Add dataset 'AISHELL-3' from OpenSLR for training mandarin TTS

[RobinDong]

What does this PR do ?

Add dataset 'AISHELL-3' from OpenSLR for training FastPitch model.

Changelog

• Add dataset 'AISHELL-3' from OpenSLR for training mandarin TTS

Usage

1. create directories

mkdir data_aishell3 mani_aishell3 sup_aishell3

2. install NeMo packages (doc)
3. run scripts to get data, process data and train the model from the data

python3 scripts/dataset_processing/tts/aishell3/get_data.py \
    --data-root data_aishell3 \
    --val-size 0.1 \
    --test-size 0.2 \
    --seed-for-ds-split 100 \
    --manifests-path mani_aishell3

python3 scripts/dataset_processing/tts/extract_sup_data.py \
    --config-path aishell3/ds_conf \
    --config-name ds_for_fastpitch_align.yaml \
    manifest_filepath=mani_aishell3/train_manifest.json \
    sup_data_path=sup_aishell3

python3 examples/tts/fastpitch.py --config-path conf/zh/ \
    --config-name fastpitch_align_multispeaker_22050.yaml \
    model.train_ds.dataloader_params.batch_size=32 \
    model.validation_ds.dataloader_params.batch_size=8 \
    train_dataset=mani_aishell3/train_manifest.json \
    validation_datasets=mani_aishell3/val_manifest.json \
    sup_data_path=sup_aishell3 \
    exp_manager.exp_dir=result \
    trainer.max_epochs=160 \
    trainer.check_val_every_n_epoch=1 \
    pitch_mean=214.61354064941406 \
    pitch_std=64.61677551269531 \
    +exp_manager.create_wandb_logger=true \
    +exp_manager.wandb_logger_kwargs.name="tutorial" \
    +exp_manager.wandb_logger_kwargs.project="aishell"

Sample

The FastPitch model trained by AISHELL-3 dataset is here. You can download it and run the snippet below:

from nemo.collections.tts.models.base import SpectrogramGenerator, Vocoder
from nemo.collections.tts.models import FastPitchModel, HifiGanModel

text = "这些新一代的中央处理器不只效能惊人，价格也十分有竞争力。"
device = "cpu"
fastpitch_model = FastPitchModel.restore_from("~/Downloads/aishell3.nemo").eval().to(device)
model = HifiGanModel.from_pretrained(model_name="tts_zh_hifigan_sfspeech").eval().to(device)

# Normalize the text and convert it into individual phonemes/tokens.
tokens = fastpitch_model.parse(text)

# Generate spectrogram from text
spectrogram = fastpitch_model.generate_spectrogram(tokens=tokens, speaker=5)

# Invert the spectrogram into audio samples
audio = model.convert_spectrogram_to_audio(spec=spectrogram)

# Convert output from pytorch tensor to numpy array
audio = audio.cpu().detach().numpy()

import soundfile as sf

sf.write("sample.wav", audio.T, 22050, format="WAV")

The sample.wav is here: aishell3.zip
Change the speaker id to 1863 speaker=1863 would make a male sound.

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you add or update any necessary documentation?
• Does the PR affect components that are optional to install? (Ex: Numba, Pynini, Apex etc)
  • Reviewer: Does the PR have correct import guards for all optional libraries?

PR Type:

• New Feature
• Bugfix
• Documentation

If you haven't finished some of the above items you can still open "Draft" PR.

Who can review?

@blisc @okuchaiev