LibriS2S

This repo contains scripts and alignment data to create a dataset build further upon librivoxDeEn such that it contains (German audio, German transcription, English audio, English transcription) quadruplets and can be used for Speech-to-Speech translation research. Because of this, the alignments are released under the same Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License

These alignments were collected by downloading the English audiobooks and using aeneas to align the book chapters to the transcripts. For more information read the original paper (Presented at LREC 2022)

Usage example

using huggingface

from datasets import load_dataset

ds = load_dataset(
   "PedroDKE/LibriS2S",
   trust_remote_code=True,
   streaming=True  # or streaming=False for full download
)
sample = next(iter(ds["train"]))
print(sample["DE_audio"])        # eg. {'path':…, 'array':…, 'sampling_rate':…}
print(sample["DE_transcript"])   # German transcript text

using pytorch

from datasets import load_dataset, Audio
from torch.utils.data import DataLoader

# Load the dataset in streaming mode
dataset = load_dataset("PedroDKE/LibriS2S", streaming=True, trust_remote_code=True)

# Audio casting
dataset = dataset.cast_column("DE_audio", Audio())
dataset = dataset.cast_column("EN_audio", Audio())

# Convert to iterable dataset (e.g. only use German audio here)
streaming_dataset = dataset["train"]

# Collate function (batching not supported for audio arrays directly)
def collate_fn(batch):
    return {
        "de_audio": [item["DE_audio"]["array"] for item in batch],
        "en_audio": [item["EN_audio"]["array"] for item in batch],
        "score": [item["score"] for item in batch],
    }

dataloader = DataLoader(streaming_dataset, batch_size=4, collate_fn=collate_fn)

for batch in dataloader:
    print("Batch shape (DE):", [x.shape for x in batch["de_audio"]])
    break

if you've cloned the dataset a more advanced PT loader can be found here.

The data

The English/German audio are available in the folder EN/DE respectively, can be downloaded from this HuggingFace repor or this onedrive. In case there are any problems with the download, feel free to open an issue here or on GitHub.
The repo structure is as follow:

• Alignments : Contains all the alignments for each book and chapter
• DE : Contains the German audio for each chapter per book.
• EN : Contains the English audio for each chapter per book.
• Example : contains example files on for the scraping and aligning explanations that were used to build this dataset.
• LibrivoxDeEn_alignments : Contains the base alignments from the LibrivoxDeEn dataset.
  

In case you feel a part of the data is missing, feel free to open an issue! The full zipfile is about 52 GB of size.

Scraping a book from Librivox

To download all chapters from a librivox url the following command can be used:

python scrape_audio_from_librivox.py \
--url https://librivox.org/undine-by-friedrich-de-la-motte-fouque/ \
--save_dir ./examples

Allign a book from Librivox with the text from LibrivoxDeEn

To allign the previously downloaded book with the txt files and tsv tables provided by LibrivoxDeEn the following command, based on the example provided with this repo, can be used:

python align_text_and_audio.py \
--text_dir ./example/en_text/ \
--audio_path ./example/audio_chapters/ \
--aeneas_path ./example/aeneas/ \
--en_audio_export_path ./example/sentence_level_audio/ \
--total_alignment_path ./example/bi-lingual-alignment/ \
--librivoxdeen_alignment ./example/undine_data.tsv \
--aeneas_head_max 120 \
--aeneas_tail_min 5 \

note: the example folder in this repo already contains the first two chapters from Undine scraped from librivox and their transcripts and (modified to only contain the first 2 chapters) tsv table retrieved from LibrivoxDeEn. Additional data to align can be scraped by using the same file shown previously and combined with the provided data from LibriVoxDeEn

Additionally with this repo the full alignment for the 8 following books with following LibrivoxDeEn id's are also given: 9, 10, 13, 18, 23, 108, 110, 120.

Other books such as 11, 36, 67 and 54 are also inside of the librivoxDeEn dataset but the chapters do not correspond in a 1:1 mannner(for example: the German version of book 67 has 27 chapters but the English version has 29 and thus need to be re-aligned before the allignment script in this repo will work). Therefore these alignments are given but might have be different if you scrape them yourselves as the re-alignments might be different for you.

Metrics on the alignment given in this repo.

Using the alignments given in this repo some metrics were collected and quickly displayed here, for this table and the next figure the books which were manually alligned, although provided in the zip, were not accounted for, but the full table can be found in the original paper.

	German	English
number of files	18868	18868
total time (hh:mm:ss)	39:11:08	40:52:31
Speakers	41	22

note: the speakers were counted for each book seperatly so some speakers might be counter more than once.

the number of hours for each book aligned in this repo:

Citation

when using this work, please cite the original paper and the LibrivoxDeEn authors

@inproceedings{jeuris-niehues-2022-libris2s,
   title = "{L}ibri{S}2{S}: A {G}erman-{E}nglish Speech-to-Speech Translation Corpus",
   author = "Jeuris, Pedro  and
     Niehues, Jan",
   booktitle = "Proceedings of the Thirteenth Language Resources and Evaluation Conference",
   month = jun,
   year = "2022",
   address = "Marseille, France",
   publisher = "European Language Resources Association",
   url = "https://aclanthology.org/2022.lrec-1.98",
   pages = "928--935",
   abstract = "Recently, we have seen an increasing interest in the area of speech-to-text translation. This has led to astonishing improvements in this area. In contrast, the activities in the area of speech-to-speech translation is still limited, although it is essential to overcome the language barrier. We believe that one of the limiting factors is the availability of appropriate training data. We address this issue by creating LibriS2S, to our knowledge the first publicly available speech-to-speech training corpus between German and English. For this corpus, we used independently created audio for German and English leading to an unbiased pronunciation of the text in both languages. This allows the creation of a new text-to-speech and speech-to-speech translation model that directly learns to generate the speech signal based on the pronunciation of the source language. Using this created corpus, we propose Text-to-Speech models based on the example of the recently proposed FastSpeech 2 model that integrates source language information. We do this by adapting the model to take information such as the pitch, energy or transcript from the source speech as additional input.",
}

@article{beilharz19,
 title = {LibriVoxDeEn: A Corpus for German-to-English Speech Translation and Speech Recognition},
 author = {Beilharz, Benjamin and Sun, Xin and Karimova, Sariya and Riezler, Stefan},
 journal = {Proceedings of the Language Resources and Evaluation Conference},
 journal-abbrev = {LREC},
 year = {2020},
 city = {Marseille, France},
 url = {https://arxiv.org/pdf/1910.07924.pdf}
}