Merge pull request #1436 from yt605155624/rename_tacotron2

[TTS]Rename tacotron2

CHANGELOG.md

 # Changelog
 
-Date: 2022-1-19, Author: yt605155624.  
-Add features to: T2S:  
-  - Add csmsc Tacotron2.  
+Date: 2022-1-29, Author: yt605155624.
+Add features to: T2S:
+  - Update aishell3 vc0 with new Tacotron2.
+  - PRLink: https://github.com/PaddlePaddle/PaddleSpeech/pull/1419
+
+Date: 2022-1-29, Author: yt605155624.
+Add features to: T2S:
+  - Add ljspeech Tacotron2.
+  - PRLink: https://github.com/PaddlePaddle/PaddleSpeech/pull/1416
+
+Date: 2022-1-24, Author: yt605155624.
+Add features to: T2S:
+  - Add csmsc WaveRNN.
+  - PRLink: https://github.com/PaddlePaddle/PaddleSpeech/pull/1379
+
+Date: 2022-1-19, Author: yt605155624.
+Add features to: T2S:
+  - Add csmsc Tacotron2.
   - PRLink: https://github.com/PaddlePaddle/PaddleSpeech/pull/1314
 
 
 Date: 2022-1-10, Author: Jackwaterveg.  
-Add features to: CLI:  
-  - Support English (librispeech/asr1/transformer).  
+Add features to: CLI:
+  - Support English (librispeech/asr1/transformer).
   - Support choosing `decode_method` for conformer and transformer models.  
   - Refactor the config, using the unified config.  
   - PRLink: https://github.com/PaddlePaddle/PaddleSpeech/pull/1297
@@ -16,8 +31,8 @@ Add features to: CLI:
 ***
 
 Date: 2022-1-17, Author: Jackwaterveg.  
-Add features to: CLI:  
-  - Support deepspeech2 online/offline model(aishell). 
+Add features to: CLI:
+  - Support deepspeech2 online/offline model(aishell).
   - PRLink: https://github.com/PaddlePaddle/PaddleSpeech/pull/1356
 
 ***

README.md

@@ -317,14 +317,15 @@ PaddleSpeech supports a series of most popular models. They are summarized in [r
     </tr>
     <tr>
       <td rowspan="4">Acoustic Model</td>
-      <td >Tacotron2</td>
-      <td rowspan="2" >LJSpeech</td>
+      <td>Tacotron2</td>
+      <td>LJSpeech / CSMSC</td>
       <td>
-      <a href = "./examples/ljspeech/tts0">tacotron2-ljspeech</a>
+      <a href = "./examples/ljspeech/tts0">tacotron2-ljspeech</a> / <a href = "./examples/csmsc/tts0">tacotron2-csmsc</a>
       </td>
     </tr>
     <tr>
       <td>Transformer TTS</td>
+      <td>LJSpeech</td>
       <td>
       <a href = "./examples/ljspeech/tts1">transformer-ljspeech</a>
       </td>
@@ -344,7 +345,7 @@ PaddleSpeech supports a series of most popular models. They are summarized in [r
       </td>
     </tr>
    <tr>
-      <td rowspan="5">Vocoder</td>
+      <td rowspan="6">Vocoder</td>
       <td >WaveFlow</td>
       <td >LJSpeech</td>
       <td>
@@ -378,7 +379,14 @@ PaddleSpeech supports a series of most popular models. They are summarized in [r
       <td>
       <a href = "./examples/csmsc/voc5">HiFiGAN-csmsc</a> 
       </td>
-    <tr>                                                                                                                                       
+    </tr>
+    <tr>
+      <td >WaveRNN</td>
+      <td >CSMSC</td>
+      <td>
+      <a href = "./examples/csmsc/voc6">WaveRNN-csmsc</a>
+      </td>
+    </tr>
     <tr>
       <td rowspan="3">Voice Cloning</td>
       <td>GE2E</td>
@@ -416,7 +424,6 @@ PaddleSpeech supports a series of most popular models. They are summarized in [r
     </tr>
   </thead>
   <tbody>
-  
   <tr>
       <td>Audio Classification</td>
       <td>ESC-50</td>
@@ -440,7 +447,6 @@ PaddleSpeech supports a series of most popular models. They are summarized in [r
     </tr>
   </thead>
   <tbody>
-  
   <tr>
       <td>Punctuation Restoration</td>
       <td>IWLST2012_zh</td>

README_cn.md

@@ -315,14 +315,15 @@ PaddleSpeech 的 **语音合成** 主要包含三个模块：文本前端、声
     </tr>
     <tr>
       <td rowspan="4">声学模型</td>
-      <td >Tacotron2</td>
-      <td rowspan="2" >LJSpeech</td>
+      <td>Tacotron2</td>
+      <td>LJSpeech / CSMSC</td>
       <td>
-      <a href = "./examples/ljspeech/tts0">tacotron2-ljspeech</a>
+      <a href = "./examples/ljspeech/tts0">tacotron2-ljspeech</a> / <a href = "./examples/csmsc/tts0">tacotron2-csmsc</a>
       </td>
     </tr>
     <tr>
       <td>Transformer TTS</td>
+      <td>LJSpeech</td>
       <td>
       <a href = "./examples/ljspeech/tts1">transformer-ljspeech</a>
       </td>
@@ -342,7 +343,7 @@ PaddleSpeech 的 **语音合成** 主要包含三个模块：文本前端、声
       </td>
     </tr>
    <tr>
-      <td rowspan="5">声码器</td>
+      <td rowspan="6">声码器</td>
       <td >WaveFlow</td>
       <td >LJSpeech</td>
       <td>
@@ -376,7 +377,14 @@ PaddleSpeech 的 **语音合成** 主要包含三个模块：文本前端、声
       <td>
       <a href = "./examples/csmsc/voc5">HiFiGAN-csmsc</a> 
       </td>
-    <tr>                                                                                                                                       
+    </tr>
+    <tr>
+      <td >WaveRNN</td>
+      <td >CSMSC</td>
+      <td>
+      <a href = "./examples/csmsc/voc6">WaveRNN-csmsc</a>
+      </td>
+    </tr>
     <tr>
       <td rowspan="3">声音克隆</td>
       <td>GE2E</td>
@@ -415,8 +423,6 @@ PaddleSpeech 的 **语音合成** 主要包含三个模块：文本前端、声
     </tr>
   </thead>
   <tbody>
-  
-
   <tr>
       <td>声音分类</td>
       <td>ESC-50</td>
@@ -440,7 +446,6 @@ PaddleSpeech 的 **语音合成** 主要包含三个模块：文本前端、声
     </tr>
   </thead>
   <tbody>
-  
   <tr>
       <td>标点恢复</td>
       <td>IWLST2012_zh</td>

docs/source/released_model.md

+
 # Released Models
 
 ## Speech-to-Text Models
@@ -32,7 +33,8 @@ Language Model | Training Data | Token-based | Size | Descriptions
 ### Acoustic Models
 Model Type | Dataset| Example Link | Pretrained Models|Static Models|Size (static)
 :-------------:| :------------:| :-----: | :-----:| :-----:| :-----:
-Tacotron2|LJSpeech|[tacotron2-vctk](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/ljspeech/tts0)|[tacotron2_ljspeech_ckpt_0.3.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/tacotron2/tacotron2_ljspeech_ckpt_0.3.zip)|||
+Tacotron2|LJSpeech|[tacotron2-ljspeech](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/ljspeech/tts0)|[tacotron2_ljspeech_ckpt_0.2.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/tacotron2/tacotron2_ljspeech_ckpt_0.2.0.zip)|||
+Tacotron2|CSMSC|[tacotron2-csmsc](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/csmsc/tts0)|[tacotron2_csmsc_ckpt_0.2.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/tacotron2/tacotron2_csmsc_ckpt_0.2.0.zip)|[tacotron2_csmsc_static_0.2.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/tacotron2/tacotron2_csmsc_static_0.2.0.zip)|103MB|
 TransformerTTS| LJSpeech| [transformer-ljspeech](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/ljspeech/tts1)|[transformer_tts_ljspeech_ckpt_0.4.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/transformer_tts/transformer_tts_ljspeech_ckpt_0.4.zip)|||
 SpeedySpeech| CSMSC | [speedyspeech-csmsc](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/csmsc/tts2) |[speedyspeech_nosil_baker_ckpt_0.5.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/speedyspeech/speedyspeech_nosil_baker_ckpt_0.5.zip)|[speedyspeech_nosil_baker_static_0.5.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/speedyspeech/speedyspeech_nosil_baker_static_0.5.zip)|12MB|
 FastSpeech2| CSMSC |[fastspeech2-csmsc](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/csmsc/tts3)|[fastspeech2_nosil_baker_ckpt_0.4.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/fastspeech2/fastspeech2_nosil_baker_ckpt_0.4.zip)|[fastspeech2_nosil_baker_static_0.4.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/fastspeech2/fastspeech2_nosil_baker_static_0.4.zip)|157MB|
@@ -52,6 +54,8 @@ Parallel WaveGAN| VCTK |[PWGAN-vctk](https://github.com/PaddlePaddle/PaddleSpeec
 |Multi Band MelGAN | CSMSC |[MB MelGAN-csmsc](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/csmsc/voc3) | [mb_melgan_csmsc_ckpt_0.1.1.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/mb_melgan/mb_melgan_csmsc_ckpt_0.1.1.zip) <br>[mb_melgan_baker_finetune_ckpt_0.5.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/mb_melgan/mb_melgan_baker_finetune_ckpt_0.5.zip)|[mb_melgan_csmsc_static_0.1.1.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/mb_melgan/mb_melgan_csmsc_static_0.1.1.zip) |8.2MB|
 Style MelGAN | CSMSC |[Style MelGAN-csmsc](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/csmsc/voc4)|[style_melgan_csmsc_ckpt_0.1.1.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/style_melgan/style_melgan_csmsc_ckpt_0.1.1.zip)| | |
 HiFiGAN | CSMSC |[HiFiGAN-csmsc](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/csmsc/voc5)|[hifigan_csmsc_ckpt_0.1.1.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/hifigan/hifigan_csmsc_ckpt_0.1.1.zip)|[hifigan_csmsc_static_0.1.1.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/hifigan/hifigan_csmsc_static_0.1.1.zip)|50MB|
+WaveRNN | CSMSC |[WaveRNN-csmsc](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/csmsc/voc6)|[wavernn_csmsc_ckpt_0.2.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/wavernn/wavernn_csmsc_ckpt_0.2.0.zip)|[wavernn_csmsc_static_0.2.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/wavernn/wavernn_csmsc_static_0.2.0.zip)|18MB|
+
 
 ### Voice Cloning
 Model Type | Dataset| Example Link | Pretrained Models

docs/source/tts/quick_start_cn.md

@@ -202,4 +202,4 @@ sf.write(
         audio_path,
         wav.numpy(),
         samplerate=fastspeech2_config.fs)
-```
+```
\ No newline at end of file

examples/aishell3/vc0/README.md

-
 # Tacotron2 + AISHELL-3 Voice Cloning
 This example contains code used to train a [Tacotron2](https://arxiv.org/abs/1712.05884) model with [AISHELL-3](http://www.aishelltech.com/aishell_3). The trained model can be used in Voice Cloning Task, We refer to the model structure of  [Transfer Learning from Speaker Verification to Multispeaker Text-To-Speech Synthesis](https://arxiv.org/pdf/1806.04558.pdf). The general steps are as follows:
 1. Speaker Encoder: We use Speaker Verification to train a speaker encoder. Datasets used in this task are different from those used in `Tacotron2` because the transcriptions are not needed, we use more datasets, refer to  [ge2e](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/other/ge2e).
@@ -17,7 +16,7 @@ mkdir data_aishell3
 tar zxvf data_aishell3.tgz -C data_aishell3
 ```
 ### Get MFA Result and Extract
-We use [MFA2.x](https://github.com/MontrealCorpusTools/Montreal-Forced-Aligner) to get durations for aishell3_fastspeech2.
+We use [MFA2.x](https://github.com/MontrealCorpusTools/Montreal-Forced-Aligner) to get phonemes for Tacotron2, the durations of MFA are not needed here.
 You can download from here [aishell3_alignment_tone.tar.gz](https://paddlespeech.bj.bcebos.com/MFA/AISHELL-3/with_tone/aishell3_alignment_tone.tar.gz), or train your MFA model reference to [mfa example](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/other/mfa) (use MFA1.x now) of our repo.
 
 ## Pretrained GE2E Model

examples/aishell3/vc0/path.sh

@@ -9,5 +9,5 @@ export PYTHONDONTWRITEBYTECODE=1
 export PYTHONIOENCODING=UTF-8
 export PYTHONPATH=${MAIN_ROOT}:${PYTHONPATH}
 
-MODEL=new_tacotron2
+MODEL=tacotron2
 export BIN_DIR=${MAIN_ROOT}/paddlespeech/t2s/exps/${MODEL}

examples/aishell3/vc1/README.md

-
 # FastSpeech2 + AISHELL-3 Voice Cloning
 This example contains code used to train a [FastSpeech2](https://arxiv.org/abs/2006.04558) model with [AISHELL-3](http://www.aishelltech.com/aishell_3). The trained model can be used in Voice Cloning Task, We refer to the model structure of  [Transfer Learning from Speaker Verification to Multispeaker Text-To-Speech Synthesis](https://arxiv.org/pdf/1806.04558.pdf). The general steps are as follows:
 1. Speaker Encoder: We use Speaker Verification to train a speaker encoder. Datasets used in this task are different from those used in `FastSpeech2` because the transcriptions are not needed, we use more datasets, refer to  [ge2e](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/other/ge2e).

examples/csmsc/tts0/README.md

@@ -212,6 +212,8 @@ optional arguments:
 Pretrained Tacotron2 model with no silence in the edge of audios:
 - [tacotron2_csmsc_ckpt_0.2.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/tacotron2/tacotron2_csmsc_ckpt_0.2.0.zip)
 
+The static model can be downloaded here [tacotron2_csmsc_static_0.2.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/tacotron2/tacotron2_csmsc_static_0.2.0.zip).
+
 
 Model | Step | eval/loss | eval/l1_loss | eval/mse_loss | eval/bce_loss| eval/attn_loss 
 :-------------:| :------------:| :-----: | :-----: | :--------: |:--------:|:---------:

examples/csmsc/tts0/local/synthesize_e2e.sh

@@ -7,6 +7,7 @@ ckpt_name=$3
 stage=0
 stop_stage=0
 
+# TODO: tacotron2 动转静的结果没有静态图的响亮, 可能还是 decode 的时候某个函数动静不对齐
 if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
     FLAGS_allocator_strategy=naive_best_fit \
     FLAGS_fraction_of_gpu_memory_to_use=0.01 \
@@ -33,7 +34,7 @@ if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
     FLAGS_allocator_strategy=naive_best_fit \
     FLAGS_fraction_of_gpu_memory_to_use=0.01 \
     python3 ${BIN_DIR}/../synthesize_e2e.py \
-        --am=fastspeech2_csmsc \
+        --am=tacotron2_csmsc \
         --am_config=${config_path} \
         --am_ckpt=${train_output_path}/checkpoints/${ckpt_name} \
         --am_stat=dump/train/speech_stats.npy \
@@ -55,7 +56,7 @@ if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
     FLAGS_allocator_strategy=naive_best_fit \
     FLAGS_fraction_of_gpu_memory_to_use=0.01 \
     python3 ${BIN_DIR}/../synthesize_e2e.py \
-        --am=fastspeech2_csmsc \
+        --am=tacotron2_csmsc \
         --am_config=${config_path} \
         --am_ckpt=${train_output_path}/checkpoints/${ckpt_name} \
         --am_stat=dump/train/speech_stats.npy \
@@ -76,7 +77,7 @@ if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
     FLAGS_allocator_strategy=naive_best_fit \
     FLAGS_fraction_of_gpu_memory_to_use=0.01 \
     python3 ${BIN_DIR}/../synthesize_e2e.py \
-        --am=fastspeech2_csmsc \
+        --am=tacotron2_csmsc \
         --am_config=${config_path} \
         --am_ckpt=${train_output_path}/checkpoints/${ckpt_name} \
         --am_stat=dump/train/speech_stats.npy \
@@ -90,3 +91,24 @@ if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
         --inference_dir=${train_output_path}/inference \
         --phones_dict=dump/phone_id_map.txt
 fi
+
+# wavernn
+if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
+    echo "in wavernn syn_e2e"
+    FLAGS_allocator_strategy=naive_best_fit \
+    FLAGS_fraction_of_gpu_memory_to_use=0.01 \
+    python3 ${BIN_DIR}/../synthesize_e2e.py \
+        --am=tacotron2_csmsc \
+        --am_config=${config_path} \
+        --am_ckpt=${train_output_path}/checkpoints/${ckpt_name} \
+        --am_stat=dump/train/speech_stats.npy \
+        --voc=wavernn_csmsc \
+        --voc_config=wavernn_csmsc_ckpt_0.2.0/default.yaml \
+        --voc_ckpt=wavernn_csmsc_ckpt_0.2.0/snapshot_iter_400000.pdz \
+        --voc_stat=wavernn_csmsc_ckpt_0.2.0/feats_stats.npy \
+        --lang=zh \
+        --text=${BIN_DIR}/../sentences.txt \
+        --output_dir=${train_output_path}/test_e2e \
+        --phones_dict=dump/phone_id_map.txt \
+        --inference_dir=${train_output_path}/inference
+fi
\ No newline at end of file

examples/csmsc/tts0/path.sh

@@ -9,5 +9,5 @@ export PYTHONDONTWRITEBYTECODE=1
 export PYTHONIOENCODING=UTF-8
 export PYTHONPATH=${MAIN_ROOT}:${PYTHONPATH}
 
-MODEL=new_tacotron2
+MODEL=tacotron2
 export BIN_DIR=${MAIN_ROOT}/paddlespeech/t2s/exps/${MODEL}

examples/csmsc/tts0/run.sh

@@ -35,3 +35,8 @@ if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
     # synthesize_e2e, vocoder is pwgan
     CUDA_VISIBLE_DEVICES=${gpus} ./local/synthesize_e2e.sh ${conf_path} ${train_output_path} ${ckpt_name} || exit -1
 fi
+
+if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
+    # inference with static model
+    CUDA_VISIBLE_DEVICES=${gpus} ./local/inference.sh ${train_output_path} || exit -1
+fi

examples/csmsc/tts2/local/synthesize_e2e.sh

@@ -92,3 +92,26 @@ if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
         --phones_dict=dump/phone_id_map.txt \
         --tones_dict=dump/tone_id_map.txt
 fi
+
+
+# wavernn
+if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
+    echo "in wavernn syn_e2e"
+    FLAGS_allocator_strategy=naive_best_fit \
+    FLAGS_fraction_of_gpu_memory_to_use=0.01 \
+    python3 ${BIN_DIR}/../synthesize_e2e.py \
+        --am=speedyspeech_csmsc \
+        --am_config=${config_path} \
+        --am_ckpt=${train_output_path}/checkpoints/${ckpt_name} \
+        --am_stat=dump/train/feats_stats.npy \
+        --voc=wavernn_csmsc \
+        --voc_config=wavernn_csmsc_ckpt_0.2.0/default.yaml \
+        --voc_ckpt=wavernn_csmsc_ckpt_0.2.0/snapshot_iter_400000.pdz \
+        --voc_stat=wavernn_csmsc_ckpt_0.2.0/feats_stats.npy \
+        --lang=zh \
+        --text=${BIN_DIR}/../sentences.txt \
+        --output_dir=${train_output_path}/test_e2e \
+        --phones_dict=dump/phone_id_map.txt \
+        --tones_dict=dump/tone_id_map.txt \
+        --inference_dir=${train_output_path}/inference
+fi

examples/csmsc/tts3/local/synthesize_e2e.sh

@@ -102,9 +102,9 @@ if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
         --am_ckpt=${train_output_path}/checkpoints/${ckpt_name} \
         --am_stat=dump/train/speech_stats.npy \
         --voc=wavernn_csmsc \
-        --voc_config=wavernn_test/default.yaml \
-        --voc_ckpt=wavernn_test/snapshot_iter_5000.pdz \
-        --voc_stat=wavernn_test/feats_stats.npy \
+        --voc_config=wavernn_csmsc_ckpt_0.2.0/default.yaml \
+        --voc_ckpt=wavernn_csmsc_ckpt_0.2.0/snapshot_iter_400000.pdz \
+        --voc_stat=wavernn_csmsc_ckpt_0.2.0/feats_stats.npy \
         --lang=zh \
         --text=${BIN_DIR}/../sentences.txt \
         --output_dir=${train_output_path}/test_e2e \

examples/csmsc/tts3/run.sh

@@ -36,3 +36,8 @@ if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
     CUDA_VISIBLE_DEVICES=${gpus} ./local/synthesize_e2e.sh ${conf_path} ${train_output_path} ${ckpt_name} || exit -1
 fi
 
+if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
+    # inference with static model
+    CUDA_VISIBLE_DEVICES=${gpus} ./local/inference.sh ${train_output_path} || exit -1
+fi
+

examples/ljspeech/tts0/README.md

+# Tacotron2 with LJSpeech-1.1
+This example contains code used to train a [Tacotron2](https://arxiv.org/abs/1712.05884) model with [LJSpeech-1.1](https://keithito.com/LJ-Speech-Dataset/)
+
+## Dataset
+### Download and Extract
+Download LJSpeech-1.1 from the [official website](https://keithito.com/LJ-Speech-Dataset/).
+
+### Get MFA Result and Extract
+We use [MFA](https://github.com/MontrealCorpusTools/Montreal-Forced-Aligner) to get phonemes for Tacotron2, the durations of MFA are not needed here.
+You can download from here [ljspeech_alignment.tar.gz](https://paddlespeech.bj.bcebos.com/MFA/LJSpeech-1.1/ljspeech_alignment.tar.gz), or train your MFA model reference to [mfa example](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/other/mfa) of our repo.
+
+## Get Started
+Assume the path to the dataset is `~/datasets/LJSpeech-1.1`.
+Assume the path to the MFA result of LJSpeech-1.1 is `./ljspeech_alignment`.
+Run the command below to
+1. **source path**.
+2. preprocess the dataset.
+3. train the model.
+4. synthesize wavs.
+    - synthesize waveform from `metadata.jsonl`.
+    - synthesize waveform from a text file.
+
+```bash
+./run.sh
+```
+You can choose a range of stages you want to run, or set `stage` equal to `stop-stage` to use only one stage, for example, running the following command will only preprocess the dataset.
+```bash
+./run.sh --stage 0 --stop-stage 0
+```
+### Data Preprocessing
+```bash
+./local/preprocess.sh ${conf_path}
+```
+When it is done. A `dump` folder is created in the current directory. The structure of the dump folder is listed below.
+
+```text
+dump
+├── dev
+│   ├── norm
+│   └── raw
+├── phone_id_map.txt
+├── speaker_id_map.txt
+├── test
+│   ├── norm
+│   └── raw
+└── train
+    ├── norm
+    ├── raw
+    └── speech_stats.npy
+```
+The dataset is split into 3 parts, namely `train`, `dev`, and` test`, each of which contains a `norm` and `raw` subfolder. The raw folder contains speech features of each utterance, while the norm folder contains normalized ones. The statistics used to normalize features are computed from the training set, which is located in `dump/train/*_stats.npy`.
+
+Also, there is a `metadata.jsonl` in each subfolder. It is a table-like file that contains phones, text_lengths, speech_lengths, durations, the path of speech features, speaker, and the id of each utterance.
+
+### Model Training
+```bash
+CUDA_VISIBLE_DEVICES=${gpus} ./local/train.sh ${conf_path} ${train_output_path}
+```
+`./local/train.sh` calls `${BIN_DIR}/train.py`.
+Here's the complete help message.
+```text
+usage: train.py [-h] [--config CONFIG] [--train-metadata TRAIN_METADATA]
+                [--dev-metadata DEV_METADATA] [--output-dir OUTPUT_DIR]
+                [--ngpu NGPU] [--phones-dict PHONES_DICT]
+
+Train a Tacotron2 model.
+
+optional arguments:
+  -h, --help            show this help message and exit
+  --config CONFIG       tacotron2 config file.
+  --train-metadata TRAIN_METADATA
+                        training data.
+  --dev-metadata DEV_METADATA
+                        dev data.
+  --output-dir OUTPUT_DIR
+                        output dir.
+  --ngpu NGPU           if ngpu == 0, use cpu.
+  --phones-dict PHONES_DICT
+                        phone vocabulary file.
+```
+1. `--config` is a config file in yaml format to overwrite the default config, which can be found at `conf/default.yaml`.
+2. `--train-metadata` and `--dev-metadata` should be the metadata file in the normalized subfolder of `train` and `dev` in the `dump` folder.
+3. `--output-dir` is the directory to save the results of the experiment. Checkpoints are saved in `checkpoints/` inside this directory.
+4. `--ngpu` is the number of gpus to use, if ngpu == 0, use cpu.
+5. `--phones-dict` is the path of the phone vocabulary file.
+
+### Synthesizing
+We use [parallel wavegan](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/ljspeech/voc1) as the neural vocoder.
+Download pretrained parallel wavegan model from [pwg_ljspeech_ckpt_0.5.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/pwgan/pwg_ljspeech_ckpt_0.5.zip) and unzip it.
+```bash
+unzip pwg_ljspeech_ckpt_0.5.zip
+```
+Parallel WaveGAN checkpoint contains files listed below.
+```text
+pwg_ljspeech_ckpt_0.5
+├── pwg_default.yaml              # default config used to train parallel wavegan
+├── pwg_snapshot_iter_400000.pdz  # generator parameters of parallel wavegan
+└── pwg_stats.npy                 # statistics used to normalize spectrogram when training parallel wavegan
+```
+`./local/synthesize.sh` calls `${BIN_DIR}/../synthesize.py`, which can synthesize waveform from `metadata.jsonl`.
+```bash
+CUDA_VISIBLE_DEVICES=${gpus} ./local/synthesize.sh ${conf_path} ${train_output_path} ${ckpt_name}
+```
+```text
+usage: synthesize.py [-h]
+                     [--am {speedyspeech_csmsc,fastspeech2_csmsc,fastspeech2_ljspeech,fastspeech2_aishell3,fastspeech2_vctk,tacotron2_csmsc}]
+                     [--am_config AM_CONFIG] [--am_ckpt AM_CKPT]
+                     [--am_stat AM_STAT] [--phones_dict PHONES_DICT]
+                     [--tones_dict TONES_DICT] [--speaker_dict SPEAKER_DICT]
+                     [--voice-cloning VOICE_CLONING]
+                     [--voc {pwgan_csmsc,pwgan_ljspeech,pwgan_aishell3,pwgan_vctk,mb_melgan_csmsc}]
+                     [--voc_config VOC_CONFIG] [--voc_ckpt VOC_CKPT]
+                     [--voc_stat VOC_STAT] [--ngpu NGPU]
+                     [--test_metadata TEST_METADATA] [--output_dir OUTPUT_DIR]
+
+Synthesize with acoustic model & vocoder
+
+optional arguments:
+  -h, --help            show this help message and exit
+  --am {speedyspeech_csmsc,fastspeech2_csmsc,fastspeech2_ljspeech,fastspeech2_aishell3,fastspeech2_vctk,tacotron2_csmsc}
+                        Choose acoustic model type of tts task.
+  --am_config AM_CONFIG
+                        Config of acoustic model. Use deault config when it is
+                        None.
+  --am_ckpt AM_CKPT     Checkpoint file of acoustic model.
+  --am_stat AM_STAT     mean and standard deviation used to normalize
+                        spectrogram when training acoustic model.
+  --phones_dict PHONES_DICT
+                        phone vocabulary file.
+  --tones_dict TONES_DICT
+                        tone vocabulary file.
+  --speaker_dict SPEAKER_DICT
+                        speaker id map file.
+  --voice-cloning VOICE_CLONING
+                        whether training voice cloning model.
+  --voc {pwgan_csmsc,pwgan_ljspeech,pwgan_aishell3,pwgan_vctk,mb_melgan_csmsc}
+                        Choose vocoder type of tts task.
+  --voc_config VOC_CONFIG
+                        Config of voc. Use deault config when it is None.
+  --voc_ckpt VOC_CKPT   Checkpoint file of voc.
+  --voc_stat VOC_STAT   mean and standard deviation used to normalize
+                        spectrogram when training voc.
+  --ngpu NGPU           if ngpu == 0, use cpu.
+  --test_metadata TEST_METADATA
+                        test metadata.
+  --output_dir OUTPUT_DIR
+                        output dir.
+```
+`./local/synthesize_e2e.sh` calls `${BIN_DIR}/../synthesize_e2e.py`, which can synthesize waveform from text file.
+```bash
+CUDA_VISIBLE_DEVICES=${gpus} ./local/synthesize_e2e.sh ${conf_path} ${train_output_path} ${ckpt_name}
+```
+```text
+usage: synthesize_e2e.py [-h]
+                         [--am {speedyspeech_csmsc,speedyspeech_aishell3,fastspeech2_csmsc,fastspeech2_ljspeech,fastspeech2_aishell3,fastspeech2_vctk,tacotron2_csmsc}]
+                         [--am_config AM_CONFIG] [--am_ckpt AM_CKPT]
+                         [--am_stat AM_STAT] [--phones_dict PHONES_DICT]
+                         [--tones_dict TONES_DICT]
+                         [--speaker_dict SPEAKER_DICT] [--spk_id SPK_ID]
+                         [--voc {pwgan_csmsc,pwgan_ljspeech,pwgan_aishell3,pwgan_vctk,mb_melgan_csmsc,style_melgan_csmsc,hifigan_csmsc}]
+                         [--voc_config VOC_CONFIG] [--voc_ckpt VOC_CKPT]
+                         [--voc_stat VOC_STAT] [--lang LANG]
+                         [--inference_dir INFERENCE_DIR] [--ngpu NGPU]
+                         [--text TEXT] [--output_dir OUTPUT_DIR]
+
+Synthesize with acoustic model & vocoder
+
+optional arguments:
+  -h, --help            show this help message and exit
+  --am {speedyspeech_csmsc,speedyspeech_aishell3,fastspeech2_csmsc,fastspeech2_ljspeech,fastspeech2_aishell3,fastspeech2_vctk,tacotron2_csmsc}
+                        Choose acoustic model type of tts task.
+  --am_config AM_CONFIG
+                        Config of acoustic model. Use deault config when it is
+                        None.
+  --am_ckpt AM_CKPT     Checkpoint file of acoustic model.
+  --am_stat AM_STAT     mean and standard deviation used to normalize
+                        spectrogram when training acoustic model.
+  --phones_dict PHONES_DICT
+                        phone vocabulary file.
+  --tones_dict TONES_DICT
+                        tone vocabulary file.
+  --speaker_dict SPEAKER_DICT
+                        speaker id map file.
+  --spk_id SPK_ID       spk id for multi speaker acoustic model
+  --voc {pwgan_csmsc,pwgan_ljspeech,pwgan_aishell3,pwgan_vctk,mb_melgan_csmsc,style_melgan_csmsc,hifigan_csmsc}
+                        Choose vocoder type of tts task.
+  --voc_config VOC_CONFIG
+                        Config of voc. Use deault config when it is None.
+  --voc_ckpt VOC_CKPT   Checkpoint file of voc.
+  --voc_stat VOC_STAT   mean and standard deviation used to normalize
+                        spectrogram when training voc.
+  --lang LANG           Choose model language. zh or en
+  --inference_dir INFERENCE_DIR
+                        dir to save inference models
+  --ngpu NGPU           if ngpu == 0, use cpu.
+  --text TEXT           text to synthesize, a 'utt_id sentence' pair per line.
+  --output_dir OUTPUT_DIR
+                        output dir.
+```
+1. `--am` is acoustic model type with the format {model_name}_{dataset}
+2. `--am_config`, `--am_checkpoint`, `--am_stat` and `--phones_dict` are arguments for acoustic model, which correspond to the 4 files in the Tacotron2 pretrained model.
+3. `--voc` is vocoder type with the format {model_name}_{dataset}
+4. `--voc_config`, `--voc_checkpoint`, `--voc_stat` are arguments for vocoder, which correspond to the 3 files in the parallel wavegan pretrained model.
+5. `--lang` is the model language, which can be `zh` or `en`.
+6. `--test_metadata` should be the metadata file in the normalized subfolder of `test`  in the `dump` folder.
+7. `--text` is the text file, which contains sentences to synthesize.
+8. `--output_dir` is the directory to save synthesized audio files.
+9. `--ngpu` is the number of gpus to use, if ngpu == 0, use cpu.
+
+
+## Pretrained Model
+Pretrained Tacotron2 model with no silence in the edge of audios:
+- [tacotron2_ljspeech_ckpt_0.2.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/tacotron2/tacotron2_ljspeech_ckpt_0.2.0.zip)
+
+
+Model | Step | eval/loss | eval/l1_loss | eval/mse_loss | eval/bce_loss| eval/attn_loss 
+:-------------:| :------------:| :-----: | :-----: | :--------: |:--------:|:---------:
+default| 1(gpu) x 60300|0.554092|0.394260|0.141046|0.018747|3.8e-05|
+
+Tacotron2 checkpoint contains files listed below.
+```text
+tacotron2_ljspeech_ckpt_0.2.0
+├── default.yaml            # default config used to train Tacotron2
+├── phone_id_map.txt        # phone vocabulary file when training Tacotron2
+├── snapshot_iter_60300.pdz # model parameters and optimizer states
+└── speech_stats.npy        # statistics used to normalize spectrogram when training Tacotron2
+```
+You can use the following scripts to synthesize for `${BIN_DIR}/../sentences_en.txt` using pretrained Tacotron2 and parallel wavegan models.
+```bash
+source path.sh
+
+FLAGS_allocator_strategy=naive_best_fit \
+FLAGS_fraction_of_gpu_memory_to_use=0.01 \
+python3 ${BIN_DIR}/../synthesize_e2e.py \
+  --am=tacotron2_ljspeech \
+  --am_config=tacotron2_ljspeech_ckpt_0.2.0/default.yaml \
+  --am_ckpt=tacotron2_ljspeech_ckpt_0.2.0/snapshot_iter_60300.pdz \
+  --am_stat=tacotron2_ljspeech_ckpt_0.2.0/speech_stats.npy  \
+  --voc=pwgan_ljspeech\
+  --voc_config=pwg_ljspeech_ckpt_0.5/pwg_default.yaml \
+  --voc_ckpt=pwg_ljspeech_ckpt_0.5/pwg_snapshot_iter_400000.pdz  \
+  --voc_stat=pwg_ljspeech_ckpt_0.5/pwg_stats.npy \
+  --lang=en \
+  --text=${BIN_DIR}/../sentences_en.txt \
+  --output_dir=exp/default/test_e2e \
+  --phones_dict=tacotron2_ljspeech_ckpt_0.2.0/phone_id_map.txt
+```

examples/ljspeech/tts0/local/synthesize_e2e.sh

+#!/bin/bash
+
+config_path=$1
+train_output_path=$2
+ckpt_name=$3
+# TODO: dygraph to static graph is not good for tacotron2_ljspeech now
+FLAGS_allocator_strategy=naive_best_fit \
+FLAGS_fraction_of_gpu_memory_to_use=0.01 \
+python3 ${BIN_DIR}/../synthesize_e2e.py \
+    --am=tacotron2_ljspeech \
+    --am_config=${config_path} \
+    --am_ckpt=${train_output_path}/checkpoints/${ckpt_name} \
+    --am_stat=dump/train/speech_stats.npy \
+    --voc=pwgan_ljspeech \
+    --voc_config=pwg_ljspeech_ckpt_0.5/pwg_default.yaml \
+    --voc_ckpt=pwg_ljspeech_ckpt_0.5/pwg_snapshot_iter_400000.pdz  \
+    --voc_stat=pwg_ljspeech_ckpt_0.5/pwg_stats.npy \
+    --lang=en \
+    --text=${BIN_DIR}/../sentences_en.txt \
+    --output_dir=${train_output_path}/test_e2e \
+    --phones_dict=dump/phone_id_map.txt \
+    # --inference_dir=${train_output_path}/inference
\ No newline at end of file

examples/ljspeech/tts0/path.sh

@@ -9,5 +9,5 @@ export PYTHONDONTWRITEBYTECODE=1
 export PYTHONIOENCODING=UTF-8
 export PYTHONPATH=${MAIN_ROOT}:${PYTHONPATH}
 
-MODEL=new_tacotron2
+MODEL=tacotron2
 export BIN_DIR=${MAIN_ROOT}/paddlespeech/t2s/exps/${MODEL}

examples/ljspeech/tts3/README.md

-# FastSpeech2 with the LJSpeech-1.1
+# FastSpeech2 with LJSpeech-1.1
 This example contains code used to train a [Fastspeech2](https://arxiv.org/abs/2006.04558) model with [LJSpeech-1.1](https://keithito.com/LJ-Speech-Dataset/).
 
 ## Dataset

paddleaudio/features/core.py

@@ -415,11 +415,11 @@ def mfcc(x,
         **kwargs)
 
     # librosa mfcc:
-    spect = librosa.feature.melspectrogram(x,sr=16000,n_fft=512,
+    spect = librosa.feature.melspectrogram(y=x,sr=16000,n_fft=512,
                                               win_length=512,
                                               hop_length=320,
                                               n_mels=64, fmin=50)
-    b = librosa.feature.mfcc(x,
+    b = librosa.feature.mfcc(y=x,
         sr=16000,
         S=spect,
         n_mfcc=20,

paddlespeech/s2t/transform/spectrogram.py

@@ -67,7 +67,7 @@ def istft(x, n_shift, win_length=None, window="hann", center=True):
     x = np.stack(
         [
             librosa.istft(
-                y=x[:, ch].T,  # [Time, Freq] -> [Freq, Time]
+                stft_matrix=x[:, ch].T,  # [Time, Freq] -> [Freq, Time]
                 hop_length=n_shift,
                 win_length=win_length,
                 window=window,

paddlespeech/t2s/audio/audio.py

@@ -53,8 +53,8 @@ class AudioProcessor(object):
 
     def _create_mel_filter(self):
         mel_filter = librosa.filters.mel(
-            self.sample_rate,
-            self.n_fft,
+            sr=self.sample_rate,
+            n_fft=self.n_fft,
             n_mels=self.n_mels,
             fmin=self.fmin,
             fmax=self.fmax)

paddlespeech/t2s/datasets/common.py

@@ -38,7 +38,7 @@ class AudioSegmentDataset(Dataset):
 
     def __getitem__(self, i):
         fpath = self.file_paths[i]
-        y, sr = librosa.load(fpath, self.sr)
+        y, sr = librosa.load(fpath, sr=self.sr)
         y, _ = librosa.effects.trim(y, top_db=self.top_db)
         y = librosa.util.normalize(y)
         y = y.astype(np.float32)
@@ -70,7 +70,7 @@ class AudioDataset(Dataset):
 
     def __getitem__(self, i):
         fpath = self.file_paths[i]
-        y, sr = librosa.load(fpath, self.sr)
+        y, sr = librosa.load(fpath, sr=self.sr)
         y, _ = librosa.effects.trim(y, top_db=self.top_db)
         y = librosa.util.normalize(y)
         y = y.astype(np.float32)

paddlespeech/t2s/exps/synthesize.py

@@ -38,9 +38,9 @@ model_alias = {
     "fastspeech2_inference":
     "paddlespeech.t2s.models.fastspeech2:FastSpeech2Inference",
     "tacotron2":
-    "paddlespeech.t2s.models.new_tacotron2:Tacotron2",
+    "paddlespeech.t2s.models.tacotron2:Tacotron2",
     "tacotron2_inference":
-    "paddlespeech.t2s.models.new_tacotron2:Tacotron2Inference",
+    "paddlespeech.t2s.models.tacotron2:Tacotron2Inference",
     # voc
     "pwgan":
     "paddlespeech.t2s.models.parallel_wavegan:PWGGenerator",

paddlespeech/t2s/exps/synthesize_e2e.py

@@ -39,9 +39,9 @@ model_alias = {
     "fastspeech2_inference":
     "paddlespeech.t2s.models.fastspeech2:FastSpeech2Inference",
     "tacotron2":
-    "paddlespeech.t2s.models.new_tacotron2:Tacotron2",
+    "paddlespeech.t2s.models.tacotron2:Tacotron2",
     "tacotron2_inference":
-    "paddlespeech.t2s.models.new_tacotron2:Tacotron2Inference",
+    "paddlespeech.t2s.models.tacotron2:Tacotron2Inference",
     # voc
     "pwgan":
     "paddlespeech.t2s.models.parallel_wavegan:PWGGenerator",
@@ -229,6 +229,11 @@ def evaluate(args):
     output_dir = Path(args.output_dir)
     output_dir.mkdir(parents=True, exist_ok=True)
     merge_sentences = False
+    # Avoid not stopping at the end of a sub sentence when tacotron2_ljspeech dygraph to static graph
+    # but still not stopping in the end (NOTE by yuantian01 Feb 9 2022)
+    if am_name == 'tacotron2':
+        merge_sentences = True
+
     for utt_id, sentence in sentences:
         get_tone_ids = False
         if am_name == 'speedyspeech':

paddlespeech/t2s/exps/new_tacotron2/train.py → paddlespeech/t2s/exps/tacotron2/train.py

@@ -30,9 +30,9 @@ from yacs.config import CfgNode
 from paddlespeech.t2s.datasets.am_batch_fn import tacotron2_multi_spk_batch_fn
 from paddlespeech.t2s.datasets.am_batch_fn import tacotron2_single_spk_batch_fn
 from paddlespeech.t2s.datasets.data_table import DataTable
-from paddlespeech.t2s.models.new_tacotron2 import Tacotron2
-from paddlespeech.t2s.models.new_tacotron2 import Tacotron2Evaluator
-from paddlespeech.t2s.models.new_tacotron2 import Tacotron2Updater
+from paddlespeech.t2s.models.tacotron2 import Tacotron2
+from paddlespeech.t2s.models.tacotron2 import Tacotron2Evaluator
+from paddlespeech.t2s.models.tacotron2 import Tacotron2Updater
 from paddlespeech.t2s.training.extensions.snapshot import Snapshot
 from paddlespeech.t2s.training.extensions.visualizer import VisualDL
 from paddlespeech.t2s.training.optimizer import build_optimizers

paddlespeech/t2s/exps/voice_cloning.py

@@ -34,9 +34,9 @@ model_alias = {
     "fastspeech2_inference":
     "paddlespeech.t2s.models.fastspeech2:FastSpeech2Inference",
     "tacotron2":
-    "paddlespeech.t2s.models.new_tacotron2:Tacotron2",
+    "paddlespeech.t2s.models.tacotron2:Tacotron2",
     "tacotron2_inference":
-    "paddlespeech.t2s.models.new_tacotron2:Tacotron2Inference",
+    "paddlespeech.t2s.models.tacotron2:Tacotron2Inference",
     # voc
     "pwgan":
     "paddlespeech.t2s.models.parallel_wavegan:PWGGenerator",

paddlespeech/t2s/exps/wavernn/synthesize.py

@@ -31,7 +31,7 @@ from paddlespeech.t2s.models.wavernn import WaveRNN
 def main():
     parser = argparse.ArgumentParser(description="Synthesize with WaveRNN.")
 
-    parser.add_argument("--config", type=str, help="GANVocoder config file.")
+    parser.add_argument("--config", type=str, help="Vocoder config file.")
     parser.add_argument("--checkpoint", type=str, help="snapshot to load.")
     parser.add_argument("--test-metadata", type=str, help="dev data.")
     parser.add_argument("--output-dir", type=str, help="output dir.")

paddlespeech/t2s/exps/wavernn/train.py

@@ -179,7 +179,7 @@ def train_sp(args, config):
 def main():
     # parse args and config and redirect to train_sp
 
-    parser = argparse.ArgumentParser(description="Train a HiFiGAN model.")
+    parser = argparse.ArgumentParser(description="Train a WaveRNN model.")
     parser.add_argument(
         "--config", type=str, help="config file to overwrite default config.")
     parser.add_argument("--train-metadata", type=str, help="training data.")

paddlespeech/t2s/frontend/phonectic.py

@@ -83,11 +83,6 @@ class English(Phonetics):
         return phonemes
 
     def _p2id(self, phonemes: List[str]) -> np.array:
-        # replace unk phone with sp
-        phonemes = [
-            phn if (phn in self.vocab_phones and phn not in self.punc) else "sp"
-            for phn in phonemes
-        ]
         phone_ids = [self.vocab_phones[item] for item in phonemes]
         return np.array(phone_ids, np.int64)
 
@@ -102,6 +97,12 @@ class English(Phonetics):
             # remove start_symbol and end_symbol
             phones = phones[1:-1]
             phones = [phn for phn in phones if not phn.isspace()]
+            # replace unk phone with sp
+            phones = [
+                phn
+                if (phn in self.vocab_phones and phn not in self.punc) else "sp"
+                for phn in phones
+            ]
             phones_list.append(phones)
 
         if merge_sentences:

paddlespeech/t2s/models/__init__.py

@@ -14,9 +14,9 @@
 from .fastspeech2 import *
 from .hifigan import *
 from .melgan import *
-from .new_tacotron2 import *
 from .parallel_wavegan import *
 from .speedyspeech import *
+from .tacotron2 import *
 from .transformer_tts import *
 from .waveflow import *
 from .wavernn import *

paddlespeech/t2s/modules/tacotron2/decoder.py

@@ -395,9 +395,6 @@ class Decoder(nn.Layer):
             iunits, odim * reduction_factor, bias_attr=False)
         self.prob_out = nn.Linear(iunits, reduction_factor)
 
-        # initialize
-        # self.apply(decoder_init)
-
     def _zero_state(self, hs):
         init_hs = paddle.zeros([paddle.shape(hs)[0], self.lstm[0].hidden_size])
         return init_hs
@@ -558,8 +555,11 @@ class Decoder(nn.Layer):
         assert len(paddle.shape(h)) == 2
         hs = h.unsqueeze(0)
         ilens = paddle.shape(h)[0]
-        maxlen = int(paddle.shape(h)[0] * maxlenratio)
-        minlen = int(paddle.shape(h)[0] * minlenratio)
+        # 本来 maxlen 和 minlen 外面有 int()，防止动转静的问题此处删除
+        maxlen = paddle.shape(h)[0] * maxlenratio
+        minlen = paddle.shape(h)[0] * minlenratio
+        # 本来是直接使用 threshold 的，此处为了防止动转静的问题把 threshold 转成 tensor
+        threshold = paddle.ones([1]) * threshold
 
         # initialize hidden states of decoder
         c_list = [self._zero_state(hs)]
@@ -645,11 +645,27 @@ class Decoder(nn.Layer):
             if use_att_constraint:
                 last_attended_idx = int(att_w.argmax())
 
+            # tacotron2 ljspeech 动转静的问题应该是这里没有正确判断 prob >= threshold 导致的
             if prob >= threshold or idx >= maxlen:
                 # check mininum length
                 if idx < minlen:
                     continue
                 break
+            """
+            仅解开 665~667 行的代码块，动转静时会卡死，但是动态图时可以正确生成音频，证明模型没问题
+            同时解开 665~667 行 和 668 ~ 670 行的代码块，动转静时不会卡死，但是生成的音频末尾有多余的噪声
+            证明动转静没有进入 prob >= threshold 的判断，但是静态图可以进入 prob >= threshold 并退出循环
+            动转静时是通过 idx >= maxlen 退出循环（所以没有这个逻辑的时候会一直循环，也就是卡死），
+            没有在模型判断该结束的时候结束，而是在超出最大长度时结束，所以合成的音频末尾有很长的额外预测的噪声
+            动转静用 prob <= threshold 的条件可以退出循环（虽然结果不正确），证明条件参数的类型本身没问题，可能是 prob 有问题
+            """
+            # if prob >= threshold:
+            #     print("prob >= threshold")
+            #     break
+            # elif idx >= maxlen:
+            #     print("idx >= maxlen")
+            #     break
+
         # (1, odim, L)
         outs = paddle.concat(outs, axis=2)
         if self.postnet is not None: