REFRAMED: Towards Realistic Audio Description Generation for Movies

Abstract

Audio Description (AD) is a verbal narration of key visual content in videos, enabling access for visually impaired audiences. Unlike standard video captioning, AD is a structured editorial task: descriptions must be inserted into gaps in dialogue and must convey only what is needed to understand the narrative being told. However, existing approaches formulate AD generation in an artificial setting where both the content and timing of descriptions are pre-specified, reducing the task to clip-level captioning. They further rely on noisy transcription and alignment pipelines, and lack the rich parallel data required for modeling narrative context. We introduce a new formulation of AD generation in which models must jointly decide what to describe and when to do it. To support this, we present REFRAMED, a high-quality dataset of 2,023 videos that span 3,302 scenes from 206 movies, with professional AD transcripts (both American and British versions), professional subtitles and aligned screenplays. We also provide a manually curated challenge set that pairs full movies with multiple AD references, together with evaluation protocols that leverage dialogue gaps and multi-reference comparisons. Experiments with state-of-the-art AD systems and multimodal LLMs show that they outperform trivial baselines but fall far short of expert human performance. Our dataset and benchmark establish a new foundation for research on video understanding.

Data format

Each row corresponds to one video excerpt. AD and subtitles are provided in SRT format. One option for processing them in Python is pysrt. Screenplays are provided as JSON-serialized dictionaries, or the string "null" when unavailable. Keys in the dictionaries correspond to scene indices in the MovieSum dataset and values contain the aligned scenes’ XML.

Decryption

The test set has been encrypted (following tell_me_a_story). The symmetric key skey.key and private_key.pem are provided in keys.zip. The following example decrypts one value from a loaded dataset (ensure you have cryptography installed)

from datasets import load_dataset

from cryptography.hazmat.primitives import hashes
from cryptography.hazmat.primitives.asymmetric import padding
from cryptography.hazmat.backends import default_backend
from cryptography.hazmat.primitives import serialization
from cryptography.fernet import Fernet

import json

test = load_dataset("igorsterner/reframed", split="test", streaming=True)
example = next(iter(test))

# Load the private key.
with open("keys/private_key.pem", "rb") as f:
    private_key = serialization.load_pem_private_key(
        f.read(),
        password=None,
        backend=default_backend()
    )

# Load the symmetrical key.
with open("keys/skey.key", "rb") as f:
    skey = f.read()

# Decrypt the symmetrical key.
unenc_skey = private_key.decrypt(
    skey,
    padding.OAEP(
        mgf=padding.MGF1(algorithm=hashes.SHA256()),
        algorithm=hashes.SHA256(),
        label=None
    )
)

# Decrypt a value as follows.
f = Fernet(unenc_skey)
encrypted_value = example["imdb_id"]
decrypted_value = json.loads(
    f.decrypt(encrypted_value.encode("ascii")).decode("utf-8")
)

print(decrypted_value)

Videos

You can view a video at https://www.rottentomatoes.com/m/{rottentomatoes_id}/videos/{video_id}. See videos_encrypted.sh for more. This bash file is encrypted with the same key as above and can be decrypted by applying f.decrypt(...) to its contents.

License

Creative Commons Attribution Non Commercial 4.0

Citation

@inproceedings{sterner2026reframed,
    title={{REFRAMED}: Towards Realistic Audio Description Generation for Movies},
    author={Igor Sterner and Mirella Lapata and Alex Lascarides and Frank Keller},
    booktitle={Third Conference on Language Modeling},
    year={2026},
    url={https://openreview.net/forum?id=8VyU9qy4Vy}
}