HiFiTTS-2: Large-Scale High Bandwidth Speech Metadata from LibriVox

HiFiTTS-2 is a large-scale speech dataset from NVIDIA, containing metadata for approximately 36.7 thousand hours of audio derived from LibriVox audiobooks. The metadata includes estimated bandwidth and corresponds to audio from 5 thousand speakers, recorded at a 48 kHz sampling rate. The dataset was last updated on the platform in November 2025.

Use Cases

• Training high-fidelity text-to-speech models based on the large-scale, high-bandwidth audio data.
• Benchmarking speech synthesis quality based on estimated bandwidth metadata.
• Studying speaker diversity and characteristics based on data from 5,000 speakers.
• Preprocessing and filtering audio datasets based on bandwidth and source metadata.

Strengths

• Large scale with metadata for approximately 36.7 thousand hours of audio.
• High speaker diversity with data from 5 thousand speakers.
• High-quality audio source with a 48 kHz sampling rate.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• The dataset contains metadata only; the actual audio files must be downloaded separately from LibriVox.

Provenance

Source

NVIDIA, derived from LibriVox audiobooks.

Collection Method

Derived from publicly available LibriVox audiobooks.

Time Range

null

Freshness

Last updated 2025-11-18 23:42:07; freshness should be verified.

Geography

null