KeisukeMiyamoto/whisper-cluster-audio

Whisper Cluster Audio

Dataset description

Whisper Cluster Audio is a Japanese audio dataset containing 1,161,120 speech clips, approximately 5,383.42 hours of audio, automatic transcripts, speech intervals, and frame-level discrete acoustic labels. It supports speech recognition training and research on representations of Japanese speech.

Audio is stored as embedded, mono, 16 kHz, 16-bit FLAC. Clip durations range from 3.00 to 29.90 seconds, with a mean of 16.69 seconds. The dataset is distributed in Parquet format with one configuration, default.

Sources and coverage

The source collection includes recordings hosted on Internet Archive and podcast, radio, and other audio distribution services. It contains diverse material, including conversational programs, broadcast recordings, and speech within entertainment recordings. Not every recording in the source collection contributes an accepted clip.

The published examples contain speech automatically identified as Japanese. This does not guarantee exclusively Japanese content: code-switching, foreign words, music, and background sounds may remain. Speaker identities, speaker counts, demographic attributes, recording dates, and a representative population distribution are not documented. Per-record source URLs, program titles, and descriptions are not included.

Splits and size

Split preparation for this release is not yet complete. The clip counts below are planned; split durations, shard counts, and the final download size remain to be confirmed before publication.

Split	Planned clips	Duration (seconds)	Duration (hours)
train	1,137,898	Pending	Pending
validation	11,611	Pending	Pending
test	11,611	Pending	Pending
Total	1,161,120	19,380,328.15	5,383.42

The total duration is the sum of the per-clip duration values, which are rounded to two decimal places.

Splits are randomized at the clip level, with approximate proportions of 98% training, 1% validation, and 1% test. They are not separated by speaker, program, or source recording. Related clips can therefore occur in different splits.

Features

Feature	Stored type	Description
id	string	Unique clip identifier.
audio	Audio; struct of binary bytes and string path	Embedded FLAC audio. The Hugging Face schema identifies this field as an audio feature.
text	string	Automatic transcript, intended as the transcription target; identical to raw_text in this release.
raw_text	string	Original automatic transcript.
speech_segments	list of structs with float64 start and end	Detected speech intervals in seconds relative to the beginning of the clip. These are speech boundaries, rather than word or speaker alignments.
has_speech	bool	Whether detected speech intervals are present.
duration	float64	Clip duration in seconds, rounded to two decimal places.
prev_id	nullable string	Identifier of the previous retained clip from the same source recording; null when no previous retained clip exists.
prev_text	nullable string	Transcript of the clip referenced by prev_id.
cluster_ids	list of uint16	Ordered acoustic cluster labels corresponding to successive encoder frames.

Previous-clip links describe retained clips, so they do not imply that clips are contiguous in the original recording. Physical row order is randomized and does not represent recording order. Previous-clip links may point to clips in another split. In particular, prev_text can expose text from another split; users should account for this when designing evaluation inputs.

Annotation properties

raw_text contains transcripts generated by OpenAI Whisper large-v3-turbo. text is identical to raw_text for all 1,161,120 clips in this release; no additional text correction was applied. Neither transcript field is a manually verified reference, and both fields may contain omissions, incorrect names, or other recognition errors.

cluster_ids uses the 500-entry codebook from KeisukeMiyamoto/whisper-cluster. Each label identifies the nearest codebook centroid to a Whisper large-v3-turbo encoder representation. Labels range from 0 to 499. Frames have a nominal spacing of 20 milliseconds, and sequences cover the clip including non-speech intervals. Encoder padding beyond the clip is excluded.

These cluster IDs are acoustic representation labels. They are not phoneme labels, text tokens, or speaker identifiers, and their numeric order has no semantic meaning.

Quality and limitations

The published clips meet automatic quality criteria for duration, detected speech, language, and transcript plausibility. Duplicate detection was also applied. The release retains 1,161,120 of 1,970,557 candidate clips. These checks do not establish human-level transcript accuracy, complete duplicate removal, or speaker-independent evaluation.

Results measured against text or raw_text measure agreement with automatic labels. They should not be treated as accuracy against independently transcribed ground truth. Source diversity and automatic selection can introduce content, recording-condition, and language-detection biases.