Skip to content

Latest commit

 

History

History
82 lines (77 loc) · 3.83 KB

dataset_README.md

File metadata and controls

82 lines (77 loc) · 3.83 KB

Large High Quality TTS Datasets

These recordings were captured as a part of the T1 project of the language technology program. When completed, this dataset will contain 8 voices diverse in age, gender and dialect. We record approximately 20 hours of speech data with each voice.

Relationship to other projects

Reading List

The reading list was generated mainly from the Gigacorpus, further details are available in this paper. The 20 hour reading list generated contains 14400 unique sentences that were selected to maximize diphone coverage while minimizing sentence length. This list covers 81% of diphones at least 20 times. The list is available here.

LOBE

All recordings were captured in LOBE, our in-house recording client.

Technical details

Recordings

By default, the recordings are captured at 44.1K-48KHz with 16-32 bit sample size. It is not guaranteed that all samples are captured in the same way as some of the earlier recordings were captured with a different recording stack. The recordings start and end with approximately 1 second of silence

Corpus structure

Each directory in this datasets corresponds to one collection. Here, a collection is a mapping between sentences and voice recordings. Each collection has the following structure:

collection_id/
    audio/
        speaker_id/
            <wav_id_1>.wav
            <wav_id_2>.wav
            ...
    text/
        <sentence_id_1>.token
        <sentence_id_2>.token
        ...
    index.tsv
    info.json
    meta.json

  • index.tsv: A direct mapping from sentences to recordings where each line is e.g.: <speaker_id>\t<wav_id>.wav\t<sentence_id>.token

  • info.json: This file contains detailed information about each recording in the collection. An example object is shown below

    "6258": { // The recording id
        "collection_info": {
            "user_name": "n/a",
            "user_id": 27,
            "session_id": 132
        },
        "text_info": {
            "id": 7182,
            "fname": "new_000007182.token",
            "score": 1.0, // Dictates in what order the sentences are read, higher score means read earlier in collection
            "text": "Samtals urðu greinar hans átján, allar markaðar fágætu innsæi og víðtækri þekkingu.",
            "pron": "s a m t a l s\tʏ r ð ʏ\tk r eiː n a r\th a n s\tauː t j au n\ta t l a r\tm a r̥ k a ð a r\tf auː c ai t ʏ\tɪ n s ai j ɪ\tɔː ɣ\tv i ð t ai k r ɪ\tθ ɛ h c i ŋ k ʏ" // the pronounciation is split on words with a \t delimiter
        },
        "recording_info": {
            "recording_fname": "2019-12-10T113456.682Z_r000006258.wav",
            "sr": 44100,
            "num_channels": 1,
            "bit_depth": 16,
            "duration": 8.5449433106576
        },
        "other": {
            "transcription": "",
            "recording_marked_bad": false, //A user has marked this recording as bad in some way
            "text_marked_bad": false //A user has marked this sentence as bad in some way
        }
    },

  • meta.json: Contains information about this collection and the speakers involved. An example file is shown below.

    {
    "speakers": [
        {
            "id": 27,
            "name": "n/a",
            "email": "n/a",
            "sex": female,
            "age": 50,
            "dialect": "n/a"
        }
    ],
    "collection": {
        "id": 7,
        "name": "n/a",
        "assigned_user_id": 27
    }
}