AudioTokenBench: 3,150 Audio Samples for Tokenizer Evaluation

3150 audio samples at 24kHz, created by bosonai and last updated on 2025-07-28. The dataset is designed for evaluating the HiggsTokenizer and contains four subsets: Speech, Music, Sound Event, and Audiophile. The Speech, Music, and Sound Event subsets each contain 1,000 ten-second clips, while the Audiophile subset contains 150 thirty-second high-fidelity clips.

Use Cases

• Benchmark audio tokenizer reconstruction quality based on the 24kHz audio samples.
• Evaluate model performance across different audio domains based on the Speech, Music, Sound Event, and Audiophile subsets.
• Test audio generation fidelity on high-quality samples based on the curated Audiophile clips.

Strengths

• Contains 3,150 total audio samples, providing a substantial evaluation corpus.
• Includes four distinct subsets (Speech, Music, Sound Event, Audiophile) for domain-specific testing.
• Audiophile subset features 150 thirty-second clips curated from high-fidelity test discs.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for individual subsets beyond the totals is unknown, which may limit suitability assessment.

Provenance

Source

Samples sourced from DAPS, MUSDB, AudioSet, and high-fidelity test discs.

Collection Method

Clips were randomly sampled from source datasets or curated from test discs.

Time Range

null

Freshness

Last updated 2025-07-28 22:03:10; freshness should be verified.

Geography

null