IRIS (Instruction-Aware Retrieval for Speech) is a dataset for training instruction-aware speech retrieval models. Each example contains an instruction, a spoken query, a positive document, negative documents, and metadata describing the retrieval criteria.

Files

File	Contents
`train.jsonl`	890,278 training examples.
`val.jsonl`	9,653 validation examples.
`audio/audio-*.tar`	391,468 WAV recordings, packaged in shards for download.

Extracting the audio archives creates audio/<id>.wav, matching the relative paths in both JSONL files. The WAV files are preserved without resampling or re-encoding.

Example format

Field	Description
`instruction`	The retrieval instruction.
`spoken_query`	Relative path to the query recording.
`positive_document`	Relative path to a positive recording.
`negative_documents`	List of relative paths to negative recordings.
`select_criteria`	The criteria used to select the positive and negative documents.
`spoken_query_metadata`	Metadata for the query.
`positive_document_metadata`	Metadata for the positive document.
`negative_documents_metadata`	Metadata for the negative documents.
`source`	The source dataset identifier.

Metadata fields include speaker, gender, emotion, accent, sentiment, environmental_sound, dialogue_id, topic_id, and text. Unavailable attributes are represented by null.

Download

After access is approved, download the files using an authenticated Hugging Face account:

hf download lca0503/IRIS-Data --repo-type dataset --local-dir IRISData

for archive in IRISData/audio/audio-*.tar; do
    tar -xf "$archive" -C IRISData
done

Resolve the audio paths relative to the downloaded IRISData directory.

Sources and terms

Source materials retain their applicable licenses and permission conditions. Access to this repository does not replace source-specific obligations. Refer to the official source pages below for their terms and attribution requirements.

Source identifier	Source
`crema_d`	CREMA-D
`emns`	EMNS
`emov_db`	EmoV-DB
`esd`	Emotional Speech Dataset
`l2arctic`	L2-ARCTIC
`librispeech_c`	LibriSpeech-C recipe, LibriSpeech, and MS-SNSD
`meld`	MELD
`msp_improve`	MSP-IMPROV
`multidialog`	MultiDialog
`spokenwoz`	SpokenWOZ
`synthetic`	MS MARCO triplets from sentence-transformers/msmarco-bm25, synthesized with Gemini 2.5 Pro TTS, Gemini 3.1 Flash TTS, and Qwen3 TTS; background sounds from UrbanSound8K
`urbansound8k`	UrbanSound8K
`vocalsound`	VocalSound
`voxceleb1`	VoxCeleb1

Downloads last month: 112

Total file size:

167 GB