MagicLuke/Redmond-Sentence-Recall

RSR segmented-latest

This dataset contains segmented child speech utterances built from the RSR raw-latest CHAT/audio tree with talkbank-toolkit.

Initial upload target: MagicLuke/Redmond-Sentence-Recall (private).

Configs

Config	Split	Rows	Source manifest
sentence	train	14563	train.sentence.jsonl
sentence	test	2065	test.sentence.jsonl
all_chi	train	15102	train.all_chi.jsonl
all_chi	test	2167	test.all_chi.jsonl

Config Meaning

• sentence is the default ASR config. It contains child utterances with parsed elicited sentence/item labels from CHAT markers such as [+ 06].
• all_chi contains every timed *CHI: utterance, including unlabeled extra child utterances. The unlabeled subset is derivable by filtering sentence_id is None.
• Rows whose segment audio was missing from the cut output are omitted from this upload.

Fields

• audio: embedded segment audio.
• audio_path: original segment-relative path under segmented-latest.
• text: audible-style transcript label.
• sentence_id: repeated prompt/item id, not a globally unique utterance id.
• segment_id: unique segment stem.
• source_cha and source_audio: paths relative to the raw RSR root.
• speaker_metadata_json: JSON-encoded CHAT speaker metadata.

Notes

The source segment cutter preserves the source sample rate and channel count. The current RSR source audio is stereo, so generated clips are stereo.

The raw full recordings and .cha transcripts are not included in this v1 upload; only segmented configs are published.

Review redistribution rights before making this dataset public.