light1726/librispeech-realtime-qa

LibriSpeech Realtime QA

A large-scale benchmark for real-time (streaming) audio comprehension built on LibriSpeech. A model receives an utterance as a live stream and must react the moment a target event occurs: repeat a keyword the instant it is spoken, call out a semantic category (a person's name, a number, a disease term, ...) as soon as one appears, or maintain a running count of occurrences.

The benchmark contains 806,070 positive samples (the event occurs; gold timestamps are provided) and 806,070 matched negative samples (a plausible query that provably never occurs), across all seven standard LibriSpeech splits. Two evaluation suites are provided with identical samples:

• streaming — the model answers while audio plays; latency from event offset to response can be measured.
• offline — the model hears the full utterance, then answers in a fixed JSON format (an answer field with the canonical gold response is included for supervised fine-tuning or exact-match scoring).

Quick start

The audio files are not included. Download LibriSpeech (e.g. from OpenSLR or the openslr/librispeech_asr mirror) and extract the archives. audio_path is relative to the directory that contains the extracted LibriSpeech/ folder — the default when running tar -xzf dev-clean.tar.gz in some directory.

import os
from datasets import load_dataset

librispeech_root = "/data/extract"  # contains LibriSpeech/dev-clean/...
# HF split names use underscores: dev_clean, dev_clean_neg, test_other, ...
ds = load_dataset("light1726/librispeech-realtime-qa", "streaming", split="dev_clean")

for ex in ds:
    audio_file = os.path.join(librispeech_root, ex["audio_path"])
    # ex: query "vigour", labels [[0.64, 1.07, "vigour"]], prompt, ...

Task suite

task_type	prompt asks the model to	labels
lexical_detection	say a target word the moment it is spoken	one [start, end, word] per occurrence
lexical_counting	keep a live count of occurrences of a word	[start, end, "one"], [start, end, "two"], ... in order
semantic_detection	call out any word of a semantic category	one [start, end, text] per occurrence (adjacent same-category words are merged into one span, e.g. "mister morton")
semantic_counting	keep a live count of category occurrences	running count words as above

Prompts are natural-language instructions, randomly chosen from 10 paraphrases per task mode (e.g. "Streaming audio incoming. Whenever "vigour" appears, repeat it right away."). Counting samples always contain at least two occurrences spaced at least 1 s apart; count words follow "one" … "seventeen" (digit strings beyond that).

Record schema

field	type	description
utterance_id	str	LibriSpeech utterance id (speaker-chapter-utterance)
split	str	LibriSpeech split name
audio_path	str	path relative to the LibriSpeech extraction directory (LibriSpeech/<split>/...flac)
task_type	str	one of the four task types above
query	str	target word (lexical) or semantic category name (semantic)
prompt	str	instruction shown to the model
labels	list	gold events; empty list for every negative sample
response_mode	str	"offline" — offline suite only
answer	str	canonical gold JSON answer (offline suite only), e.g. [{"word":"vigour","start":0.64,"end":1.07}] (times rounded to 2 decimals; [] for negatives)

Positive samples live in {split}.jsonl, their matched negatives in {split}.neg.jsonl (same utterances, same order). The negative files for the test-clean and test-other splits are withheld from the public release (their positives are included); running the test splits therefore requires suppressing false positives without seeing paired negatives in advance. Negative queries are chosen to be plausible but provably absent: for lexical tasks the query matches no word under exact case-insensitive match, nor under inflection/possessive variants (e.g. a query is rejected if the utterance contains its plural or possessive form); for semantic tasks the category never occurs in the gold annotation. Negative labels are always [] — the correct streaming behavior is to stay silent and the correct offline answer is [].

Label provenance and quality

• Word timings come from Montreal Forced Aligner alignments of LibriSpeech (gilkeyio/librispeech-alignments). Expect tens-of-milliseconds alignment error; evaluation with a small timestamp tolerance (e.g. ±0.25 s) is recommended.
• Semantic categories were assigned per word by an LLM-based classifier (DeepSeek) over the aligned transcripts, then post-processed. Manual spot-checks show the large majority of labels are correct; occasional borderline annotations exist (e.g. "evening" in "evening papers" tagged clock_time). No manual correction was applied.
• Utterances lacking finite alignments or containing MFA special tokens were excluded entirely.
• Every positive was machine-verified: the query occurs in the gold annotation, and each label's timestamps and text exactly reproduce the alignment data. Every negative was machine-verified for absence as described above.

Statistics

806,070 samples per polarity per suite (1,612,140 per suite; 3,224,280 total), covering 268,848 unique LibriSpeech utterances. The public release contains 3,201,964 of these records: the 11,158 test-split negative samples per suite are withheld.

split	unique utts	lexical det.	lexical cnt.	semantic det.	semantic cnt.	per polarity
dev-clean	2,524	2,511	243	2,437	309	5,500
dev-other	2,634	2,623	239	2,300	260	5,422
test-clean	2,415	2,399	237	2,392	277	5,305
test-other	2,672	2,658	270	2,636	289	5,853
train-clean-100	26,601	26,529	7,639	40,260	8,014	82,442
train-clean-360	92,836	92,593	28,804	137,485	27,584	286,466
train-other-500	139,166	138,800	38,282	199,111	38,889	415,082
total	268,848	268,113	75,714	386,621	75,622	806,070

Semantic categories (19): person_name, body_part, cardinal, profession, kinship, animal, ordinal, city, color, food, plant, clock_time, country, organization, weather, money, date, disease, state_or_province.

Intended use and limitations

Intended for training and evaluating streaming speech models on latency-critical detection and counting. Limitations to keep in mind:

• Labels inherit forced-alignment error; timestamps are not human-verified.
• Semantic labels are LLM-generated and contain annotation noise; they are consistency-checked, not human-corrected.
• Queries are drawn from a single English read-speech corpus; domain coverage is limited accordingly.
• Lexical matching is case-insensitive on the aligned (normalized) transcript; homophones of the query that are transcribed differently count as absent.

License and attribution

The LibriSpeech audio is distributed by its authors under CC BY 4.0. These annotations are released under CC BY 4.0 as a derivative work. If you use this dataset, please cite:

@misc{zhan2026librispeechrealqa,
  title = {LibriSpeech-Real-QA: A Benchmark for Streaming Question Answering},
  author = {Zewen Zhan and Hui Lu and Xueyuan Chen and Huimeng Wang and Yikai Huang and Xixin Wu and Shiyin Kang and Bin Xia and Zhiyong Wu},
  year = {2026},
  eprint = {2611.06496},
  archivePrefix = {arXiv},
  primaryClass = {cs.CL},
  url = {https://arxiv.org/abs/2611.06496},
  doi = {10.48550/arXiv.2611.06496},
}