LibriSpeech 10-Hour Audio Transcription Corpus

Librispeech10H is a 10-hour subset of the LibriSpeech corpus, containing English speech audio and corresponding transcriptions. It was created by user ahazeemi and published on Hugging Face in April 2022. The dataset is structured for machine learning tasks, as indicated by its platform tags.

Use Cases

• Train an automatic speech recognition model to predict text transcriptions from audio waveforms.
• Benchmark the performance of speech-to-text models on a standardized 10-hour English audio corpus.
• Fine-tune pre-trained audio models using aligned audio and text pairs for speaker adaptation or domain-specific tasks.

Strengths

• Contains 10 hours of audio data, providing a substantial training volume for speech models.
• Includes aligned text transcriptions, enabling supervised learning for speech recognition.

Limitations

• Limited to 10 hours of audio, which is small compared to full-scale ASR training corpora.
• Potential bias towards the specific books and speakers represented in the original LibriSpeech source.

Provenance

Source

Subset of the LibriSpeech corpus.

Collection Method

Curated and hosted on Hugging Face by ahazeemi.

Freshness

Last updated in April 2022.

Geography

Region tag suggests US English speech.