google/mind-the-gap-inline-ad

Mind The Gap (MTG-AD): Dataset and Fine-grained Evaluation for End-to-End Inline Audio Descriptions

Mind The Gap (MTG-AD) is a professionally annotated dataset of 433 YouTube user-generated videos (~36 hours) paired with inline audio descriptions (AD) co-created by paired blind and sighted AD professionals. Inline AD must fit within the natural silences of a video, so the dataset benchmarks the full end-to-end inline AD task: generating coherent timed narration scripts that describe the important visual content without overlapping the original speech.

This repository accompanies the NeurIPS 2026 (Evaluations & Datasets Track) paper Mind The Gap: Dataset and Fine-grained Evaluation for End-to-End Inline Audio Descriptions. The evaluation code (MTG-AD-Eval), AD generation prompts, and TTS-timing toolkit are at github.com/google/mtg-ad.

Paper · Dataset · Code

Dataset Summary

Property	Value
Videos	433 YouTube user-generated videos (332 validation / 101 test)
Total duration	~35.9 hours
Average video length	~5 minutes
Ground-truth AD segments	8,126 total (6,214 validation / 1,912 test; median 16 per video)
Categories	Vlog (201), Education (71), Cooking (49), Nature (48), Travel (36), Science (23), Fashion (4), Other (1)
Language	English
License	CC BY 4.0

Configurations and Splits

The repository has two configurations (subsets). They have different schemas and granularity, so they are kept as separate configs of the same dataset.

Config	Split	Rows	Granularity	Contents
annotations (default)	validation	332	one row per video	ground-truth AD, dense visual info, ASR transcript, metadata
annotations (default)	test	101	one row per video	ground-truth AD, dense visual info, ASR transcript, metadata
model_outputs	test	1,193	one row per (video, model configuration)	ADs generated by 12 model/prompt configurations + per-video metric scores

from datasets import load_dataset

ann = load_dataset("google/mind-the-gap-inline-ad", "annotations")       # validation + test
val, test = ann["validation"], ann["test"]

preds = load_dataset("google/mind-the-gap-inline-ad", "model_outputs", split="test")
gemini = preds.filter(lambda r: r["config_name"] == "gemini_2p5_flash_timing_asr")

The test split is the 101-video evaluation subset used for all multi-model results in the paper. The validation split contains the remaining 332 videos and can be used for prompt development, few-shot example selection, or fine-tuning/validation. There is no training split.

Coverage of auxiliary files

	validation (332)	test (101)
Ground-truth AD	332	101
Dense visual info	330 (305 Gemini 2.5 Flash, 25 Gemini 3.7 Flash)	101 (Gemini 2.5 Flash)
ASR transcript	332 (94 empty)	101 (20 empty)
Model outputs + metrics	–	101 per configuration (82 for qwen3p5_omni_flash), 12 configurations

An empty transcript ("", empty file) means the video has no transcribed speech. Where a visual-info file is unavailable, the visual_info field is null and there is no file under raw/.

Data Fields

annotations

Field	Type	Description
video_id	string	11-character YouTube video ID
youtube_url	string	https://www.youtube.com/watch?v={video_id}
split	string	validation or test
category	string	Content category of the video
duration	string	Video duration as M:SS
duration_seconds	int	Video duration in seconds
num_ad_segments	int	Number of ground-truth AD segments
gt_ad	string	Raw ground-truth AD script (one segment per line, see format below)
gt_ad_segments	list of {start_time, start_seconds, ad_type, text}	Parsed ground-truth AD segments
visual_info	string or null	Dense, timestamped visual information generated by a Gemini model (visual events, on-screen text, people, summary)
visual_info_model	string or null	Model that generated visual_info: gemini-2.5-flash or gemini-3.7-flash
transcript	string	Raw timestamped YouTube ASR transcript (empty string if the video has no transcribed speech)
transcript_segments	list of {start_time, end_time, start_seconds, end_seconds, text}	Parsed transcript segments (empty list if no speech)

model_outputs

Field	Type	Description
video_id	string	YouTube video ID (test split)
config_name	string	Model/prompt configuration identifier (see table below)
model	string	Model name
prompt	string	Prompting strategy
model_ad	string	Model response as produced (not edited), including any preamble/reasoning the model emitted before the timestamped AD script
model_ad_source	string	Where the response comes from (original for the first 9 configurations; otherwise the source directory)
metrics_source	string or null	Where the metric file comes from; null if no metric file exists for that row
include_in_scoring	bool	false for rows that are not part of the paper's evaluation and should be left out of aggregate metrics (the 25 unstandardized qwen3p5_omni_flash responses)
exclusion_reason	string or null	Why a row is excluded from scoring; null if included
metric columns	float or null	Per-video numeric metric values (see Metrics); null when a metric could not be computed for that video
metrics_json	string or null	Verbatim content of the per-video metric JSON file (newer configurations include extra fields such as the AD text, transcript, and judge reasoning)

config_name	Model	Prompt
gemini_2p5_flash_timing_asr	Gemini 2.5 Flash	Guideline + Timing ASR
gemini_2p5_flash_guideline_asr	Gemini 2.5 Flash	Guideline + ASR
gemini_2p5_flash_guideline	Gemini 2.5 Flash	Guideline only
gemini_2p5_flash_simple	Gemini 2.5 Flash	Simple AD
claude_4_5	Claude 4.5	Guideline + Timing ASR
claude_4	Claude 4	Guideline + Timing ASR
gpt_4_1_mini	GPT-4.1 Mini	Guideline + Timing ASR
gpt_5_2	GPT 5.2	Guideline + Timing ASR
qwen_vl_max	Qwen-2-VL-Max	Guideline + Timing ASR
gemini_3p5_flash_timing_asr	Gemini 3.5 Flash	Guideline + Timing ASR
qwen3p5_omni_flash	Qwen-3.5-Omni-Flash	Guideline + Timing ASR
qwen3p5_omni_plus	Qwen-3.5-Omni-Plus	Guideline + Timing ASR

Raw File Formats

The same content is also provided as plain-text files under raw/, in the layout expected by the evaluation toolkit:

raw/
├── validation/
│   ├── gt_ad/output_{VIDEO_ID}.txt
│   ├── visual_info/output_{VIDEO_ID}.txt
│   └── transcript/output_{VIDEO_ID}.txt
└── test/
    ├── gt_ad/output_{VIDEO_ID}.txt
    ├── visual_info/output_{VIDEO_ID}.txt
    ├── transcript/output_{VIDEO_ID}.txt
    ├── model_ads/{CONFIG_NAME}/output_{VIDEO_ID}.txt
    ├── model_metrics/{CONFIG_NAME}/output_{VIDEO_ID}.json
    └── scoring_exclusions.json   # {config: [video_id, ...]} rows to leave out of aggregate metrics

Video ID lists: video_ids.txt (all 433), validation_video_ids.txt (332), test_video_ids.txt (101).

Ground-truth AD (gt_ad): one segment per line; the timestamp is the segment start time, followed by the AD type and a Unicode line separator (U+2028) before the description text:

HH:MM:SS.mmm-STANDARD<U+2028>description text

Visual info (visual_info): free-form model output containing timestamped visual events (HH:MM:SS.mmm- event description), on-screen text, descriptions of people, and a video summary.

Transcript (transcript): one ASR segment per line:

HH:MM:SS.mmm--HH:MM:SS.mmm spoken text

Model metrics (model_metrics/*.json): a flat JSON object of metric values. In the JSONL model_outputs config, . in key names is replaced with _ (e.g. metric_G1.1_precision_avg_ad_in_visual_info → metric_G1_1_precision_avg_ad_in_visual_info).

Metrics

The evaluation suite (MTG-AD-Eval) measures generated AD along dimensions derived from professional AD guidelines. Main per-video metric columns and the paper metrics they correspond to:

Paper metric	Column	Direction
Prioritization (G1.1)	metric_G1_1_precision_avg_ad_in_visual_info	↑
No Redundancy (G1.2)	metric_2_1_does_not_restate	↑
On-Screen Text (G1.3)	metric_1_2_narrates_onscreen_text (meaningful when onscreen_text_present = 1)	↑
Timeliness (G2)	metric_g_2_avg_diff_time (seconds)	↓
Audio Overlaps (G7)	total_overlap / tts_ad_transcript_overlap_seconds (seconds, TTS at tts_rate = 1.2×)	↓
Avg. Silence Duration (G8)	avg_silence_duration_w_ad_tts (seconds)	↓
Audio Coverage Ratio (G8)	ad_total_speaking_time_per_silent_second	≤ 1 (higher is better up to 1; values > 1 mean AD time exceeds available silence)
VideoA11y-Ref	videoa11y_descriptive, videoa11y_objective, videoa11y_accurate, videoa11y_clear (1–5 Likert)	↑

Additional columns describe AD/GT word counts and durations (ad_total_words, gt_total_time, …), speech/silence statistics of the video (total_silence_time, dialog_density, …), and diagnostics (metric_g_2_matched_visual_info_ratio, metric_g_2_diffs_over_ten, metric_G1_1_precision_avg_ad_in_visual_info_std).

Aggregate numbers in the paper are computed from these per-video values with the aggregation in the code release (analysis/aggregate.py, analysis/paper_tables_and_figures.ipynb): rows with include_in_scoring = false are dropped, On-Screen Text only counts videos with on-screen text, timing values longer than the video are ignored, the prompt comparison uses only the videos that have outputs from every prompt, and the model comparisons score each model on all test videos it has output for. Naive column means therefore differ from the reported numbers.

Data Collection and Curation

Video selection

Videos were sourced from YouTube channels with at least 1,000 subscribers, selecting videos around 3–7 minutes long. An initial candidate pool of 557 videos was reviewed by the professional AD team. Videos were excluded if they contained non-stop narration leaving no temporal gaps, or skipped for redundancy after annotating 3–5 videos of the same theme.

Annotation process

Audio descriptions were written by pairs of professional AD writers: one blind and one sighted.

1. The sighted writer watched each video in full, then re-watched it in segments to draft concise, time-aligned descriptions of visual information not conveyed by the existing audio, including on-screen text.
2. The draft was imported into an editing environment to test and refine timing against the original audio track so that descriptions fit within natural pauses.
3. The blind writing partner performed a comprehensive quality-control review of clarity and effectiveness.

Visual information and transcripts

Dense visual information is synthetic (model-generated). It was generated with Gemini 2.5 Flash for 406 videos (including all test videos) and with Gemini 3.7 Flash, using the visual-info prompt in the evaluation code, for 25 validation videos that lacked it (see visual_info_model). Two validation videos (6efEHcwNMZw, Ay3oHsgFgdE) have no visual info because the model could not access them by YouTube URL. ASR transcripts are YouTube automatic speech recognition transcripts.

Model-generated ADs (test split)

For the 101 test videos, ADs were generated by frontier multimodal models with several prompting strategies (see the config_name table) and scored with the evaluation suite. Model responses and metric files are released as produced, without edits. Notes on the newer configurations (gemini_3p5_flash_timing_asr, qwen3p5_omni_flash, qwen3p5_omni_plus):

• qwen3p5_omni_flash covers 82 of the 101 test videos (no output is available for the other 19). For 57 videos the response is a version with timestamps standardized to H:MM:SS.mmm-STANDARD text; these are the responses evaluated in the paper. The remaining 25 are the model's original responses as-is: they have format or validity issues (e.g. time ranges without descriptions, malformed JSON, or API error messages instead of an AD), and their metrics were computed on those original responses. They are kept for reference but marked include_in_scoring = false (and listed in raw/test/scoring_exclusions.json); leave them out when aggregating.
• qwen3p5_omni_plus has no metric file for 1 video (CJv4d-XUFDQ) and an empty metric file for 1 video (68p4YCz7Ga4); metric columns are null for both.
• Metric files for these configurations come from a separate evaluation run and include extra fields (kept as-is). With the exclusions above, they reproduce the paper's model comparison tables and figures exactly (see below).

Key Results (from the paper)

Mean ± 95% CI half-width (1.96 · SEM) over videos, as in the paper's appendix tables. ↑ higher is better, ↓ lower is better; Audio Coverage is better when higher only up to 1 (values above 1 mean the AD is longer than all available silence). Bold marks the best value in each column.

Prompt strategy comparison (Gemini-2.5 Flash; paper Tab. 4)

The 372 validation + test videos for which every prompt has output. The Human AD row is the professional ground-truth AD, rendered with the same TTS (1.2×) and scored with the same metrics.

Prompt	Prioritization ↑	No Redundancy ↑	On-Screen Text ↑	Timeliness (s) ↓	Overlaps w. TTS (s) ↓	Audio Coverage (≤1)	Avg. Silence (s) ↓
Simple AD	0.75 ± 0.02	0.77 ± 0.03	0.47 ± 0.03	6.46 ± 3.03	15.07 ± 3.01	0.41 ± 0.06	17.95 ± 3.46
Guideline only	0.79 ± 0.02	0.72 ± 0.02	0.69 ± 0.03	3.27 ± 1.29	59.19 ± 7.28	1.23 ± 0.11	8.85 ± 3.13
Guideline + ASR	0.79 ± 0.02	0.71 ± 0.02	0.65 ± 0.03	5.11 ± 2.07	21.34 ± 4.64	0.83 ± 0.08	6.71 ± 2.47
Gd. + Timing ASR	0.76 ± 0.02	0.69 ± 0.03	0.63 ± 0.03	4.43 ± 1.17	12.77 ± 3.88	0.78 ± 0.05	8.97 ± 2.80
Human AD	0.73 ± 0.02	0.72 ± 0.02	0.61 ± 0.03	2.90 ± 0.48	7.51 ± 1.42	0.61 ± 0.03	4.19 ± 1.13

VideoA11y-Ref scores (1–5) for the prompt comparison

Prompt	Descriptive ↑	Objective ↑	Accurate ↑	Clear ↑
Simple AD	3.46 ± 0.13	3.55 ± 0.13	1.82 ± 0.11	2.59 ± 0.13
Guideline only	3.28 ± 0.14	3.59 ± 0.13	2.18 ± 0.12	2.85 ± 0.14
Guideline + ASR	2.81 ± 0.11	3.27 ± 0.12	1.74 ± 0.10	2.57 ± 0.12
Gd. + Timing ASR	3.03 ± 0.10	2.98 ± 0.13	1.64 ± 0.09	2.68 ± 0.11
Human AD	4.72 ± 0.07	4.93 ± 0.03	4.98 ± 0.02	4.96 ± 0.03

VideoA11y-Ref compares each AD with the human AD, so the Human AD row only serves as an upper bound.

The released model_outputs cover the test split only, so recomputing this comparison from them uses the 101 test videos and gives slightly different values.

Initial frontier models (Gd. + Timing ASR prompt; paper Tab. 5)

Each model is scored on all 101 test videos.

Model	Prioritization ↑	No Redundancy ↑	On-Screen Text ↑	Timeliness (s) ↓	Overlaps w. TTS (s) ↓	Audio Coverage (≤1)	Avg. Silence (s) ↓
Gemini-2.5 Flash	0.76 ± 0.04	0.70 ± 0.05	0.64 ± 0.06	5.00 ± 3.07	8.92 ± 5.62	0.74 ± 0.07	6.63 ± 2.16
Qwen-2-VL-Max	0.59 ± 0.05	0.71 ± 0.04	0.59 ± 0.06	6.92 ± 3.91	46.53 ± 11.16	0.94 ± 0.15	8.81 ± 2.78
GPT-4.1 mini	0.52 ± 0.05	0.70 ± 0.05	0.58 ± 0.06	65.28 ± 18.59	21.46 ± 4.77	1.48 ± 0.46	19.37 ± 5.89
Claude 4	0.51 ± 0.05	0.69 ± 0.05	0.56 ± 0.05	19.83 ± 4.39	7.45 ± 2.14	0.62 ± 0.06	18.57 ± 8.78
Claude 4.5	0.59 ± 0.05	0.70 ± 0.04	0.61 ± 0.05	20.50 ± 4.30	12.86 ± 3.31	0.70 ± 0.07	10.40 ± 5.86
GPT-5.2	0.55 ± 0.05	0.73 ± 0.04	0.63 ± 0.05	22.56 ± 4.76	11.14 ± 2.74	0.65 ± 0.07	14.85 ± 7.31

VideoA11y-Ref scores (1–5)

Model	Descriptive ↑	Objective ↑	Accurate ↑	Clear ↑
Gemini-2.5 Flash	3.10 ± 0.19	2.92 ± 0.22	1.63 ± 0.15	2.69 ± 0.20
Qwen-2-VL-Max	2.58 ± 0.22	2.77 ± 0.25	1.33 ± 0.14	2.21 ± 0.21
GPT-4.1 mini	3.28 ± 0.24	2.82 ± 0.22	1.31 ± 0.14	2.19 ± 0.23
Claude 4	2.32 ± 0.18	2.13 ± 0.20	1.16 ± 0.10	2.01 ± 0.18
Claude 4.5	2.69 ± 0.21	2.77 ± 0.24	1.26 ± 0.13	2.21 ± 0.22
GPT-5.2	2.89 ± 0.21	2.94 ± 0.26	1.35 ± 0.14	2.54 ± 0.25

Next-generation omni-modal models (Gd. + Timing ASR prompt; paper Tab. 6)

Each model is scored on all test videos it has output for: 101 for Gemini-2.5/3.5 Flash and Qwen-2-VL-Max, 99 for Qwen-3.5-Omni-Plus, and 57 for Qwen-3.5-Omni-Flash (after the scoring exclusions).

Model	Prioritization ↑	No Redundancy ↑	On-Screen Text ↑	Timeliness (s) ↓	Overlaps w. TTS (s) ↓	Audio Coverage (≤1)	Avg. Silence (s) ↓
Gemini-2.5 Flash	0.76 ± 0.04	0.70 ± 0.05	0.64 ± 0.06	5.00 ± 3.07	8.92 ± 5.62	0.74 ± 0.07	6.63 ± 2.16
Gemini-3.5 Flash†	0.93 ± 0.03	0.90 ± 0.03	0.57 ± 0.05	2.99 ± 0.91	0.26 ± 0.13	0.45 ± 0.03	4.85 ± 1.20
Qwen-2-VL-Max	0.59 ± 0.05	0.71 ± 0.04	0.59 ± 0.06	6.92 ± 3.91	46.53 ± 11.16	0.94 ± 0.15	8.81 ± 2.78
Qwen-3.5-Omni-Flash†	0.94 ± 0.02	0.88 ± 0.05	0.61 ± 0.06	4.91 ± 3.64	52.14 ± 15.89	0.94 ± 0.20	16.61 ± 11.88
Qwen-3.5-Omni-Plus†	0.79 ± 0.06	0.88 ± 0.04	0.58 ± 0.06	2.67 ± 1.04	30.45 ± 8.69	0.72 ± 0.14	9.49 ± 2.12

VideoA11y-Ref scores (1–5)

Model	Descriptive ↑	Objective ↑	Accurate ↑	Clear ↑
Gemini-2.5 Flash	3.10 ± 0.19	2.92 ± 0.22	1.63 ± 0.15	2.69 ± 0.20
Gemini-3.5 Flash†	4.02 ± 0.15	4.07 ± 0.15	3.79 ± 0.18	4.02 ± 0.17
Qwen-2-VL-Max	2.58 ± 0.22	2.77 ± 0.25	1.33 ± 0.14	2.21 ± 0.21
Qwen-3.5-Omni-Flash†	4.43 ± 0.23	4.41 ± 0.20	3.98 ± 0.25	4.27 ± 0.26
Qwen-3.5-Omni-Plus†	3.82 ± 0.27	4.03 ± 0.26	3.23 ± 0.25	3.81 ± 0.27

† LLM-judged metrics (Prioritization, No Redundancy, Timeliness, VideoA11y-Ref) for these models use Gemini-3.5 Flash as the judge (the other models use Gemini-2.5 Flash). The segment-level metrics appear robust to this change, but VideoA11y-Ref values are not directly comparable across the two groups. Programmatic metrics (On-Screen Text, Overlaps, Coverage, Avg. Silence) do not depend on the judge.

Tab. 5 and Tab. 6 (and the paper's model comparison figures) are recomputed exactly from the released per-video metrics with analysis/paper_tables_and_figures.ipynb in the code release.

How to Use

Download the original videos

The dataset contains YouTube video IDs, not video files. Videos can be downloaded with yt-dlp, subject to YouTube's Terms of Service and video availability:

pip install yt-dlp
while IFS= read -r vid_id; do
  yt-dlp "https://www.youtube.com/watch?v=$vid_id" -o "videos/%(id)s.%(ext)s"
done < test_video_ids.txt

Work with ground-truth segments

from datasets import load_dataset

test = load_dataset("google/mind-the-gap-inline-ad", "annotations", split="test")
row = test[0]
for seg in row["gt_ad_segments"]:
    print(seg["start_seconds"], seg["text"])

Limitations

• Scale: 433 videos is a modest scale; results may not generalize to all video styles.
• Language: all videos and annotations are in English.
• Domain: YouTube user-generated content; may not generalize to professional broadcast media.
• Visual information: generated by multimodal models and may reflect their biases and errors; 25 validation videos use a different model (Gemini 3.7 Flash) than the rest, and 2 validation videos have none.
• ASR transcripts: automatic and may contain recognition errors; 114 videos have empty transcripts (no transcribed speech).
• Video availability: videos are hosted on YouTube and may become unavailable over time (link rot).
• Metric coverage: automated metrics evaluate a prioritized subset of guidelines (G1.1–G1.3, G2, G7, G8); aspects such as active voice (G5) and prosody (G10) are not measured programmatically.

Ethical Considerations

Personal and sensitive information. The videos show real people in public user-generated YouTube content. No personally identifiable information beyond what is publicly available in the original videos is included. AD annotations may describe people visible in the frames. People appearing in the videos may not have specifically consented to research use of their likeness.

Annotators. Annotations were produced by professional AD teams of sighted and blind writers. Annotator demographic information is not released to protect privacy.

Social impact. The dataset supports video accessibility research for blind and low-vision (BLV) audiences by providing a high-quality benchmark and actionable evaluation for automated inline audio description.

Citation

@inproceedings{venugopalan2026mindthegap,
  title     = {Mind The Gap ({MTG-AD}): Dataset and Fine-grained Evaluation for End-to-End Inline Audio Descriptions},
  author    = {Venugopalan, Subhashini and Tan, Jenny and Roper, Taylor and Tobin, Jimmy and
               Kast, Anton and Martin, Alicia and Sepah, Sam and Pavel, Amy},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Evaluations and Datasets Track},
  year      = {2026}
}

License

The annotations in this dataset are released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. The original videos are not redistributed and remain subject to their owners' rights and YouTube's Terms of Service.

Creator: Google LLC