ASR Language Model Training Corpus Version 2.0

Asr Book Lm V2.0 is a text corpus for training language models in automatic speech recognition systems. The dataset was created by author Jiejie and was last updated on March 14, 2022. Its size is categorized as 1K<n<10K, indicating it contains between 1,000 and 10,000 entries.

Use Cases

• Train a language model to predict the next word in transcribed speech sequences.
• Fine-tune an acoustic model by aligning phoneme sequences with text transcriptions from the corpus.
• Generate synthetic speech training data by using the text corpus as input for a text-to-speech system.
• Evaluate word error rates by comparing ASR system outputs against reference text from this dataset.

Strengths

• Dataset size is confirmed to be between 1,000 and 10,000 entries (1K<n<10K).
• Data is stored in Parquet format, which is efficient for columnar storage and fast loading.

Limitations

• The exact number of rows, columns, and total data size is unknown.
• The source, collection method, and license for the data are not specified.
• The data may be geographically biased, as a platform tag suggests a US region focus.

Provenance

Source

Jiejie (author on Hugging Face)

Collection Method

Inferred to be a curated text corpus, but exact gathering method is null.

Time Range

null

Freshness

Last updated in March 2022, which may be stale for rapidly evolving ASR tasks.

Geography

Platform tags suggest a potential US regional focus, but exact coverage is null.