Japanese Speech Dataset With 380 Speakers And 1.2 Million Samples

Filtered GOL Dataset is a Japanese text-to-speech resource containing approximately 1.2 million audio samples totaling 1,880 hours from 380 speakers. It was filtered by tts-dataset for TTS training, applying rules on text length, audio duration, and speaker minimums. The audio is in FLAC format at 44.1kHz and is packaged as a WebDataset.

Use Cases

• Train a multi-speaker TTS model using the 380 distinct speaker identities and their associated FLAC audio.
• Fine-tune an ASR model on Japanese speech using the 1.2 million samples of filtered text and audio pairs.
• Analyze speaker characteristics or build a voice cloning system leveraging the dataset's requirement of at least 5 hours of audio per speaker.
• Preprocess and filter raw speech data by applying similar rules on text length (3+ characters) and audio duration (1-60 seconds).

Strengths

• Large scale with approximately 1.2 million samples and 1,880 total hours of audio.
• Includes 380 speakers, each contributing a minimum of 5 hours of data, supporting diverse voice modeling.
• Applied rigorous filtering to exclude non-linguistic text, emoticons, and repetitive characters, improving data quality for TTS.

Limitations

• Audio is limited to monaural FLAC format at 44.1kHz, which may not be suitable for applications requiring stereo or different sample rates.
• The dataset is exclusively in Japanese (Language:ja), limiting its utility for multilingual TTS research.
• Filtering conditions, such as capping audio at under 60 seconds, exclude longer-form speech samples.

Provenance

Source

Filtered version of the midralab/gol-dataset from Hugging Face, curated by tts-dataset.

Collection Method

Filtered from an original dataset using text length, audio duration, speaker minimums, and text content rules.

Freshness

Last updated on 2026-01-15.

Geography

Region is tagged as 'us', but primary language is Japanese (ja), indicating potential geographic bias or mixed provenance.