[TTS]Cantonese FastSpeech2 Training, test=tts (#2907)

examples/canton/tts3/README.md

+# FastSpeech2 with Cantonese language
+
+## Dataset
+### Download and Extract
+If you don't have the Cantonese datasets mentioned above, please download and unzip  [Guangzhou_Cantonese_Scripted_Speech_Corpus_Daily_Use_Sentence](https://magichub.com/datasets/guangzhou-cantonese-scripted-speech-corpus-daily-use-sentence/) and [Guangzhou_Cantonese_Scripted_Speech_Corpus_in_Vehicle](https://magichub.com/datasets/guangzhou-cantonese-scripted-speech-corpus-in-the-vehicle/) under `~/datasets/`.
+
+To obtain better performance, please combine these two datasets together as follows:
+
+```bash
+mkdir -p ~/datasets/canton_all/WAV
+cp -r ~/datasets/Guangzhou_Cantonese_Scripted_Speech_Corpus_Daily_Use_Sentence/WAV/* ~/datasets/canton_all/WAV
+cp -r ~/datasets/Guangzhou_Cantonese_Scripted_Speech_Corpus_in_Vehicle/WAV/* ~/datasets/canton_all/WAV
+```
+
+After that, it should be look like:
+```
+~/datasets/canton_all
+│   └── WAV
+│       └──G0001
+│       └──G0002
+│       ...
+│       └──G0071
+│       └──G0072
+```
+
+
+### Get MFA Result and Extract
+We use [MFA1.x](https://github.com/MontrealCorpusTools/Montreal-Forced-Aligner) to get durations for canton_fastspeech2.
+You can train your MFA model reference to [canton_mfa example](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/other/mfa) (use MFA1.x now) of our repo.
+We here provide the MFA results of these two datasets. [canton_alignment.zip](https://paddlespeech.bj.bcebos.com/MFA/Canton/canton_alignment.zip)
+
+## Get Started
+Assume the path to the Cantonese MFA result of the two datsets mentioned above is `./canton_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 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
+    ├── energy_stats.npy
+    ├── norm
+    ├── pitch_stats.npy
+    ├── 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、pitch and energy 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, the path of pitch features, a path of energy features, speaker, and id of each utterance.
+
+### Training details can refer to the script of [examples/aishell3/tts3](../../aishell3/tts3).
+
+## Pretrained Model(Waiting========)
+Pretrained FastSpeech2 model with no silence in the edge of audios:
+- [fastspeech2_aishell3_ckpt_1.1.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/fastspeech2/fastspeech2_aishell3_ckpt_1.1.0.zip)
+- [fastspeech2_conformer_aishell3_ckpt_0.2.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/fastspeech2/fastspeech2_conformer_aishell3_ckpt_0.2.0.zip) (Thanks for [@awmmmm](https://github.com/awmmmm)'s contribution)
+
+
+FastSpeech2 checkpoint contains files listed below.
+
+```text
+fastspeech2_aishell3_ckpt_1.1.0
+├── default.yaml            # default config used to train fastspeech2
+├── energy_stats.npy        # statistics used to normalize energy when training fastspeech2
+├── phone_id_map.txt        # phone vocabulary file when training fastspeech2
+├── pitch_stats.npy         # statistics used to normalize pitch when training fastspeech2
+├── snapshot_iter_96400.pdz # model parameters and optimizer states
+├── speaker_id_map.txt      # speaker id map file when training a multi-speaker fastspeech2
+└── speech_stats.npy        # statistics used to normalize spectrogram when training fastspeech2
+```
+You can use the following scripts to synthesize for `${BIN_DIR}/../sentences.txt` using pretrained fastspeech2 and parallel wavegan models.
+```bash
+source path.sh
+
+FLAGS_allocator_strategy=naive_best_fit \
+FLAGS_fraction_of_gpu_memory_to_use=0.01 \
+python3 ${BIN_DIR}/../synthesize_e2e.py \
+  --am=fastspeech2_aishell3 \
+  --am_config=fastspeech2_aishell3_ckpt_1.1.0/default.yaml \
+  --am_ckpt=fastspeech2_aishell3_ckpt_1.1.0/snapshot_iter_96400.pdz \
+  --am_stat=fastspeech2_aishell3_ckpt_1.1.0/speech_stats.npy \
+  --voc=pwgan_aishell3 \
+  --voc_config=pwg_aishell3_ckpt_0.5/default.yaml \
+  --voc_ckpt=pwg_aishell3_ckpt_0.5/snapshot_iter_1000000.pdz \
+  --voc_stat=pwg_aishell3_ckpt_0.5/feats_stats.npy \
+  --lang=zh \
+  --text=${BIN_DIR}/../sentences.txt \
+  --output_dir=exp/default/test_e2e \
+  --phones_dict=fastspeech2_aishell3_ckpt_1.1.0/phone_id_map.txt \
+  --speaker_dict=fastspeech2_aishell3_ckpt_1.1.0/speaker_id_map.txt \
+  --spk_id=0 \
+  --inference_dir=exp/default/inference
+```

examples/canton/tts3/conf/default.yaml

+###########################################################
+#                FEATURE EXTRACTION SETTING               #
+###########################################################
+
+fs: 24000          # sr
+n_fft: 2048        # FFT size (samples).
+n_shift: 300       # Hop size (samples). 12.5ms
+win_length: 1200   # Window length (samples). 50ms
+                   # If set to null, it will be the same as fft_size.
+window: "hann"     # Window function.
+
+# Only used for feats_type != raw
+
+fmin: 80           # Minimum frequency of Mel basis.
+fmax: 7600         # Maximum frequency of Mel basis.
+n_mels: 80         # The number of mel basis.
+
+# Only used for the model using pitch features (e.g. FastSpeech2)
+# The canton datasets we use are different from others like Databaker or LJSpeech, 
+# we set it to 110 to avoid too many zero-pitch problem. 
+# Reference: https://github.com/JeremyCCHsu/Python-Wrapper-for-World-Vocoder/issues/38
+f0min: 110          # Minimum f0 for pitch extraction.
+f0max: 400         # Maximum f0 for pitch extraction.
+
+
+###########################################################
+#                       DATA SETTING                      #
+###########################################################
+batch_size: 32
+num_workers: 2
+
+
+###########################################################
+#                       MODEL SETTING                     #
+###########################################################
+model:
+    adim: 384         # attention dimension
+    aheads: 2         # number of attention heads
+    elayers: 4        # number of encoder layers
+    eunits: 1536      # number of encoder ff units
+    dlayers: 4        # number of decoder layers
+    dunits: 1536      # number of decoder ff units
+    positionwise_layer_type: conv1d   # type of position-wise layer
+    positionwise_conv_kernel_size: 3  # kernel size of position wise conv layer
+    duration_predictor_layers: 2      # number of layers of duration predictor
+    duration_predictor_chans: 256     # number of channels of duration predictor
+    duration_predictor_kernel_size: 3 # filter size of duration predictor
+    postnet_layers: 5                 # number of layers of postnset
+    postnet_filts: 5                  # filter size of conv layers in postnet
+    postnet_chans: 256                # number of channels of conv layers in postnet
+    use_scaled_pos_enc: True          # whether to use scaled positional encoding
+    encoder_normalize_before: True    # whether to perform layer normalization before the input
+    decoder_normalize_before: True    # whether to perform layer normalization before the input
+    reduction_factor: 1               # reduction factor
+    init_type: xavier_uniform         # initialization type
+    init_enc_alpha: 1.0               # initial value of alpha of encoder scaled position encoding
+    init_dec_alpha: 1.0               # initial value of alpha of decoder scaled position encoding
+    transformer_enc_dropout_rate: 0.2            # dropout rate for transformer encoder layer
+    transformer_enc_positional_dropout_rate: 0.2 # dropout rate for transformer encoder positional encoding
+    transformer_enc_attn_dropout_rate: 0.2       # dropout rate for transformer encoder attention layer
+    transformer_dec_dropout_rate: 0.2            # dropout rate for transformer decoder layer
+    transformer_dec_positional_dropout_rate: 0.2 # dropout rate for transformer decoder positional encoding
+    transformer_dec_attn_dropout_rate: 0.2       # dropout rate for transformer decoder attention layer
+    pitch_predictor_layers: 5                  # number of conv layers in pitch predictor
+    pitch_predictor_chans: 256                 # number of channels of conv layers in pitch predictor
+    pitch_predictor_kernel_size: 5             # kernel size of conv leyers in pitch predictor
+    pitch_predictor_dropout: 0.5               # dropout rate in pitch predictor
+    pitch_embed_kernel_size: 1                 # kernel size of conv embedding layer for pitch
+    pitch_embed_dropout: 0.0                   # dropout rate after conv embedding layer for pitch
+    stop_gradient_from_pitch_predictor: True   # whether to stop the gradient from pitch predictor to encoder
+    energy_predictor_layers: 2                 # number of conv layers in energy predictor
+    energy_predictor_chans: 256                # number of channels of conv layers in energy predictor
+    energy_predictor_kernel_size: 3            # kernel size of conv leyers in energy predictor
+    energy_predictor_dropout: 0.5              # dropout rate in energy predictor
+    energy_embed_kernel_size: 1                # kernel size of conv embedding layer for energy
+    energy_embed_dropout: 0.0                  # dropout rate after conv embedding layer for energy
+    stop_gradient_from_energy_predictor: False # whether to stop the gradient from energy predictor to encoder
+    spk_embed_dim: 256                         # speaker embedding dimension
+    spk_embed_integration_type: concat         # speaker embedding integration type
+
+
+
+###########################################################
+#                       UPDATER SETTING                   #
+###########################################################
+updater:
+    use_masking: True                 # whether to apply masking for padded part in loss calculation
+
+
+###########################################################
+#                     OPTIMIZER SETTING                   #
+###########################################################
+optimizer:
+    optim: adam               # optimizer type
+    learning_rate: 0.001      # learning rate
+
+###########################################################
+#                     TRAINING SETTING                    #
+###########################################################
+max_epoch: 1000
+num_snapshots: 5
+
+
+###########################################################
+#                       OTHER SETTING                     #
+###########################################################
+seed: 10086

examples/canton/tts3/local/preprocess.sh

+#!/bin/bash
+
+stage=0
+stop_stage=100
+
+config_path=$1
+
+if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
+    # get durations from MFA's result
+    echo "Generate durations.txt from MFA results ..."
+    python3 ${MAIN_ROOT}/utils/gen_duration_from_textgrid.py \
+        --inputdir=./canton_alignment \
+        --output durations.txt \
+        --config=${config_path}
+fi
+
+if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
+    # extract features
+    echo "Extract features ..."
+    python3 ${BIN_DIR}/preprocess.py \
+        --dataset=canton \
+        --rootdir=~/datasets/canton_all \
+        --dumpdir=dump \
+        --dur-file=durations.txt \
+        --config=${config_path} \
+        --num-cpu=20 \
+        --cut-sil=True
+fi
+
+if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
+    # get features' stats(mean and std)
+    echo "Get features' stats ..."
+    python3 ${MAIN_ROOT}/utils/compute_statistics.py \
+        --metadata=dump/train/raw/metadata.jsonl \
+        --field-name="speech"
+
+    python3 ${MAIN_ROOT}/utils/compute_statistics.py \
+        --metadata=dump/train/raw/metadata.jsonl \
+        --field-name="pitch"
+
+    python3 ${MAIN_ROOT}/utils/compute_statistics.py \
+        --metadata=dump/train/raw/metadata.jsonl \
+        --field-name="energy"
+fi
+
+if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
+    # normalize and covert phone/speaker to id, dev and test should use train's stats
+    echo "Normalize ..."
+    python3 ${BIN_DIR}/normalize.py \
+        --metadata=dump/train/raw/metadata.jsonl \
+        --dumpdir=dump/train/norm \
+        --speech-stats=dump/train/speech_stats.npy \
+        --pitch-stats=dump/train/pitch_stats.npy \
+        --energy-stats=dump/train/energy_stats.npy \
+        --phones-dict=dump/phone_id_map.txt \
+        --speaker-dict=dump/speaker_id_map.txt
+
+    python3 ${BIN_DIR}/normalize.py \
+        --metadata=dump/dev/raw/metadata.jsonl \
+        --dumpdir=dump/dev/norm \
+        --speech-stats=dump/train/speech_stats.npy \
+        --pitch-stats=dump/train/pitch_stats.npy \
+        --energy-stats=dump/train/energy_stats.npy \
+        --phones-dict=dump/phone_id_map.txt \
+        --speaker-dict=dump/speaker_id_map.txt
+
+    python3 ${BIN_DIR}/normalize.py \
+        --metadata=dump/test/raw/metadata.jsonl \
+        --dumpdir=dump/test/norm \
+        --speech-stats=dump/train/speech_stats.npy \
+        --pitch-stats=dump/train/pitch_stats.npy \
+        --energy-stats=dump/train/energy_stats.npy \
+        --phones-dict=dump/phone_id_map.txt \
+        --speaker-dict=dump/speaker_id_map.txt
+fi

examples/canton/tts3/local/synthesize.sh

+../../../aishell3/tts3/local/synthesize.sh
\ No newline at end of file

examples/canton/tts3/local/train.sh

+../../../aishell3/tts3/local/train.sh
\ No newline at end of file

examples/canton/tts3/path.sh

+../../csmsc/tts3/path.sh
\ No newline at end of file

examples/canton/tts3/run.sh

+#!/bin/bash
+
+set -e
+source path.sh
+
+gpus=0
+stage=0
+stop_stage=100
+
+conf_path=conf/default.yaml
+train_output_path=exp/default
+ckpt_name=snapshot_iter_112793.pdz
+
+# with the following command, you can choose the stage range you want to run
+# such as `./run.sh --stage 0 --stop-stage 0`
+# this can not be mixed use with `$1`, `$2` ...
+source ${MAIN_ROOT}/utils/parse_options.sh || exit 1
+
+if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
+    # prepare data
+    ./local/preprocess.sh ${conf_path} || exit -1
+fi
+
+if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
+    # train model, all `ckpt` under `train_output_path/checkpoints/` dir
+    CUDA_VISIBLE_DEVICES=${gpus} ./local/train.sh ${conf_path} ${train_output_path} || exit -1
+fi
+
+if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
+    # synthesize, vocoder is pwgan by default
+    CUDA_VISIBLE_DEVICES=${gpus} ./local/synthesize.sh ${conf_path} ${train_output_path} ${ckpt_name} || exit -1
+fi
+
+if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
+    # synthesize_e2e, vocoder is pwgan by default
+    CUDA_VISIBLE_DEVICES=${gpus} ./local/synthesize_e2e.sh ${conf_path} ${train_output_path} ${ckpt_name} || exit -1
+fi
+
+if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
+    # inference with static model, vocoder is pwgan by default
+    CUDA_VISIBLE_DEVICES=${gpus} ./local/inference.sh ${train_output_path} || exit -1
+fi
+
+if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
+    # install paddle2onnx
+    version=$(echo `pip list |grep "paddle2onnx"` |awk -F" " '{print $2}')
+    if [[ -z "$version" || ${version} != '1.0.0' ]]; then
+        pip install paddle2onnx==1.0.0
+    fi
+    ./local/paddle2onnx.sh ${train_output_path} inference inference_onnx fastspeech2_aishell3
+    # considering the balance between speed and quality, we recommend that you use hifigan as vocoder
+    ./local/paddle2onnx.sh ${train_output_path} inference inference_onnx pwgan_aishell3
+    # ./local/paddle2onnx.sh ${train_output_path} inference inference_onnx hifigan_aishell3
+    
+fi
+
+# inference with onnxruntime, use fastspeech2 + pwgan by default
+if [ ${stage} -le 6 ] && [ ${stop_stage} -ge 6 ]; then
+    ./local/ort_predict.sh ${train_output_path}
+fi
+
+if [ ${stage} -le 7 ] && [ ${stop_stage} -ge 7 ]; then
+    ./local/export2lite.sh ${train_output_path} inference pdlite fastspeech2_aishell3 x86
+    ./local/export2lite.sh ${train_output_path} inference pdlite pwgan_aishell3 x86
+    # ./local/export2lite.sh ${train_output_path} inference pdlite hifigan_aishell3 x86
+fi
+
+if [ ${stage} -le 8 ] && [ ${stop_stage} -ge 8 ]; then
+    CUDA_VISIBLE_DEVICES=${gpus} ./local/lite_predict.sh ${train_output_path} || exit -1
+fi

paddlespeech/t2s/datasets/get_feats.py

@@ -102,7 +102,7 @@ class Pitch():
 
     def _convert_to_continuous_f0(self, f0: np.ndarray) -> np.ndarray:
         if (f0 == 0).all():
-            print("All frames seems to be unvoiced.")
+            print("All frames seems to be unvoiced, this utt will be removed.")
             return f0
 
         # padding start and end of f0 sequence

paddlespeech/t2s/exps/fastspeech2/preprocess.py

@@ -54,8 +54,15 @@ def process_sentence(config: Dict[str, Any],
     record = None
     if utt_id in sentences:
         # reading, resampling may occur
-        wav, _ = librosa.load(str(fp), sr=config.fs)
-        if len(wav.shape) != 1:
+        wav, _ = librosa.load(
+            str(fp), sr=config.fs,
+            mono=False) if "canton" in str(fp) else librosa.load(
+                str(fp), sr=config.fs)
+        if len(wav.shape) == 2 and "canton" in str(fp):
+            # Remind that Cantonese datasets should be placed in ~/datasets/canton_all. Otherwise, it may cause problem.
+            wav = wav[0]
+            wav = np.ascontiguousarray(wav)
+        elif len(wav.shape) != 1:
             return record
         max_value = np.abs(wav).max()
         if max_value > 1.0:
@@ -102,6 +109,8 @@ def process_sentence(config: Dict[str, Any],
         np.save(mel_path, logmel)
         # extract pitch and energy
         f0 = pitch_extractor.get_pitch(wav, duration=np.array(durations))
+        if (f0 == 0).all():
+            return None
         assert f0.shape[0] == len(durations)
         f0_dir = output_dir / "data_pitch"
         f0_dir.mkdir(parents=True, exist_ok=True)
@@ -282,7 +291,20 @@ def main():
                 test_wav_files += wav_files[-sub_num_dev:]
             else:
                 train_wav_files += wav_files
-
+    elif args.dataset == "canton":
+        sub_num_dev = 5
+        wav_dir = rootdir / "WAV"
+        train_wav_files = []
+        dev_wav_files = []
+        test_wav_files = []
+        for speaker in os.listdir(wav_dir):
+            wav_files = sorted(list((wav_dir / speaker).rglob("*.wav")))
+            if len(wav_files) > 100:
+                train_wav_files += wav_files[:-sub_num_dev * 2]
+                dev_wav_files += wav_files[-sub_num_dev * 2:-sub_num_dev]
+                test_wav_files += wav_files[-sub_num_dev:]
+            else:
+                train_wav_files += wav_files
     elif args.dataset == "ljspeech":
         wav_files = sorted(list((rootdir / "wavs").rglob("*.wav")))
         # split data into 3 sections