README.md

dutch-vl-tts

This dataset contains 15.000 audio fragments of a male Dutch Flemish voice, the sentences read are extracted from the Mozilla Common Voice project.

Dataset: Google Drive (1.5GB)

New: the dataset is now also available for preview on Kaggle

To use this dataset with Mozilla TTS, append the following fragment to TTS/tts/datasets/preprocess.py:

def rdh_flemish(root_path, meta_file):
    txt_file = os.path.join(root_path, meta_file)
    speaker_name = "rdh_flemish"
    items = []
    with open(txt_file, 'r', encoding="utf-8") as f:
        for line in f:
            cols = line.split("|")
            text = cols[1]
            wav_file = os.path.join(root_path, cols[0] + ".wav")
            items.append([text, wav_file, speaker_name])
    return items

Audio

Files in the dataset are 16-bit, 22050Hz downsampled from 44.1kHz, mono, wave.

The audio samples unfortunately may vary slightly over recording sessions.

• De plant van de aardappel is giftig
• De Straat van Gibraltar is een zee-engte aan het uiteinde van de Middellandse Zee.
• Dat kan met betrekkelijk weinig geld.
• Ik geloof in deze dialoog en we zullen deze dialoog voeren.
• We zullen de uitkomsten op deze werkterreinen afwachten.

Trained models

Models with their corresponding synthesised audio samples are provided in the links below.

Original dataset:

• Tacotron DDC: Google Drive

• Glow: Google Drive

Other dataset:

• Tacotron DDC with transfer learning: Google Drive (subpar results)

Due to a severe lack of quality data (4.000 noise gated fragments) the second dataset hasn't been released. The first model was used for transfer learning, although this still proved to be insufficient.