LongCapBench-TAC-Long

LongCapBench-TAC-Long is the long-form timestamped audio captioning subset of LongCapBench, the benchmark released with LongCap. It contains four full-length recordings, two films and two variety shows in English and Chinese, totaling 7.05 hours. All 6,520 events were labeled by human annotators: every speech, music, and sound event in each recording, with start and end times and structured attributes.

Paper (arXiv, coming soon) | Project page | Code | Model, FP8 | Model, BF16

Recordings

Key	Recording	Language	Duration	Events
01_movie_en	A Star Is Born (1937)	English	110.7 min	1,892
01_movie_zh	Personal Tailor (私人订制)	Chinese	117.1 min	1,956
08_standup_variety_en	The 92nd Academy Awards (2020)	English	93.6 min	1,187
08_standup_variety_zh	Roast! (吐槽大会) S3E05	Chinese	101.5 min	1,485
Total			7.05 h	6,520

The audio and video are not included; sources.jsonl links to each source.

Annotation

Every event was labeled by human annotators. Working from draft events produced by Qwen3.5-Omni-Plus, annotators first corrected each event's start and end times, then reviewed each of its attributes. Batches went to specialist groups, for example for music, that each handled a small set of related attributes. The final labels depart substantially from the drafts: about 53% of matched event boundaries were moved, and 91% of attribute entries were added, removed, or modified.

Attribute names and values follow LongCap's expert-designed taxonomy and are written in Chinese; transcripts keep the spoken language. attribute_keys.json gives the English name of every key. Intervals that annotators left unlabeled are listed in annotation_notes.json.

Example. Three consecutive reference events 58.5 minutes into the 92nd Academy Awards: applause, a metallic screech, and a short exclamation.

Show the events

[
  {
    "start": 3510.0,
    "end": 3525.5,
    "attributes": {
      "type": "sound_event",
      "声音事件识别": "掌声",
      "声源真实性": "真实",
      "响度轮廓": "高响度且持续",
      "音色与质感": "密集、嘈杂、具有空间感",
      "混响特性": "明显的厅堂混响",
      "前景-背景音区分": "前景主导声音",
      "声音功能": "表达赞赏和庆祝"
    }
  },
  {
    "start": 3513.5,
    "end": 3525.5,
    "attributes": {
      "type": "sound_event",
      "声音事件识别": "金属摩擦/刹车声",
      "声源真实性": "真实录音",
      "响度轮廓": "持续高响度",
      "音色与质感": "尖锐、粗糙、高频、金属感",
      "混响特性": "微弱环境反射",
      "前景-背景音区分": "单一主导声源",
      "声音功能": "警示/噪音"
    }
  },
  {
    "start": 3525.589,
    "end": 3525.912,
    "attributes": {
      "type": "speech",
      "说话人年龄": "青年",
      "说话人性别": "男",
      "文字转写": "Wow.",
      "语言识别": "英文/其他",
      "响度轮廓": "突发",
      "语速": "快",
      "音高语调": "上扬",
      "说话人意图": "抒发获奖的惊讶情绪"
    }
  }
]

Files

File	Contents
references.jsonl	Human-annotated events, one line per recording
predictions/longcap-35b-a3b-fp8.jsonl	Output of LongCap-35B-A3B-FP8 on the same recordings (6,505 events, model-generated)
sources.jsonl	Source links and checksums of the evaluated audio
SCHEMA.md	Record format and field descriptions
validate.py, verify_audio.py	Checks for the files and for audio you prepare locally
Other .json and .jsonl files	Annotation notes, event provenance, prediction status, and hashes

Each line is one recording: key, duration in seconds, and events, a list of objects with start, end (seconds from the start of the recording), and attributes. Events may overlap.

The predictions are the model's single pass over each full recording, kept with their errors; three of the four are partial. Scored against the references with the paper's metrics, they reach 44.84% TER, 11.46% GER, and 50.85% tcpCapSim.

Load the data

import json
from huggingface_hub import hf_hub_download

def load(filename):
    path = hf_hub_download("JinchuanTian/LongCapBench-TAC-Long", filename, repo_type="dataset")
    with open(path, encoding="utf-8") as f:
        return [json.loads(line) for line in f if line.strip()]

references = load("references.jsonl")
predictions = load("predictions/longcap-35b-a3b-fp8.jsonl")

Read the files as JSON lines, as above. Attribute values mix strings, lists, and objects, so datasets.load_dataset with automatic schema inference can alter them; the dataset viewer is off for the same reason.

Source audio

Download each recording from its source_url in sources.jsonl, under the provider's terms, and convert it to 16 kHz mono WAV:

ffmpeg -i source-video.mp4 -vn -ac 1 -ar 16000 -c:a pcm_s16le -fflags +bitexact -map_metadata -1 01_movie_en.wav
python verify_audio.py 01_movie_en 01_movie_en.wav

verify_audio.py checks that your file matches the evaluated audio exactly. A match is not guaranteed, because a source may have been re-encoded or edited since we prepared it.

Disclaimer

This dataset is intended for research use. The references are human annotations and can still contain errors or omissions, and the included predictions are AI-generated and can be wrong. You are responsible for how you use the data, including obtaining the source media lawfully. The annotations, predictions, and scripts are provided "as is", without warranty of any kind, under the Apache License 2.0. The source audio and video are not part of this dataset and remain under their owners' terms.

Citation

Please cite the LongCap paper; the BibTeX entry is in the code repository.