KeisukeMiyamoto/lambda-audio-courpus

Lambda Audio Corpus

This is a combined Japanese automatic speech recognition corpus containing 2,292,233 clips (8,428.40 hours). It combines broadcast, archive, radio news, podcast, and ReazonSpeech data in a common schema. Audio is embedded as 16 kHz mono FLAC, and stored durations range from 0.46 to 30 seconds.

Sources

origin	Included clips	Duration	Source
nhk-archive-audio-30s	137,106	1,067.26 hours	NHK Archives Audio 30s
nhk-radio-news-audio-30s	4,462	30.55 hours	NHK Radio News Audio 30s
reazon-speech-v2	1,023,030	1,821.55 hours	ReazonSpeech v2
tech-podcast-audio-30s	1,127,635	5,509.03 hours	Tech Podcast Audio 30s

The table describes the records included here and does not imply that every upstream dataset is included in full.

Splits

Split	Clips	Duration	Ratio by rows
train	2,259,220	8,331.22 hours	98.56%
test	16,648	48.80 hours	0.73%
validation	16,365	48.37 hours	0.71%

The source data was assigned to splits by origin. After Japanese-language filtering, mixing, exact-text deduplication, and ReazonSpeech subsampling, the split ratios are shown above. Half of the ReazonSpeech rows in each split were randomly removed with seed 42. Rows were locally shuffled with seed 42 before deduplication. When text appeared more than once, one row was kept in test, then validation, then train priority order. The splits are not speaker-, program-, episode-, or source-disjoint.

Columns

Column	Description
id	Unique segment ID.
audio	Embedded 16 kHz mono FLAC audio.
text	Transcript used as the ASR target. Its provenance and correction process depend on origin.
raw_text	Transcript before correction, when supplied by the source dataset. It is null for reazon-speech-v2.
speech_segments	Speech intervals in seconds relative to the stored clip.
has_speech	Whether speech is present; true for all included records.
duration	Stored clip duration in seconds.
prev_id	Previous related segment ID, when available.
prev_text	Transcript of the previous related segment, when available.
origin	Source dataset identifier shown in the Sources table.

For the NHK and podcast sources, raw_text was generated with Whisper large-v3-turbo and text contains LLM-assisted corrections. For reazon-speech-v2, text is derived from the supplied ReazonSpeech transcription and raw_text is unavailable. A record referenced by prev_id may be in another split or may have been removed during filtering or deduplication; prev_text can therefore cause contextual leakage.

Usage

from datasets import load_dataset

dataset = load_dataset(
    "KeisukeMiyamoto/lambda-audio-courpus",
    split="train",
    streaming=True,
)

record = next(iter(dataset))
print(record["id"], record["origin"], record["duration"], record["text"])

Streaming is recommended because the Parquet files occupy approximately 575.0 GiB.

Source Terms and License

This combined dataset uses license: other because rights and conditions differ by source. Copyright in the original recordings remains with the respective creators and rights holders.

• NHK-derived data remains subject to the applicable NHK terms, including the NHK Creative Library usage rules.
• ReazonSpeech is distributed under CDLA-Sharing-1.0 and its dataset card additionally restricts use to the purpose permitted by Article 30-4 of the Japanese Copyright Act. Review and accept the ReazonSpeech terms before use.
• Podcast recordings remain the property of their creators and publishers; users must verify that their intended use is permitted.

Users are responsible for reviewing every upstream dataset card and ensuring compliance with all applicable licenses, terms, copyright restrictions, privacy requirements, and laws.