LibriSpeech ASR Adversarial Audio Corpus

Packed with approximately 1000 hours of read English speech audio, prepared by Vassil Panayotov with assistance from Daniel Povey. It is derived from LibriVox audiobooks, segmented and aligned, with a 16 kHz sampling rate.

Use Cases

• Train automatic speech recognition models on 1000 hours of read English speech audio.
• Benchmark adversarial robustness of ASR systems using carefully segmented and aligned audio.
• Analyze speech patterns and alignment from LibriVox audiobook recordings.

Strengths

• Approximately 1000 hours of audio data provides substantial material for model training.
• Audio is carefully segmented and aligned, ensuring structured data for ASR tasks.
• Data is derived from the LibriVox project, a known source of public domain audiobooks.

Limitations

• Audio is stored in FLAC format, requiring conversion to float32 arrays for typical ML use, adding a preprocessing step.
• The dataset consists solely of read speech from audiobooks, which may not represent spontaneous conversational speech patterns.
• Specific details on adversarial perturbations or modifications are not described in the provided input.

Provenance

Source

LibriVox project audiobooks.

Collection Method

Derived from read audiobooks, carefully segmented and aligned.

Freshness

The dataset was last updated on 2025-04-03.

Geography

Region tag suggests US focus, but specific coverage is unknown.