Default Format

This describes, how a corpus with the default format is saved on disk. Every corpus is a folder with a bunch of files.


This file contains a list of every audio file in the corpus. Every file is identified by a unique id. Every line in the file contains the mapping from file-id to the file-path for a single file. The filepath is the path to the audio file relative to the corpus folder.

<recording-id> <wav-file-path>


2014-03-17-09-45-16_Kinect-Beam train/2014-03-17-09-45-16_Kinect-Beam.wav
2014-03-17-09-45-16_Realtek train/2014-03-17-09-45-16_Realtek.wav
2014-03-17-09-45-16_Yamaha train/2014-03-17-09-45-16_Yamaha.wav
2014-03-17-10-26-07_Realtek train/2014-03-17-10-26-07_Realtek.wav


This file contains all utterances in the corpus. An utterance is a part of a file (A file can contain one or more utterances). Every line in this file defines a single utterance, which consists of utterance-id, file-id, start and end. Start and end are measured in seconds within the file. If end is -1 it is considered to be the end of the file (If the utterance is the full length of the file, start and end are 0/-1).

<utterance-id> <recording-id> <start> <end>


1_hello 2014-03-17-09-45-16_Kinect-Beam
1_hello_sam 2014-03-17-09-45-16_Realtek 0 -1
2_this_is 2014-03-17-09-45-16_Yamaha 0 5
3_goto 2014-03-17-09-45-16_Yamaha 5 -1


This file contains the mapping from utterance to issuers, which gives the information who/what is the origin of a given utterance (e.g. the speaker). Every line contains one mapping from utterance-id to issuer-id.

<utterance-id> <issuer-id>


1_hello marc
1_hello_sam marc
2_this_is sam
3_goto jenny


This file contains additional information about the issuers. Depending on the type of the issuer (Issuer, Speaker, Artist which are defined via type parameter) different parameters can be set.

  • info: An arbitrary info dictionary.
  • gender (MALE/FEMALE): The gender of a speaker
  • age_group (child, youth, adult, senior): The age group of a speaker
  • native_language (language code ISO 639-3): The language natively spoken
  • name: The name of the artist/band/group
  "speaker-1": {
    "info": {},
    "type": "speaker",
    "gender": "MALE",
    "age_group": "ADULT",
    "native_language": "deu"
  "speaker-2": {
    "info": {},
    "type": "artist",
    "name": "Ohooo"
  "speaker-3": {
    "info": {
      "region": "zh"


There can be multiple label-lists in a corpus (e.g. text-transcription, raw-text-transcription - with punctuation, audio classification type, …). Every label-list is saved in a separate file with the prefix labels_. A single file contains labels of a specific type for all utterances. A label-list of an utterance can contain one or more labels (e.g. in a text segmentation every word could be a label). A label optionally can have a start and end time (in seconds within the utterance). For labels without start/end defined 0/-1 is set. Every line in the file defines one label. The labels are stored in order per utterance (e.g. 1. word, 2. word, 3. word, …). Optionally addtional meta-information can be stored per label. This has to be a json string in square brackets.

<utterance-id> <start> <end> <label-value> [<label-meta>]


1_hello 0 -1 hi
1_hello 0 -1 this
1_hello 0 -1 is
1_hello_sam 0 -1 hello
1_hello_sam 0 -1 sam
2_this_is 0 -1 this
2_this_is 0 -1 is [{"prio": 3}]
2_this_is 0 -1 me [{"stress": true}]
3_goto 0 -1 go
3_goto 0 -1 to
3_goto 0 -1 the
3_goto 0 -1 mall


Contains a list of stored features. A corpus can have different feature containers. Every container contains the features of all utterances of a given type (e.g. MFCC features). A feature container is a h5py file which contains a dataset per utterance. Every line contains one container of features.

<feature-name> <relative-path>


mfcc mfcc_features
fbank fbank_features


Contains a list of tracks that are stored in audio-containers. Every entry consists of a track-id, the relative path to the container and a key that identifies the track in the audio-container.

<track-id> <audio-container-path> <audio-container-key>


track-1 ../audio.hdf5 track-1
track-2 ../audio.hdf5 track-2x