])
                 lengths.append(os.path.getsize(f'{self.datasets_root}{os.sep}audio{os.sep}{wav_fpath}') // (2 * self.hop_length))
-        print("found sid:%d", sid)
         self.audio_metadata = audio_metadata_new
         self.lengths = lengths
 
@@ -74,50 +64,31 @@ class VitsDataset(torch.utils.data.Dataset):
         wav_fpath, text, sid = audio_metadata[0], audio_metadata[5], audio_metadata[6]
         text = self.get_text(text)
 
-        # TODO: add original audio data root for loading
-        file_name = wav_fpath.split("_00")[0].split('-')[1]
-        spec, wav = self.get_audio(f'{self.datasets_root}{os.sep}..{os.sep}..{os.sep}magicdata{os.sep}train{os.sep}{"_".join(file_name.split("_")[:2])}{os.sep}{file_name}')
-
-        # spec, wav = self.get_audio(f'{self.datasets_root}{os.sep}audio{os.sep}{wav_fpath}')
+        spec, wav = self.get_audio(f'{self.datasets_root}{os.sep}audio{os.sep}{wav_fpath}')
         sid = self.get_sid(sid)
         emo = torch.FloatTensor(np.load(f'{self.datasets_root}{os.sep}emo{os.sep}{wav_fpath.replace("audio", "emo")}'))
         return (text, spec, wav, sid, emo)
 
     def get_audio(self, filename):
-        audio, sampling_rate = load_wav_to_torch(filename)
-        if sampling_rate != self.sampling_rate:
-            raise ValueError("{} {} SR doesn't match target {} SR".format(
-                sampling_rate, self.sampling_rate))
-        audio_norm = audio / self.max_wav_value
-        audio_norm = audio_norm.unsqueeze(0)
+        # Load preprocessed wav npy instead of reading from wav file
+        audio = torch.FloatTensor(np.load(filename))
+        audio_norm = audio.unsqueeze(0)
+
+        spec_filename = filename.replace(".wav", ".spec")
+        if os.path.exists(spec_filename):
+            spec = torch.load(spec_filename)
+        else:
             spec = spectrogram(audio_norm, self.filter_length,self.hop_length, self.win_length,
                 center=False)
+            torch.save(spec, spec_filename)
         spec = torch.squeeze(spec, 0)
         return spec, audio_norm
 
-        # print("Loading", filename)
-        # # audio = torch.FloatTensor(np.load(filename).astype(np.float32)) 
-        # audio = audio.unsqueeze(0)
-        # audio_norm = audio / self.max_wav_value
-        # audio_norm = audio_norm.unsqueeze(0)
-        # # spec_filename = filename.replace(".wav", ".spec.pt")
-        # # if os.path.exists(spec_filename):
-        # #     spec = torch.load(spec_filename)
-        # # else:
-        # #     spec = spectrogram(audio, self.filter_length,self.hop_length, self.win_length,
-        # #         center=False)
-        # #     spec = torch.squeeze(spec, 0)
-        # #     torch.save(spec, spec_filename)
-        # spec = spectrogram(audio, self.filter_length, self.hop_length, self.win_length,
-        #     center=False)
-        # spec = torch.squeeze(spec, 0)
-        # return spec, audio
-
     def get_text(self, text):
         if self.cleaned_text:
             text_norm = text_to_sequence(text, self.text_cleaners)
         if self.add_blank:
-            text_norm = intersperse(text_norm, 0)
+            text_norm = intersperse(text_norm, 0) # 在所有文本数值序列中的元素前后都补充一个0 - 不适用于中文
         text_norm = torch.LongTensor(text_norm)
         return text_norm
 
@@ -188,7 +159,7 @@ class VitsDatasetCollate():
             emo[i, :] = row[4]
 
         if self.return_ids:
-            return text_padded, text_lengths, spec_padded, spec_lengths, wav_padded, wav_lengths, sid, ids_sorted_decreasing
+            return text_padded, text_lengths, spec_padded, spec_lengths, wav_padded, wav_lengths, sid, ids_sorted_decreasing, emo
         return text_padded, text_lengths, spec_padded, spec_lengths, wav_padded, wav_lengths, sid, emo
 
 

monotonic_align/__init__.py

@@ -0,0 +1,19 @@
+import numpy as np
+import torch
+from .monotonic_align.core import maximum_path_c
+
+
+def maximum_path(neg_cent, mask):
+  """ Cython optimized version.
+  neg_cent: [b, t_t, t_s]
+  mask: [b, t_t, t_s]
+  """
+  device = neg_cent.device
+  dtype = neg_cent.dtype
+  neg_cent = neg_cent.data.cpu().numpy().astype(np.float32)
+  path = np.zeros(neg_cent.shape, dtype=np.int32)
+
+  t_t_max = mask.sum(1)[:, 0].data.cpu().numpy().astype(np.int32)
+  t_s_max = mask.sum(2)[:, 0].data.cpu().numpy().astype(np.int32)
+  maximum_path_c(path, neg_cent, t_t_max, t_s_max)
+  return torch.from_numpy(path).to(device=device, dtype=dtype)

monotonic_align/core.pyx

@@ -0,0 +1,42 @@
+cimport cython
+from cython.parallel import prange
+
+
+@cython.boundscheck(False)
+@cython.wraparound(False)
+cdef void maximum_path_each(int[:,::1] path, float[:,::1] value, int t_y, int t_x, float max_neg_val=-1e9) nogil:
+  cdef int x
+  cdef int y
+  cdef float v_prev
+  cdef float v_cur
+  cdef float tmp
+  cdef int index = t_x - 1
+
+  for y in range(t_y):
+    for x in range(max(0, t_x + y - t_y), min(t_x, y + 1)):
+      if x == y:
+        v_cur = max_neg_val
+      else:
+        v_cur = value[y-1, x]
+      if x == 0:
+        if y == 0:
+          v_prev = 0.
+        else:
+          v_prev = max_neg_val
+      else:
+        v_prev = value[y-1, x-1]
+      value[y, x] += max(v_prev, v_cur)
+
+  for y in range(t_y - 1, -1, -1):
+    path[y, index] = 1
+    if index != 0 and (index == y or value[y-1, index] < value[y-1, index-1]):
+      index = index - 1
+
+
+@cython.boundscheck(False)
+@cython.wraparound(False)
+cpdef void maximum_path_c(int[:,:,::1] paths, float[:,:,::1] values, int[::1] t_ys, int[::1] t_xs) nogil:
+  cdef int b = paths.shape[0]
+  cdef int i
+  for i in prange(b, nogil=True):
+    maximum_path_each(paths[i], values[i], t_ys[i], t_xs[i])

monotonic_align/setup.py

@@ -0,0 +1,9 @@
+from distutils.core import setup
+from Cython.Build import cythonize
+import numpy
+
+setup(
+  name = 'monotonic_align',
+  ext_modules = cythonize("core.pyx"),
+  include_dirs=[numpy.get_include()]
+)

pre.py

@@ -1,4 +1,4 @@
-from models.synthesizer.preprocess import create_embeddings, preprocess_dataset
+from models.synthesizer.preprocess import create_embeddings, preprocess_dataset, create_emo
 from models.synthesizer.hparams import hparams
 from pathlib import Path
 import argparse
@@ -64,7 +64,7 @@ if __name__ == "__main__":
                 "noise removal and is recommended. Please install and try again. If installation fails, "
                 "use --no_trim to disable this error message.")
     encoder_model_fpath = args.encoder_model_fpath
-    del args.no_trim, args.encoder_model_fpath
+    del args.no_trim
    
     args.hparams = hparams.parse(args.hparams)
     n_processes_embed = args.n_processes_embed
@@ -73,3 +73,5 @@ if __name__ == "__main__":
     
     create_embeddings(synthesizer_root=args.out_dir, n_processes=n_processes_embed, encoder_model_fpath=encoder_model_fpath)
     
+    if args.emotion_extract:
+        create_emo(synthesizer_root=args.out_dir, n_processes=n_processes_embed, skip_existing=args.skip_existing, hparams=args.hparams)

utils/audio_utils.py

@@ -17,28 +17,6 @@ def load_wav_to_torch(full_path):
   sampling_rate, data = read(full_path)
   return torch.FloatTensor(data.astype(np.float32)), sampling_rate
 
-def spectrogram1(y, n_fft, sampling_rate, hop_size, win_size, center=False):
-    if torch.min(y) < -1.:
-        print('min value is ', torch.min(y))
-    if torch.max(y) > 1.:
-        print('max value is ', torch.max(y))
-
-    global hann_window
-    dtype_device = str(y.dtype) + '_' + str(y.device)
-    wnsize_dtype_device = str(win_size) + '_' + dtype_device
-    if wnsize_dtype_device not in hann_window:
-        hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(dtype=y.dtype, device=y.device)
-
-    y = torch.nn.functional.pad(y.unsqueeze(1), (int((n_fft-hop_size)/2), int((n_fft-hop_size)/2)), mode='reflect')
-    y = y.squeeze(1)
-
-    spec = torch.stft(y, n_fft, hop_length=hop_size, win_length=win_size, window=hann_window[wnsize_dtype_device],
-                      center=center, pad_mode='reflect', normalized=False, onesided=True, return_complex=True)
-
-    spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6)
-    return spec
-
-
 def spectrogram(y, n_fft, hop_size, win_size, center=False):
     if torch.min(y) < -1.:
         print('min value is ', torch.min(y))

vits.ipynb

@@ -17,7 +17,7 @@
     "from models.synthesizer.utils.text import text_to_sequence\n",
     "\n",
     "\n",
-    "hps = load_hparams_json(\"data/ckpt/synthesizer/vits2/config.json\")\n",
+    "hps = load_hparams_json(\"data/ckpt/synthesizer/vits5/config.json\")\n",
     "print(hps.train)\n",
     "model = Vits(\n",
     "    len(symbols),\n",
@@ -27,7 +27,7 @@
     "    **hps[\"model\"])\n",
     "_ = model.eval()\n",
     "device = torch.device(\"cpu\")\n",
-    "checkpoint = torch.load(str(\"data/ckpt/synthesizer/vits2/G_120000.pth\"), map_location=device)\n",
+    "checkpoint = torch.load(str(\"data/ckpt/synthesizer/vits5/G_56000.pth\"), map_location=device)\n",
     "if \"model_state\" in checkpoint:\n",
     "    state = checkpoint[\"model_state\"]\n",
     "else:\n",
@@ -35,17 +35,17 @@
     "model.load_state_dict(state, strict=False)\n",
     "\n",
     "# 随机抽取情感参考音频的根目录\n",
-    "random_emotion_root = \"D:\\\\audiodata\\\\aidatatang_200zh\\\\corpus\\\\train\\\\G0017\"\n",
+    "random_emotion_root = \"D:\\\\audiodata\\\\SV2TTS\\\\synthesizer\\\\emo\\\\\"\n",
     "import random, re\n",
     "from pypinyin import lazy_pinyin, Style\n",
     "\n",
     "import os\n",
     "\n",
     "def tts(txt, emotion, sid=0):\n",
-    "    txt = \" \".join(lazy_pinyin(txt, style=Style.TONE3, neutral_tone_with_five=True))\n",
+    "    txt = \" \".join(lazy_pinyin(txt, style=Style.TONE3, neutral_tone_with_five=False))\n",
     "    text_norm = text_to_sequence(txt, hps[\"data\"][\"text_cleaners\"])\n",
-    "    if hps[\"data\"][\"add_blank\"]:\n",
-    "        text_norm = intersperse(text_norm, 0)\n",
+    "    # if hps[\"data\"][\"add_blank\"]:\n",
+    "    # text_norm = intersperse(text_norm, 0)\n",
     "    stn_tst = torch.LongTensor(text_norm)\n",
     "\n",
     "    with torch.no_grad(): #inference mode\n",
@@ -57,6 +57,13 @@
     "            import librosa\n",
     "            wav, sr = librosa.load(emotion, 16000)\n",
     "            emo = torch.FloatTensor(extract_emo(np.expand_dims(wav, 0), sr, embeddings=True))\n",
+    "        elif emotion == \"random_sample\":\n",
+    "            rand_emo = random.sample(os.listdir(random_emotion_root), 1)[0]\n",
+    "            print(rand_emo)\n",
+    "            emo = torch.FloatTensor(np.load(f\"{random_emotion_root}\\\\{rand_emo}\")).unsqueeze(0)\n",
+    "        elif emotion.endswith(\"npy\"):\n",
+    "            print(emotion)\n",
+    "            emo = torch.FloatTensor(np.load(f\"{random_emotion_root}\\\\{emotion}\")).unsqueeze(0)\n",
     "        else:\n",
     "            print(\"emotion参数不正确\")\n",
     "\n",
@@ -80,8 +87,55 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "txt = \"随机抽取的音频文件路径可以用于使用该情感合成其他句子\"\n",
-    "tts(txt, emotion='C:\\\\Users\\\\babys\\\\Desktop\\\\voicecollection\\\\secondround\\\\美玉.wav', sid=2)"
+    "txt = \"我们将其拓展到文本驱动数字人形象领域\"\n",
+    "#正常: \n",
+    "tts(txt, emotion='emo-T0055G4906S0052.wav_00.npy', sid=100)\n",
+    "#快速：emo-T0055G2323S0179.wav_00.npy\n",
+    "\n",
+    "#难过：\n",
+    "tts(txt, emotion='emo-15_4581_20170825202626.wav_00.npy', sid=100)\n",
+    "\n",
+    "#开心：T0055G2412S0498.wav\n",
+    "tts(txt, emotion='emo-T0055G2412S0498.wav_00.npy', sid=100)\n",
+    "\n",
+    "#愤怒 T0055G1371S0363.wav T0055G1344S0160.wav\n",
+    "tts(txt, emotion='emo-T0055G1344S0160.wav_00.npy', sid=100)\n",
+    "\n",
+    "#疲惫\n",
+    "tts(txt, emotion='emo-T0055G2294S0476.wav_00.npy', sid=100)\n",
+    "\n",
+    "#着急\n",
+    "tts(txt, emotion='emo-T0055G1671S0170.wav_00.npy', sid=100)\n",
+    "\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "txt = \"我们将其拓展到文本驱动数字人形象领域\"\n",
+    "tts(txt, emotion='random_sample', sid=100)\n",
+    "tts(txt, emotion='random_sample', sid=100)\n",
+    "tts(txt, emotion='random_sample', sid=100)\n",
+    "tts(txt, emotion='random_sample', sid=100)\n",
+    "tts(txt, emotion='random_sample', sid=100)\n",
+    "tts(txt, emotion='random_sample', sid=100)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "txt = \"我们将其拓展到文本驱动数字人形象领域\"\n",
+    "types = [\"平淡\", \"激动\", \"疲惫\", \"兴奋\", \"沮丧\", \"开心\"]\n",
+    "for t in types:\n",
+    "    print(t)\n",
+    "    tts(txt, emotion=f'C:\\\\Users\\\\babys\\\\Music\\\\{t}.wav', sid=100)\n",
+    "# tts(txt, emotion='D:\\\\audiodata\\\\aidatatang_200zh\\\\corpus\\\\train\\\\G1858\\\\T0055G1858S0342.wav', sid=5)"
    ]
   },
   {
@@ -128,7 +182,7 @@
     "        use_lws = False,                            # \"Fast spectrogram phase recovery using local weighted sums\"\n",
     "        symmetric_mels = True,                      # Sets mel range to [-max_abs_value, max_abs_value] if True,\n",
     "                                                    #               and [0, max_abs_value] if False\n",
-    "        trim_silence = True,                        # Use with sample_rate of 16000 for best results\n",
+    "        trim_silence = False,                        # Use with sample_rate of 16000 for best results\n",
     "\n",
     ")\n",
     "preprocess_dataset(datasets_root=datasets_root, \n",
@@ -137,7 +191,7 @@
     "        skip_existing=True, \n",
     "        hparams=hparams, \n",
     "        no_alignments=False, \n",
-    "        dataset=\"magicdata\", \n",
+    "        dataset=\"aidatatang_200zh\", \n",
     "        emotion_extract=True)"
    ]
   },
@@ -213,6 +267,14 @@
     "metadata_file.close()"
    ]
   },
+  {
+   "attachments": {},
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "从训练集中抽取10%作为测试集"
+   ]
+  },
   {
    "cell_type": "code",
    "execution_count": null,
@@ -220,36 +282,239 @@
    "outputs": [],
    "source": [
     "from pathlib import Path\n",
-    "import os\n",
-    "import shutil\n",
-    "emo_root = Path('../audiodata/SV2TTS/synthesizer').joinpath('emo')\n",
-    "# raw_root = Path('../audiodata/aidatatang_200zh/corpus/train')\n",
-    "# emo_file_list = emo_root.glob(\"**/*.npy\")\n",
-    "# for emo_file in emo_file_list:\n",
-    "#     if emo_file.name.endswith('wav__00.npy'):\n",
-    "#         folder = emo_file.parent\n",
-    "#         os.rename(emo_file, folder.joinpath(emo_file.name.replace(\"__00\", \"_00\")))\n",
-    "    # shutil.move(emo_file, emo_root.joinpath(emo_file.name))\n",
-    "\n",
     "root = Path('../audiodata/SV2TTS/synthesizer')\n",
-    "dict_info = []\n",
+    "dict_info1 = []\n",
+    "dict_info2 = []\n",
+    "count = 1\n",
     "with open(root.joinpath(\"train.txt\"), \"r\", encoding=\"utf-8\") as dict_meta:\n",
     "    for raw in dict_meta:\n",
     "        if not raw:\n",
     "            continue\n",
-    "        v = raw.split(\"|\")[0].replace(\"audio\",\"emo\")\n",
-    "        emo_fpath = root.joinpath(\"emo\").joinpath(v)\n",
-    "        if emo_fpath.exists():\n",
-    "            dict_info.append(raw)\n",
-    "        # else:\n",
-    "        #     print(emo_fpath)\n",
+    "        if count % 10 == 0:\n",
+    "            dict_info2.append(raw)\n",
+    "        else:\n",
+    "            dict_info1.append(raw)\n",
+    "        count += 1\n",
     "# Iterate over each wav\n",
-    "meta2 = Path('../audiodata/SV2TTS/synthesizer/train2.txt')\n",
+    "meta1 = Path('../audiodata/SV2TTS/synthesizer/train1.txt')\n",
+    "metadata_file = meta1.open(\"w\", encoding=\"utf-8\")\n",
+    "for new_info in dict_info1:\n",
+    "    metadata_file.write(new_info)\n",
+    "metadata_file.close()\n",
+    "\n",
+    "meta2 = Path('../audiodata/SV2TTS/synthesizer/eval.txt')\n",
     "metadata_file = meta2.open(\"w\", encoding=\"utf-8\")\n",
-    "for new_info in dict_info:\n",
+    "for new_info in dict_info2:\n",
     "    metadata_file.write(new_info)\n",
     "metadata_file.close()"
    ]
+  },
+  {
+   "attachments": {},
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "evaluation"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from pathlib import Path\n",
+    "root = Path('../audiodata/SV2TTS/synthesizer')\n",
+    "spks = []\n",
+    "spk_id = {}\n",
+    "rows = []\n",
+    "with open(root.joinpath(\"eval.txt\"), \"r\", encoding=\"utf-8\") as dict_meta:\n",
+    "    for raw in dict_meta:\n",
+    "        speaker_name = raw.split(\"-\")[1][6:10]\n",
+    "        if speaker_name not in spk_id:\n",
+    "            spks.append(speaker_name)\n",
+    "            spk_id[speaker_name] = 1\n",
+    "        rows.append(raw)\n",
+    "i = 0\n",
+    "spks.sort()\n",
+    "\n",
+    "for sp in spks:\n",
+    "    spk_id[sp] = str(i)\n",
+    "    i = i + 1\n",
+    "print(len(spks))\n",
+    "meta2 = Path('../audiodata/SV2TTS/synthesizer/eval2.txt')\n",
+    "metadata_file = meta2.open(\"w\", encoding=\"utf-8\")\n",
+    "for row in rows:\n",
+    "    speaker_n = row.split(\"-\")[1][6:10]\n",
+    "    metadata_file.write(row.strip()+\"|\"+spk_id[speaker_n]+\"\\n\")\n",
+    "metadata_file.close()\n"
+   ]
+  },
+  {
+   "attachments": {},
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "[Not Recommended]\n",
+    "Try to transcript map to detailed format:\n",
+    "ni3 hao3 -> n i3 <pad> h ao3\n",
+    "\n",
+    "After couple of tests, I think this method will not improve the quality of result and may cause the crash of monotonic alignment."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "\n",
+    "from pathlib import Path\n",
+    "datasets_root = Path(\"../audiodata/SV2TTS/synthesizer/\")\n",
+    "\n",
+    "dictionary_fp = Path(\"../audiodata/ProDiff/processed/mandarin_pinyin.dict\")\n",
+    "dict_map = {}\n",
+    "for l in open(dictionary_fp, encoding='utf-8').readlines():\n",
+    "    item = l.split(\"\\t\")\n",
+    "    dict_map[item[0]] = item[1].replace(\"\\n\",\"\")\n",
+    "\n",
+    "with datasets_root.joinpath('train2.txt').open(\"w+\", encoding='utf-8') as f:\n",
+    "    for l in open(datasets_root.joinpath('train.txt'), encoding='utf-8').readlines():\n",
+    "        items = l.strip().replace(\"\\n\",\"\").replace(\"\\t\",\" \").split(\"|\")\n",
+    "        phs_str = \"\"\n",
+    "        for word in items[5].split(\" \"):\n",
+    "            if word in dict_map:\n",
+    "                phs_str += dict_map[word] \n",
+    "            else:\n",
+    "                phs_str += word\n",
+    "            phs_str += \" _ \"\n",
+    "        items[5] = phs_str\n",
+    "        # if not os.path.exists(mfa_input_root.joinpath('train.txt')):\n",
+    "        #     with open(mfa_input_root.joinpath(fileName + 'lab'), 'w+', encoding=\"utf-8\") as f:\n",
+    "        f.write(\"|\".join(items) + \"\\n\")"
+   ]
+  },
+  {
+   "attachments": {},
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "预处理后的数据可视化"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import matplotlib.pyplot as plt\n",
+    "import librosa.display\n",
+    "import librosa, torch\n",
+    "import numpy as np\n",
+    "from utils.audio_utils import spectrogram, mel_spectrogram, load_wav_to_torch, spec_to_mel\n",
+    "\n",
+    "# x, sr = librosa.load(\"D:\\audiodata\\SV2TTS\\synthesizer\\audio\\audio-T0055G2333S0196.wav_00.npy\")\n",
+    "x = np.load(\"D:\\\\audiodata\\\\SV2TTS\\\\synthesizer\\\\audio\\\\audio-T0055G1858S0342.wav_00.npy\")\n",
+    "\n",
+    "plt.figure(figsize=(14, 5))\n",
+    "librosa.display.waveplot(x)\n",
+    "\n",
+    "X = librosa.stft(x)\n",
+    "Xdb = librosa.amplitude_to_db(abs(X))\n",
+    "plt.figure(figsize=(14, 5))\n",
+    "librosa.display.specshow(Xdb,  x_axis='time', y_axis='hz')\n",
+    "\n",
+    "# spectrogram = np.load(\"D:\\\\audiodata\\\\SV2TTS\\\\synthesizer\\\\mels\\\\mel-T0055G1858S0342.wav_00.npy\")\n",
+    "audio = torch.from_numpy(x.astype(np.float32))\n",
+    "\n",
+    "# audio, sampling_rate = load_wav_to_torch(\"D:\\\\audiodata\\\\aidatatang_200zh\\\\corpus\\\\train\\\\G1858\\\\T0055G1858S0342.wav\")\n",
+    "# audio_norm = audio / 32768.0\n",
+    "audio_norm = audio.unsqueeze(0)\n",
+    "spec = spectrogram(audio_norm, 1024, 256, 1024, center=False)\n",
+    "# spec = spec_to_mel()\n",
+    "spec = torch.squeeze(spec, 0)\n",
+    "mel = spec_to_mel(spec, 1024, 80, 16000, 0, None)\n",
+    "\n",
+    "fig = plt.figure(figsize=(10, 8))\n",
+    "ax2 = fig.add_subplot(211)\n",
+    "im = ax2.imshow(mel, interpolation=\"none\")"
+   ]
+  },
+  {
+   "attachments": {},
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "情感聚类"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "\n",
+    "# from sklearn import metrics\n",
+    "# from sklearn.mixture import GaussianMixture  # 高斯混合模型\n",
+    "import os\n",
+    "import numpy as np\n",
+    "import librosa\n",
+    "import IPython.display as ipd\n",
+    "from random import sample\n",
+    "\n",
+    "embs = []\n",
+    "wavnames = []\n",
+    "emo_root_path = \"D:\\\\audiodata\\\\SV2TTS\\\\synthesizer\\\\emo\\\\\"\n",
+    "wav_root_path = \"D:\\\\audiodata\\\\aidatatang_200zh\\\\corpus\\\\train\\\\\"\n",
+    "for idx, emo_fpath in enumerate(sample(os.listdir(emo_root_path), 10000)):\n",
+    "    if emo_fpath.endswith(\".npy\") and emo_fpath.startswith(\"emo-T\"):\n",
+    "        embs.append(np.expand_dims(np.load(emo_root_path + emo_fpath), axis=0))\n",
+    "        wav_fpath = wav_root_path + emo_fpath[9:14] + \"\\\\\" + emo_fpath.split(\"_00\")[0][4:]\n",
+    "        wavnames.append(wav_fpath)\n",
+    "print(len(embs))\n",
+    "\n",
+    "\n",
+    "x = np.concatenate(embs, axis=0)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# 聚类算法类的数量\n",
+    "n_clusters = 20\n",
+    "from sklearn.cluster import *\n",
+    "# model = KMeans(n_clusters=n_clusters, random_state=10)\n",
+    "# model = DBSCAN(eps=0.002, min_samples=2)\n",
+    "# 可以自行尝试各种不同的聚类算法\n",
+    "# model = Birch(n_clusters= n_clusters, threshold= 0.2)\n",
+    "# model = SpectralClustering(n_clusters=n_clusters)\n",
+    "model = AgglomerativeClustering(n_clusters= n_clusters)\n",
+    "import random\n",
+    "\n",
+    "y_predict = model.fit_predict(x)\n",
+    "\n",
+    "def disp(wavname):\n",
+    "    wav, sr =librosa.load(wavname, 16000)\n",
+    "    display(ipd.Audio(wav, rate=sr))\n",
+    "\n",
+    "classes=[[] for i in range(y_predict.max()+1)]\n",
+    "\n",
+    "for idx, wavname in enumerate(wavnames):\n",
+    "    classes[y_predict[idx]].append(wavname)\n",
+    "\n",
+    "for i in range(y_predict.max()+1):\n",
+    "    print(\"类别:\", i, \"本类中样本数量:\", len(classes[i]))\n",
+    "    \"\"\"每一个类只预览2条音频\"\"\"\n",
+    "    for j in range(2):\n",
+    "        idx = random.randint(0, len(classes[i]) - 1)\n",
+    "        print(classes[i][idx])\n",
+    "        disp(classes[i][idx])"
+   ]
   }
  ],
  "metadata": {