README.md

---
pretty_name: LibriTTS Corpus with Forced Alignments
annotations_creators:
- crowdsourced
language: en
tags:
- speech
- audio
- automatic-speech-recognition
- text-to-speech
license:
- cc-by-4.0
task_categories:
- automatic-speech-recognition
- text-to-speech
extra_gated_prompt: "When using this dataset to download LibriTTS, you agree to the terms on https://www.openslr.org"
---

> This dataset is identical to **[cdminix/libritts-aligned](https://huggingface.co/datasets/cdminix/libritts-aligned)** except it uses the newly released LibriTTS-R corpus. Please cite **[Y. Koizumi, et al., "LibriTTS-R: Restoration of a Large-Scale Multi-Speaker TTS Corpus", Interspeech 2023](https://google.github.io/df-conformer/librittsr/)**

*When using this dataset to download LibriTTS-R, make sure you agree to the terms on https://www.openslr.org*

# Dataset Card for LibriTTS-R with Forced Alignments (and Measures)

This dataset downloads LibriTTS-R and preprocesses it on your machine to create alignments using [montreal forced aligner](https://montreal-forced-aligner.readthedocs.io/en/latest/).
You need to run ``pip install alignments phones`` before using this dataset.
When running this the first time, it can take an hour or two, but subsequent runs will be lightning fast.

## Requirements
- ``pip install alignments phones`` **(required)**
- ``pip install speech-collator`` (optional)

*Note: version >=0.0.15 of alignments is required for this corpus*

## Example Item

```json
{
    'id': '100_122655_000073_000002.wav',
    'speaker': '100',
    'text': 'the day after, diana and mary quitted it for distant b.',
    'start': 0.0,
    'end': 3.6500000953674316, 
    'phones': ['[SILENCE]', 'ð', 'ʌ', '[SILENCE]', 'd', 'eɪ', '[SILENCE]', 'æ', 'f', 't', 'ɜ˞', '[COMMA]', 'd', 'aɪ', 'æ', 'n', 'ʌ', '[SILENCE]', 'æ', 'n', 'd', '[SILENCE]', 'm', 'ɛ', 'ɹ', 'i', '[SILENCE]', 'k', 'w', 'ɪ', 't', 'ɪ', 'd', '[SILENCE]', 'ɪ', 't', '[SILENCE]', 'f', 'ɜ˞', '[SILENCE]', 'd', 'ɪ', 's', 't', 'ʌ', 'n', 't', '[SILENCE]', 'b', 'i', '[FULL STOP]'], 
    'phone_durations': [5, 2, 4, 0, 5, 13, 0, 16, 7, 5, 20, 2, 6, 9, 15, 4, 2, 0, 11, 3, 5, 0, 3, 8, 9, 8, 0, 13, 3, 5, 3, 6, 4, 0, 8, 5, 0, 9, 5, 0, 7, 5, 6, 7, 4, 5, 10, 0, 3, 35, 9],
    'audio': '/dev/shm/metts/train-clean-360-alignments/100/100_122655_000073_000002.wav'
}
```

The phones are IPA phones, and the phone durations are in frames (assuming a hop length of 256, sample rate of 22050 and window length of 1024). These attributes can be changed using the ``hop_length``, ``sample_rate`` and ``window_length`` arguments to ``LibriTTSAlign``.

## Data Collator

This dataset comes with a data collator which can be used to create batches of data for training.
It can be installed using ``pip install speech-collator`` ([MiniXC/speech-collator](https://www.github.com/MiniXC/speech-collator)) and can be used as follows:

```python
import json
from datasets import load_dataset
from speech_collator import SpeechCollator
from torch.utils.data import DataLoader

dataset = load_dataset('cdminix/libritts-aligned', split="train")

speaker2ixd = json.load(open("speaker2idx.json"))
phone2ixd = json.load(open("phone2idx.json"))

collator = SpeechCollator(
    speaker2ixd=speaker2idx,
    phone2ixd=phone2idx ,
)
dataloader = DataLoader(dataset, collate_fn=collator.collate_fn, batch_size=8)
```

You can either download the ``speaker2idx.json`` and ``phone2idx.json`` files from [here](https://huggingface.co/datasets/cdminix/libritts-aligned/tree/main/data) or create them yourself using the following code:

```python
import json
from datasets import load_dataset
from speech_collator import SpeechCollator, create_speaker2idx, create_phone2idx

dataset = load_dataset("cdminix/libritts-aligned", split="train")

# Create speaker2idx and phone2idx
speaker2idx = create_speaker2idx(dataset, unk_idx=0)
phone2idx = create_phone2idx(dataset, unk_idx=0)

# save to json
with open("speaker2idx.json", "w") as f:
    json.dump(speaker2idx, f)
with open("phone2idx.json", "w") as f:
    json.dump(phone2idx, f)
```

### Measures

When using ``speech-collator`` you can also use the ``measures`` argument to specify which measures to use. The following example extracts Pitch and Energy on the fly.

```python
import json
from torch.utils.data import DataLoader
from datasets import load_dataset
from speech_collator import SpeechCollator, create_speaker2idx, create_phone2idx
from speech_collator.measures import PitchMeasure, EnergyMeasure

dataset = load_dataset("cdminix/libritts-aligned", split="train")

speaker2idx = json.load(open("data/speaker2idx.json"))
phone2idx = json.load(open("data/phone2idx.json"))

# Create SpeechCollator
speech_collator = SpeechCollator(
    speaker2idx=speaker2idx,
    phone2idx=phone2idx,
    measures=[PitchMeasure(), EnergyMeasure()],
    return_keys=["measures"]
)

# Create DataLoader
dataloader = DataLoader(
    dataset,
    batch_size=8,
    collate_fn=speech_collator.collate_fn,
)
```

COMING SOON: Detailed documentation on how to use the measures at [MiniXC/speech-collator](https://www.github.com/MiniXC/speech-collator).

## Splits

This dataset has the following splits:
- ``train``: All the training data, except one sample per speaker which is used for validation.
- ``dev``: The validation data, one sample per speaker.
- ``train.clean.100``: Training set derived from the original materials of the train-clean-100 subset of LibriSpeech.
- ``train.clean.360``: Training set derived from the original materials of the train-clean-360 subset of LibriSpeech.
- ``train.other.500``: Training set derived from the original materials of the train-other-500 subset of LibriSpeech.
- ``dev.clean``: Validation set derived from the original materials of the dev-clean subset of LibriSpeech.
- ``dev.other``: Validation set derived from the original materials of the dev-other subset of LibriSpeech.
- ``test.clean``: Test set derived from the original materials of the test-clean subset of LibriSpeech.
- ``test.other``: Test set derived from the original materials of the test-other subset of LibriSpeech.

## Environment Variables

There are a few environment variable which can be set.

- ``LIBRITTS_VERBOSE``: If set, will print out more information about the dataset creation process.
- ``LIBRITTS_MAX_WORKERS``: The number of workers to use when creating the alignments. Defaults to ``cpu_count()``.
- ``LIBRITTS_PATH``: The path to download LibriTTS to. Defaults to the value of ``HF_DATASETS_CACHE``.

# Citation

When using LibriTTS-R please cite the following papers:
- [LibriTTS-R: Restoration of a Large-Scale Multi-Speaker TTS Corpus](https://google.github.io/df-conformer/librittsr/)
- [LibriTTS: A Corpus Derived from LibriSpeech for Text-to-Speech](https://arxiv.org/abs/1904.02882)
- [Montreal Forced Aligner: Trainable text-speech alignment using Kaldi](https://www.researchgate.net/publication/319185277_Montreal_Forced_Aligner_Trainable_Text-Speech_Alignment_Using_Kaldi)

When using the Measures please cite the following paper (ours):
- [Evaluating and reducing the distance between synthetic and real speech distributions](https://arxiv.org/abs/2211.16049)