Update requirements and README.md of lstm_tacotron2

modules/audio/audio_classification/PANNs/cnn10/module.py

@@ -31,7 +31,7 @@ from paddlehub.utils.log import logger
     name="panns_cnn10",
     version="1.0.0",
     summary="",
-    author="Baidu",
+    author="paddlepaddle",
     author_email="",
     type="audio/sound_classification",
     meta=AudioClassifierModule)

modules/audio/audio_classification/PANNs/cnn14/module.py

@@ -31,7 +31,7 @@ from paddlehub.utils.log import logger
     name="panns_cnn14",
     version="1.0.0",
     summary="",
-    author="Baidu",
+    author="paddlepaddle",
     author_email="",
     type="audio/sound_classification",
     meta=AudioClassifierModule)

modules/audio/audio_classification/PANNs/cnn6/module.py

@@ -31,7 +31,7 @@ from paddlehub.utils.log import logger
     name="panns_cnn6",
     version="1.0.0",
     summary="",
-    author="Baidu",
+    author="paddlepaddle",
     author_email="",
     type="audio/sound_classification",
     meta=AudioClassifierModule)

modules/audio/voice_cloning/lstm_tacotron2/README.md

+```shell
+$ hub install lstm_tacotron2==1.0.0
+```
+
+## 概述
+
+声音克隆是指使用特定的音色，结合文字的读音合成音频，使得合成后的音频具有目标说话人的特征，从而达到克隆的目的。
+
+在训练语音克隆模型时，目标音色作为Speaker Encoder的输入，模型会提取这段语音的说话人特征（音色）作为Speaker Embedding。接着，在训练模型重新合成此类音色的语音时，除了输入的目标文本外，说话人的特征也将成为额外条件加入模型的训练。
+
+在预测时，选取一段新的目标音色作为Speaker Encoder的输入，并提取其说话人特征，最终实现输入为一段文本和一段目标音色，模型生成目标音色说出此段文本的语音片段。
+
+![](https://ai-studio-static-online.cdn.bcebos.com/982ab955b87244d3bae3b003aff8e28d9ec159ff0d6246a79757339076dfe7d4)
+
+`lstm_tacotron2`是一个支持中文的语音克隆模型，分别使用了LSTMSpeakerEncoder、Tacotron2和WaveFlow模型分别用于语音特征提取、目标音频特征合成和语音波形转换。
+
+关于模型的详请可参考[Parakeet](https://github.com/PaddlePaddle/Parakeet/tree/release/v0.3/parakeet/models)。
+
+
+## API
+
+```python
+def __init__(speaker_audio: str = None,
+             output_dir: str = './')
+```
+初始化module，可配置模型的目标音色的音频文件和输出的路径。
+
+**参数**
+- `speaker_audio`(str): 目标说话人语音音频文件(*.wav)的路径，默认为None(使用默认的女声作为目标音色)。
+- `output_dir`(str): 合成音频的输出文件，默认为当前目录。
+
+
+```python
+def get_speaker_embedding()
+```
+获取模型的目标说话人特征。
+
+**返回**
+* `results`(numpy.ndarray): 长度为256的numpy数组，代表目标说话人的特征。
+
+```python
+def set_speaker_embedding(speaker_audio: str)
+```
+设置模型的目标说话人特征。
+
+**参数**
+- `speaker_audio`(str): 必填，目标说话人语音音频文件(*.wav)的路径。
+
+```python
+def generate(data: List[str], batch_size: int = 1, use_gpu: bool = False):
+```
+根据输入文字，合成目标说话人的语音音频文件。
+
+**参数**
+- `data`(List[str]): 必填，目标音频的内容文本列表，目前只支持中文，不支持添加标点符号。
+- `batch_size`(int): 可选，模型合成语音时的batch_size，默认为1。
+- `use_gpu`(bool): 是否使用gpu执行计算，默认为False。
+
+
+**代码示例**
+
+```python
+import paddlehub as hub
+
+model = hub.Module(name='lstm_tacotron2', output_dir='./', speaker_audio='/data/man.wav')  # 指定目标音色音频文件
+texts = [
+    '语音的表现形式在未来将变得越来越重要$',
+    '今天的天气怎么样$',  ]
+wavs = model.generate(texts, use_gpu=True)
+
+for text, wav in zip(texts, wavs):
+    print('='*30)
+    print(f'Text: {text}')
+    print(f'Wav: {wav}')
+```
+
+输出
+```
+==============================
+Text: 语音的表现形式在未来将变得越来越重要$
+Wav: /data/1.wav
+==============================
+Text: 今天的天气怎么样$
+Wav: /data/2.wav
+```
+
+
+## 查看代码
+
+https://github.com/PaddlePaddle/Parakeet
+
+## 依赖
+
+paddlepaddle >= 2.0.0
+
+paddlehub >= 2.1.0
+
+## 更新历史
+
+* 1.0.0
+
+  初始发布

modules/audio/voice_cloning/lstm_tacotron2/module.py

@@ -23,14 +23,9 @@ from paddlehub.env import MODULE_HOME
 from paddlehub.module.module import moduleinfo
 from paddlehub.utils.log import logger
 from paddlenlp.data import Pad
-import soundfile as sf
-
-if not importlib.util.find_spec('parakeet'):
-    raise ImportError('The module requires additional dependencies: "parakeet".\n'
-                      'You can install parakeet via "git clone https://github.com'
-                      '/PaddlePaddle/Parakeet -b release/v0.3 && pip install -e Parakeet"')
 from parakeet.models import ConditionalWaveFlow, Tacotron2
 from parakeet.models.lstm_speaker_encoder import LSTMSpeakerEncoder
+import soundfile as sf
 
 from .audio_processor import SpeakerVerificationPreprocessor
 from .chinese_g2p import convert_sentence
@@ -41,7 +36,7 @@ from .preprocess_transcription import voc_phones, voc_tones, phone_pad_token, to
     name="lstm_tacotron2",
     version="1.0.0",
     summary="",
-    author="Baidu",
+    author="paddlepaddle",
     author_email="",
     type="audio/voice_cloning",
 )

modules/audio/voice_cloning/lstm_tacotron2/preprocess_transcription.py

@@ -19,7 +19,6 @@ import re
 
 from parakeet.frontend import Vocab
 import tqdm
-import yaml
 
 zh_pattern = re.compile("[\u4e00-\u9fa5]")
 
@@ -180,75 +179,3 @@ def split_syllable(syllable: str):
         phones.append(syllable)
         tones.append(tone)
     return phones, tones
-
-
-def load_aishell3_transcription(line: str):
-    sentence_id, pinyin, text = line.strip().split("|")
-    syllables = pinyin.strip().split()
-
-    results = []
-
-    for syllable in syllables:
-        if syllable in _pauses:
-            results.append(syllable)
-        elif not ernized(syllable):
-            results.append(syllable)
-        else:
-            results.append(syllable[:-2] + syllable[-1])
-            results.append('&r5')
-
-    phones = []
-    tones = []
-    for syllable in results:
-        p, t = split_syllable(syllable)
-        phones.extend(p)
-        tones.extend(t)
-    for p in phones:
-        assert p in _phones, p
-    return {"sentence_id": sentence_id, "text": text, "syllables": results, "phones": phones, "tones": tones}
-
-
-def process_aishell3(dataset_root, output_dir):
-    dataset_root = Path(dataset_root).expanduser()
-    output_dir = Path(output_dir).expanduser()
-    output_dir.mkdir(parents=True, exist_ok=True)
-
-    prosody_label_path = dataset_root / "label_train-set.txt"
-    with open(prosody_label_path, 'rt') as f:
-        lines = [line.strip() for line in f]
-
-    records = lines[5:]
-
-    processed_records = []
-    for record in tqdm.tqdm(records):
-        new_record = load_aishell3_transcription(record)
-        processed_records.append(new_record)
-        print(new_record)
-
-    with open(output_dir / "metadata.pickle", 'wb') as f:
-        pickle.dump(processed_records, f)
-
-    with open(output_dir / "metadata.yaml", 'wt', encoding="utf-8") as f:
-        yaml.safe_dump(processed_records, f, default_flow_style=None, allow_unicode=True)
-
-    print("metadata done!")
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser(
-        description="Preprocess transcription of AiShell3 and save them in a compact file(yaml and pickle).")
-    parser.add_argument(
-        "--input",
-        type=str,
-        default="~/datasets/aishell3/train",
-        help="path of the training dataset,(contains a label_train-set.txt).")
-    parser.add_argument(
-        "--output",
-        type=str,
-        help="the directory to save the processed transcription."
-        "If not provided, it would be the same as the input.")
-    args = parser.parse_args()
-    if args.output is None:
-        args.output = args.input
-
-    process_aishell3(args.input, args.output)

modules/audio/voice_cloning/lstm_tacotron2/requirements.txt

-librosa
-nltk
-pypinyin
-scipy
-soundfile
-webrtcvad
-yaml
+paddle-parakeet

modules/text/text_generation/plato-mini/module.py

@@ -30,7 +30,7 @@ from plato_mini.utils import select_response
     name="plato-mini",
     version="1.0.0",
     summary="",
-    author="PaddlePaddle",
+    author="paddlepaddle",
     author_email="",
     type="nlp/text_generation",
 )

modules/text/text_generation/unified_transformer-12L-cn-luge/module.py

@@ -30,7 +30,7 @@ from unified_transformer_12L_cn_luge.utils import select_response
     name="unified_transformer_12L_cn_luge",
     version="1.0.0",
     summary="",
-    author="PaddlePaddle",
+    author="paddlepaddle",
     author_email="",
     type="nlp/text_generation",
 )

modules/text/text_generation/unified_transformer-12L-cn/module.py

@@ -30,7 +30,7 @@ from unified_transformer_12L_cn.utils import select_response
     name="unified_transformer_12L_cn",
     version="1.0.0",
     summary="",
-    author="PaddlePaddle",
+    author="paddlepaddle",
     author_email="",
     type="nlp/text_generation",
 )