mariannedhk/librispeech_phones

Summary

Phone annotations for the LibriSpeech corpus. This dataset can for example be used in combination with audio from the librispeech_asr dataset to extract phone embeddings from an audio encoder model.

Data sources

Phone start and end times are extracted from the LibriSpeech Alignments, obtained using the using the Montreal Forced Aligner by Lugosch et al. (2019). Phone position is derived from the same source, using the word alignments to enumerate phones within each word start and end time. missing_alignments.json lists identifiers of files in the LibriSpeech corpus for which alignments are not available (by dataset split). Speaker sex is inferred from the SPEAKERS.TXT metadata file released with the LibriSpeech corpus.

Columns

• phone phone label in ARPAbet transcription format (excluding stress marker)
• phone_stress phone label in ARPAbet transcription format (including stress marker)
• phone_ipa phone label in International Phonetic Alphabet transcription format
• phone_position phone position within a word
• start_time phone start time relative to audio file onset
• end_time phone end time relative to audio file onset
• speaker_id unique identifier for each speaker in the LibriSpeech corpus
• speaker_sex speaker sex as reported in the LibriSpeech metadata
• file_id unique identifier for each file in the LibriSpeech corpus
• subset subset of the LibriSpeech corpus

Example usage

from datasets import load_dataset

# download phone annotations for the full librispeech corpus
libri_phones = load_dataset("mariannedhk/librispeech_phones")

# download only phone annotations for the development sets
# similarly, specify "all_train" or "all_test" for downloading only phone annotations from the train or test sets, respectively
libri_dev_phones = load_dataset("mariannedhk/librispeech_phones", "all_dev")

# load annotations for only the dev.clean split
# (this may still download the full dataset first)
libri_dev_clean_phones = load_dataset("mariannedhk/librispeech", split="dev.clean")