Skip to content

Latest commit

 

History

History
286 lines (220 loc) · 8.53 KB

DATASET_SCHEMA.md

File metadata and controls

286 lines (220 loc) · 8.53 KB

Spatial LibriSpeech Schema

This document describes the metadata schema for Spatial LibriSpeech. The metadata is organised as columns of a pandas' DataFrame. We first describe how to load the dataframe, and then list all its columns.

Loading the metadata

Once you have downloaded any the dataset you can load the dataset with the following command:

import pandas as pd
metadata = pd.read_parquet("<path_to_dataset>/metadata.parquet")

Alternatively, you can download the metadata directly from the internet:

import pandas as pd
metadata = pd.read_parquet("https://docs-assets.developer.apple.com/ml-research/datasets/spatial-librispeech/v1/metadata.parquet")

The rest of this document describes the metadata schema for Spatial LibriSpeech (organised by columns in the dataframe)

Sample identification metadata

  • sample_id

    • Type: int
    • Range: [0, 22]
    • Dataloader feature: spatial_librispeech.Feature.SAMPLE_ID
    • Numeric identifier for sample. Corresponding audio files will be named {sample_id:06}.flac.

  • split

    • Type: string
    • Values: [train, test]
    • Determines whether current sample belongs to the train or the test set.

  • lite_version

    • Type: boolean
    • If true, current sample is part of the lite version of the dataset.

Acoustics metadata

  • acoustics/frequency_bins

    • Type: numpy.array of 33 floats
    • Unit: hertz
    • Dataloader feature: spatial_librispeech.Feature.FREQUENCY_BINS
    • Mean frequency values of the third octave bins used for all acoustic features.

  • acoustics/c50_db

    • Type: numpy.array of 33 floats
    • Unit: decibels
    • Dataloader feature: spatial_librispeech.Feature.C50_DB
    • Third octave values of speech clarity (C50) in decibels.

  • acoustics/drr_db

    • Type: numpy.array of 33 floats
    • Unit: decibels
    • Dataloader feature: spatial_librispeech.Feature.DRR_DB
    • Third octave values of direct-to-reverberant-ratio (DRR) in decibels.

  • acoustics/edt_ms

    • Type: numpy.array of 33 floats
    • Unit: milliseconds
    • Dataloader feature: spatial_librispeech.Feature.EDT_MS
    • Third octave values of early-decay time (EDT) in milliseconds.

  • acoustics/t20_ms

    • Type: numpy.array of 33 floats
    • Unit: milliseconds
    • Dataloader feature: spatial_librispeech.Feature.T20_MS
    • Third octave values of 20dB decay duration (T20) in milliseconds.

  • acoustics/t30_ms

    • Type: numpy.array of 33 floats
    • Unit: milliseconds
    • Dataloader feature: spatial_librispeech.Feature.T30_MS
    • Third octave values of 30dB decay duration (T30) in milliseconds.

Audio information

  • audio_info/duration

    • Type: float
    • Unit: seconds
    • Sample duration in seconds.

  • audio_info/frames

    • Type: int
    • Unit: seconds
    • Number of frames in audio sample. Note sample rate is 16kHz.

  • audio_info/size/ambisonics

    • Type: int
    • Unit: bytes
    • Size of speech ambisonics sample in bytes.

  • audio_info/size/noise_ambisonics

    • Type: int
    • Unit: bytes
    • Size of distractor noise ambisonics sample in bytes.

  • audio_info/checksum/ambisonics

    • Type: string
    • Hexadecimal representation of sha-256 checksum of speech ambisonics sample.

  • audio_info/checksum/noise_ambisonics

    • Type: string
    • Hexadecimal representation of sha-256 checksum of distractor noise ambisonics sample.

Speech metadata

  • speech/azimuth

    • Type: float
    • Unit: radians
    • Dataloader feature: spatial_librispeech.Feature.AZIMUTH
    • Horizontal angle between speech source and microphone array.

  • speech/elevation

    • Type: float
    • Unit: radians
    • Dataloader feature: spatial_librispeech.Feature.ELEVATION
    • Vertical angle between speech source and microphone array.

  • speech/distance

    • Type: float
    • Unit: meters
    • Dataloader feature: spatial_librispeech.Feature.DISTANCE
    • Distance between speech source and microphone array.

  • speech/speaking_azimuth

    • Type: float
    • Unit: radians
    • Dataloader feature: spatial_librispeech.Feature.SPEAKING_AZIMUTH
    • Horizontal rotation of speech source with respect to microphone array. Zero is in direct line of array.

  • speech/speaking_elevation

    • Type: float
    • Unit: radians
    • Dataloader feature: spatial_librispeech.Feature.SPEAKING_ELEVATION
    • Vertical rotation of speech source with respect to microphone array. Zero is in direct line of array.

  • speech/mrp

    • Type: float
    • Unit: decibels at random active speech level (dB-ASL)
    • Signal power for speech at mouth reference point (25 mm in front of the lip plane,cf. ITU-T P.58)

  • speech/source_id

    • Type: float
    • Range: [1, 20]
    • Numeric identifier for speech source location in the room.

  • speech/directivity_id

    • Type: int
    • Range: [0, 15]
    • Dataloader feature: spatial_librispeech.Feature.DIRECTIVITY_ID
    • Numeric identifier for the different directivity profiles applied to speech.

Original LibriSpeech metadata

  • speech/librispeech_metadata/book_id"

    • Type: int
      • Dataloader feature: spatial_librispeech.Feature.BOOK_ID
    • Numeric identifier for book being read.

  • speech/librispeech_metadata/chapter_id"

    • Type: int
    • Dataloader feature: spatial_librispeech.Feature.CHAPTER_ID
    • Numeric identifier for chapter being read.

  • speech/librispeech_metadata/chapter_title"

    • Type: string
    • Name of chapter being read.

  • speech/librispeech_metadata/project_id"

    • Type: int
    • Dataloader feature: spatial_librispeech.Feature.PROJECT_ID
    • Numeric identifier for project being read.

  • speech/librispeech_metadata/project_title"

    • Type: string
    • Name of project being read.

  • speech/librispeech_metadata/reader_id"

    • Type: int
    • Numeric identifier for reader.

  • speech/librispeech_metadata/reader_name"

    • Type: string
    • Name (or alias) of reader.

  • speech/librispeech_metadata/reader_sex"

    • Type: string
    • Dataloader feature: spatial_librispeech.Feature.READER_SEX
    • Values: [m, f] # TODO: verify these
    • Sex of reader.

  • speech/librispeech_metadata/sequence_number"

    • Type: int
    • Dataloader feature: spatial_librispeech.Feature.SEQUENCE_NUMBER
    • Sequence identifier for utterance (will be ordered for same project, book, and chapter).

  • speech/librispeech_metadata/transcription"

    • Type: string
    • Dataloader feature: spatial_librispeech.Feature.TRANSCRIPTION
    • Transcription of text being read.

  • speech/librispeech_metadata/subset"

    • Type: string
    • Values: [train-clean-100, train-clean-360, test-clean, test-other]
    • Original librispeech subset for utterance.

Noise metadata

  • noise/azimuth

    • Type: float
    • Unit: radians
    • Dataloader feature: spatial_librispeech.Feature.NOISE_AZIMUTH
    • Horizontal angle between distractor noise source and microphone array.

  • noise/elevation

    • Type: float
    • Unit: radians
    • Dataloader feature: spatial_librispeech.Feature.NOISE_ELEVATION
    • Vertical angle between distractor noise source and microphone array.

  • noise/distance

    • Type: float
    • Unit: meters
    • Dataloader feature: spatial_librispeech.Feature.NOISE_DISTANCE
    • Distance between distractor noise source and microphone array.

  • noise/snr

    • Type: float
    • Unit: decibels
    • Dataloader feature: spatial_librispeech.Feature.SNR_DB
    • Signal-to-noise ratio for distractor noise in decibels.

  • noise/source_id

    • Type: float
    • Range: [1, 20]
    • Numeric identifier for distractor noise source location in the room.

  • noise/deep_noise_suppression_metadata/filename

    • Type: string
    • Filename of noise in deep noise suppression.

  • noise/deep_noise_suppression_metadata/is_audioset

    • Type: boolean
    • If true, noise is part of Audioset dataset, otherwise it is part of Freesound dataset.

  • noise/deep_noise_suppression_metadata/is_audioset

    • Type: string
    • Comma-separated labels for noise, for Audioset you may need to consult lookup table.

  • noise/deep_noise_suppression_metadata/youtube_id

    • Type: string
    • Only available for Audioset samples, youtube id of video from where noise was extracted.

Room metadata

  • room/room_id

    • Type: int
    • Dataloader feature: spatial_librispeech.Feature.ROOM_ID
    • Numeric identifier for simulated simulated room.

  • room/floor_area

    • Type: float
    • Unit: squared meters
    • Room's floor surface area.

  • room/surface_area

    • Type: float
    • Unit: squared meters
    • Sum of the area of all surface (floor, walls, ceiling) in the room.

  • room/volume

    • Type: float
    • Unit: cubic meters
    • Room's total volume.