An exmpale of the data structure required by ESPnet. The required format can be found in data

|-- data
|   |-- test
|   |   |-- spk2utt  # Mapping a speaker-ID to a list of utterance-IDs
|   |   |-- text     # Mapping a utterance-ID to a text
|   |   |-- utt2spk  # Mappinng a utterance-ID to a speaker-ID
|   |   `-- wav.scp  # Mappinng a utterance-ID to a path of audio file
|   |-- train
|   |   |-- spk2utt
|   |   |-- text
|   |   |-- utt2spk
|   |   `-- wav.scp
|   `-- valid
|       |-- spk2utt
|       |-- text
|       |-- utt2spk
|       `-- wav.scp

ESPnet follows the data strcutre developed by Kaldi-asr: A data-directory must contain some texts, wav.scp, text, and etc. which have common format to describe DNN corpus. The format is space separated and must be two columns. The first column is taken as ID and the second is some value.

<ID> <value>
<ID> <value>
...

e.g. wav.scp shows <Sample ID> <Wave file path>.

• The directory name in data is arbitrary: train, valid, and test can be tr, cv, and eval for example.

• The path for wav.scp can be both absolute path or relative path from the base directory ( egs2/<corpus-name>/<task-name>, e.g. egs2/an4/asr1).

  uttidA /absolute/path/uttidA.wav
  uttidB ./relative/path/uttidB.wav
  

  • Maybe, we assume monaural and 16bit-signed-integer-pcm audio file. Any sampling rates are okay.
    • Please check your audio format in advance using sox or ffmpeg for example.

      soxi youraudio.wav
      

  • (THIS IS EXTENSION BY ESPNET, NOT KALDI FORMAT) Not only wav, but also flac can be used (Supported audio format can be extended)

    uttidA a.flac
    uttidB b.flac
    

  • wavs directory exists at the same level of data in this example, but this is not requirement.

• If you don't have speaker information, it can be dummy data because actually most recipes don't use speaker information.

  uttidA dummy
  uttidB dummy
  uttidC dummy
  ...
  

• spk2utt can be gernated from utt2spk using utils/utt2spk_to_spk2utt.pl, and utt2spk can be gernated from from spk2utt using utils/spk2utt_to_utt2spk.pl.

  utils/utt2spk_to_spk2utt.pl data/train/utt2spk > data/train/spk2utt
  utils/spk2utt_to_utt2spk.pl data/train/spk2utt > data/train/utt2spk

• To check and force your directory to satisfy expected format, use utils/validate_data_dir.sh and utils/fix_data_dir.sh

  # check
  utils/validate_data_dir.sh --no-feats data/train
  # Force format (This is irreversible change)
  utils/fix_data_dir.sh data/train
  

• It's okay to contains the other files (foo.txt in the following example) in the data directory. They are not referred.

  data
    `-- test
      |-- foo.txt
      |-- spk2utt
      |-- text 
      |-- utt2spk
      `-- wav.scp
  

|-- segments
|-- spk2utt
|-- text
|-- utt2spk
`-- wav.scp

If your audio data is long recording and each audio file includes multiple utterances, you need to put segments file to specify the start time and end time of each utterance. The format is <utterance_id> <wav_id> <start_time> <end_time> (in seconds).

sw02001-A_000098-001156 sw02001-A 0.98 11.56
...

Note that if using segments, wav.scp has <wav_id> instead of utterance_id.

sw02001-A /path/to/sw02001-A.wav
...

wav.scp provides a feature to describe the file format of wavfiles converted by an arbitral command without saving the files actually. The usage is as following:

foo_id some_command /path/to/foo.wav |

Note that in our style, the environment variable for scripts are set by path.sh, so please check that the command exists in the ${PATH}.

If a line ends with |, it indicates using this pipeline mechanism and our Python script derives the output data from the command via pipeline (We are using https://github.com/nttcslab-sp/kaldiio).

e.g. Change sampling rate, encoding, bits, etc.
sox input.wav -b16 -e unsigned-integer -r 16000 -t wav - |

e.g. Channel selection
sox stereo.wav -c 1 -t wav - |
sox stereo.wav -c 2 -t wav - |