KeisukeMiyamoto/nhk-archive-audio

NHK Archives Audio

NHK Archives Audio is a Japanese audio corpus built from records in the public NHK Archives search service. It contains the audio tracks of archive video and audio records together with titles, descriptions, genres, broadcast metadata, regions, source pages, direct stream URLs, and duration metadata.

The source streams were converted to 16 kHz mono FLAC and embedded directly in Parquet files for use with the Hugging Face Dataset Viewer. This dataset does not contain transcripts.

This is an independently constructed research dataset and is not an official NHK dataset.

Purpose

This dataset is intended for Japanese speech and audio research, including speech recognition experiments, voice activity detection, speaker diarization, acoustic event analysis, archive search, and broadcast-media analysis.

Source Data

The records and metadata come from the public NHK Archives search service. Source metadata was read from KeisukeMiyamoto/nhk-archive-meta.

Dataset Size

Items	Duration	Parquet shards	Download size
32,414	2,185.65 hours	444	241.00 GB

Duration distribution

Statistic	Seconds
Minimum	0.256
Median	103.654
Mean	242.745
Maximum	10,840.004

All 32,414 record IDs and direct stream URLs are unique.

Audio Format

Property	Value
Format	FLAC
Sampling rate	16,000 Hz
Channels	Mono

Columns

Column	Type	Description
id	string	NHK Archives record ID.
title	string	Record title from NHK Archives.
audio	Audio	Embedded 16 kHz mono FLAC audio. The stored filename is <id>.flac.
duration	float64	Duration of the stored FLAC audio in seconds.
description	string	Record description. May be missing.
program_name	string	Program or work title associated with the record. May be missing.
subtitle	string	Subtitle associated with the program or work. May be missing.
genres	list[string]	Leaf genre labels from the NHK Archives genre hierarchy. May be missing.
broadcast	string	Broadcast year or year range as supplied by NHK Archives. May be missing and is not always a single year.
region	list[string]	Prefecture or other regional labels. May be missing.
page_url	string	Corresponding NHK Archives record page.
direct_url	string	Source HLS playlist URL used during collection. It may later become unavailable.

Missing metadata

Column	Missing rows	Missing rate
description	7,955	24.54%
program_name	9,816	30.28%
subtitle	20,352	62.79%
genres	4,089	12.61%
broadcast	7,872	24.29%
region	15,266	47.10%

Missing values reflect the source metadata and were not inferred or filled.

Text Processing

All textual metadata is recursively normalized with Unicode NFKC before storage. This standardizes compatibility characters such as full-width Latin letters, digits, and punctuation.

Usage

from datasets import load_dataset

dataset = load_dataset(
    "KeisukeMiyamoto/nhk-archive-audio",
    split="train",
)

Because the complete dataset is large, streaming is recommended when only sequential access is required:

from datasets import load_dataset

dataset = load_dataset(
    "KeisukeMiyamoto/nhk-archive-audio",
    split="train",
    streaming=True,
)

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

License

Please read the official NHK Creative Library usage rules.