LibriSpeech Audio Samples for Speech Recognition Testing

LibriSpeech ASR Dummy is a small-scale dataset from Hugging Face's internal testing, containing audio-text pairs for English speech recognition. It was created by hf-internal-testing and last updated in June 2024. The dataset is categorized as 'n1 K', indicating it contains approximately 1,000 samples.

Use Cases

• Train a speech-to-text model to predict text transcripts from audio waveforms.
• Benchmark ASR model performance on clean, read English speech audio samples.
• Validate audio preprocessing pipelines using standardized audio-text pairs.
• Fine-tune a pre-trained acoustic model on a small, curated English audio corpus.

Strengths

• Approximately 1,000 samples, providing a manageable size for quick testing.
• Contains both audio and text modalities, enabling supervised learning for speech recognition.
• Data is in Parquet format, facilitating efficient loading with libraries like pandas and polars.

Limitations

• Small sample size (~1k rows) limits its use for training robust production models.
• Scope is restricted to English language and US region, lacking linguistic and geographic diversity.
• As a 'dummy' set, it may lack the complexity and noise of real-world conversational audio.

Provenance

Source

Hugging Face (hf-internal-testing)

Collection Method

Derived from or inspired by the LibriSpeech corpus for internal platform testing.

Time Range

null

Freshness

Last updated on 2024-06 19.

Geography

Primarily US region.