[TTS][DE] Multi-speaker fastpitch model training recipe on HUI-Audio-Corpus-German

[XuesongYang]

What does this PR do ?

• created a dataset preparation recipe for HUI-ACG.
  • re-write get_data.py to support multiprocessing and decouple text normalization.
• added a script to convert graphemes into phonemes.
• created a model config for multi-speaker Fastpitch model training.

Collection: [Note which collection this PR will affect]

Changelog

• Add specific line by line info of high level changes in this PR.

Usage

• You can potentially add a usage example below

# Download the dataset, normalize texts, create train/val/test splits of JSON manifests.
$ python scripts/dataset_processing/tts/hui_acg/get_data.py \
        --data-root /home/xueyang/datasets \
        --normalize-text
$ ls /home/xueyang/datasets/HUI-Audio-Corpus-German-clean/*_text_normed.json
test_manifest_text_normed.json  train_manifest_text_normed.json  val_manifest_text_normed.json

# run phonemizer to convert graphemes into phonemes, resulting in new files with name like train_manifest_text_normed_phonemes.json
$ python scripts/dataset_processing/tts/hui_acg/phonemizer.py \
        --json-manifests /home/xueyang/datasets/HUI-Audio-Corpus-German-clean/train_manifest_text_normed.json \
                                     /home/xueyang/datasets/HUI-Audio-Corpus-German-clean/val_manifest_text_normed.json \
                                     /home/xueyang/datasets/HUI-Audio-Corpus-German-clean/test_manifest_text_normed.json \
        --preserve-punctuation
$ ls /home/xueyang/datasets/HUI-Audio-Corpus-German-clean/*_phonemes.json
test_manifest_text_normed_phonemes.json  train_manifest_text_normed_phonemes.json  val_manifest_text_normed_phonemes.json

$ python scripts/dataset_processing/tts/extract_sup_data.py \
        --config-path hui_acg/ds_conf \
        --config-name ds_for_fastpitch_align.yaml \
        manifest_filepath=/home/xueyang/datasets/HUI-Audio-Corpus-German-clean/Friedrich_Clean/train_manifest_text_normed_phonemes.json \
        sup_data_path=/home/xueyang/datasets/HUI-Audio-Corpus-German-clean/sup_data_phonemes

# you would see the information in the terminal
-- Set up ENVs...
a867b5ef4cd08fe0db216937b2d3e767fe71b3e7
-- Started ...
[NeMo I 2022-06-17 15:15:57 tokenize_and_classify:81] Creating ClassifyFst grammars. This might take some time...
[NeMo I 2022-06-17 15:16:15 data:186] Loading dataset from /data/HUI-Audio-Corpus-German-clean/train_manifest_text_normed_phonemes.json.
[NeMo I 2022-06-17 15:16:22 data:221] Loaded dataset with 173622 files.
[NeMo I 2022-06-17 15:16:22 data:223] Dataset contains 403.30 hours.
[NeMo I 2022-06-17 15:16:22 data:321] Pruned 0 files. Final dataset contains 173622 files
[NeMo I 2022-06-17 15:16:22 data:323] Pruned 0.00 hours. Final dataset contains 403.30 hours.
Processing /home/xueyang/datasets/HUI-Audio-Corpus-German-clean/train_manifest_text_normed_phonemes.json:
PITCH_MEAN=155.3916473388672, PITCH_STD=102.26656341552734
PITCH_MIN=65.4063949584961, PITCH_MAX=2093.004638671875

# train FastPitch model. copy & paste PITCH_MEAN, PITCH_STD, PITCH_MIN, and PITCH_MAX to the params.
$ python examples/tts/fastpitch.py \
        --config-path conf/de \
        --config-name fastpitch_align_44100 \
        model.train_ds.dataloader_params.num_workers=16 \
        model.validation_ds.dataloader_params.num_workers=16 \
        model.train_ds.dataloader_params.batch_size=24 \
        model.validation_ds.dataloader_params.batch_size=24 \
        train_dataset=/home/xueyang/datasets/HUI-Audio-Corpus-German-clean/train_manifest_text_normed_phonemes.json \
        validation_datasets=/home/xueyang/datasets/HUI-Audio-Corpus-German-clean/val_manifest_text_normed_phonemes.json \
        sup_data_path=/home/xueyang/datasets/HUI-Audio-Corpus-German-clean/sup_data_phonemes \
        pitch_mean=155.3916473388672 \
        pitch_std=102.26656341552734 \
        pitch_fmin=65.4063949584961 \
        pitch_fmax=2093.004638671875 \
        trainer.devices=2 \
        trainer.num_nodes=1 \
        trainer.accumulate_grad_batches=2 \
        trainer.log_every_n_steps=10 \
        exp_manager.exp_dir=/home/xueyang/experiments/GermanTTS/multi_spk_tts_de/fastpitch-train/results \
        exp_manager.resume_if_exists=True \
        exp_manager.resume_ignore_no_checkpoint=True \
        +exp_manager.create_wandb_logger=True \
        +exp_manager.wandb_logger_kwargs.name=fastpitch-train \
        +exp_manager.wandb_logger_kwargs.project=multi_spk_tts_de 

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?

Anyone in the community is free to review the PR once the checks have passed.
Contributor guidelines contains specific people who can review PRs to various areas.

Additional Information

• Related to # (issue)