xg-chu/TalkVid_Fallingwater

Dataset Card for TalkVid-Fallingwater

Paired speech + 3D facial motion data: 145.6 hours of 24 kHz waveforms aligned frame-for-frame with 108-d FLAME motion codes at 25 fps. This is the training and evaluation data for the Fallingwater speech-driven facial motion model.

Dataset Details

Dataset Description

TalkVid-Fallingwater is a processed, model-ready derivative of TalkVid. The source videos are not redistributed here. Instead, each clip has been reduced to the two streams a motion model actually consumes — a mono speech waveform and a per-frame parametric face/head motion code — and packed into a single LMDB store for fast random access during training.

The 24 kHz audio rate is not arbitrary: Fallingwater encodes speech with Mimi, which operates natively at 24 kHz, so waveforms are stored at the rate the model consumes. Likewise the 108-d motion layout unpacks directly into FLAME parameters with no remapping at load time.

• Curated by: Xuangeng Chu
• Language(s): English (10,531 clips), Chinese (1,689), Spanish (201), Hindi (57), other/unknown (610)
• License: CC BY-NC 4.0, inherited from upstream TalkVid
• Derived from: FreedomIntelligence/TalkVid

Clips	13,111
Duration	145.6 h (13,106,974 motion frames)
Clip length	12 s – 22.7 min (mean 40 s)
Motion	float16, (T, 108), 25 fps
Audio	int16 mono, 24 kHz, exactly T × 960 samples
Store size	26.1 GiB

Dataset Sources

• Upstream dataset: FreedomIntelligence/TalkVid
• Upstream paper: TalkVid: A Large-Scale Diversified Audio-Visual Dataset for Talking Head Synthesis (arXiv:2508.13618)
• Original videos: YouTube; per-clip links in metadata.json under info["Video Link"]

Uses

Direct Use

• Training and evaluating speech-driven 3D facial animation models (audio → FLAME expression, jaw, eye and head pose)
• Motion tokenization / discrete motion codec research (the data is what Fallingwater's BSQ codec stage is fit on)
• Speaker-conditioned or style-conditioned motion generation, using info["Person ID"] to group clips by identity
• Analysis of co-speech head and face dynamics at scale

Out-of-Scope Use

• Any commercial use. The CC BY-NC 4.0 license forbids it.
• Photorealistic identity reconstruction. Motion codes carry no FLAME shape parameters, no appearance, and no video frames — the dataset does not support reconstructing what a speaker looks like.
• Impersonation or synthetic media of the source speakers. These are identifiable real people who uploaded videos to YouTube and did not consent to being animated.
• Speech recognition or speaker identification. Audio is stored in 12 s – 22.7 min clips without transcripts and is included as a motion-conditioning signal, not as an ASR/SID corpus.
• Treating demographic labels as ground truth. Gender, Age Group and Ethnicity are inherited upstream annotations over coarse categories, not self-reported identity.

Dataset Structure

data.mdb              # LMDB store, 13,111 entries keyed by clip id
metadata.json         # per-clip annotations (list of 13,111 records)
split.json            # {"train": [...], "val": [...], "test": [...]}
talkvid_stats.json    # per-dim mean/std over the 108 motion dims, + 112->108 dim mapping

data.mdb and metadata.json cover exactly the same set of clip ids.

Data Instances

Every LMDB value is a serialized .npz archive with two arrays:

Array	dtype	Shape	Notes
motion_code	float16	(T, 108)	25 fps; T equals num_frames in metadata.json
audio	int16	(T * 960,)	mono 24 kHz, frame-aligned with motion_code

Audio is integer PCM; divide by 32768.0 for [-1, 1] floats.

Motion Code Layout

Each frame is a 108-d vector that unpacks into FLAME parameters (n_shape=300, n_exp=100) as [100, 3, 1, 4]:

Slice	Dim	Content
[..., 0:100]	100	FLAME expression coefficients
[..., 100:103]	3	global head pose (axis-angle rotation)
[..., 103:104]	1	jaw opening — FLAME's jaw_pose[0], expanded to [jaw, 0, 0] at decode time
[..., 104:108]	4	eye pose — 2 per eye, expanded to [l_x, l_y, 0, r_x, r_y, 0] at decode time

Identity (FLAME shape) is deliberately not included: the motion code is identity-agnostic, and a shape code can be supplied separately at render time.

This layout is the tracker's original 112-d output with 4 constant/degenerate dimensions removed — the two unused jaw axes and one rotation axis per eye. talkvid_stats.json records the exact original_to_new_mapping (112 → 108), dropped_original_dims ([104, 105, 108, 111]), and the per-dimension mean / std used for normalization.

Data Fields

Each record in metadata.json:

Field	Description
id	clip id, matching the LMDB key
num_frames	motion sequence length T
info	upstream annotations: Person ID, Gender, Age Group, Ethnicity, Language, Video Category, Video Link
head_detail	per-clip head-tracking quality scores (movement, rotation, completeness, resolution)
dover_scores	DOVER video quality score
cotracker_ratio	CoTracker point-tracking consistency ratio
descriptions	natural-language descriptions of the speaker and scene

Data Splits

Split	Clips	Frames	Hours
train	12,626	12,642,433	140.47
test	485	459,191	5.10
val	8	5,350	0.06

train and test partition the dataset (12,626 + 485 = 13,111). val is an 8-clip subset of test, intended as a cheap monitoring set during training — it is not a third disjoint split, and metrics on it are not independent of test.

Splits are clip-level, not speaker-level: the same Person ID can appear in both train and test. If you need speaker-disjoint evaluation, re-split by grouping on info["Person ID"].

Dataset Creation

Curation Rationale

TalkVid ships full videos with rich annotations, which is far more than a parametric motion model needs and far too slow to stream during training. This derivative keeps only what Fallingwater consumes, at the exact rates it consumes them, so that a training step is a single LMDB read and two tensor slices — no decoding, no resampling, no per-sample alignment arithmetic.

Source Data

Data Collection and Processing

Starting from TalkVid's curated clips, a face-tracking pipeline extracted per-frame FLAME parameters, which were reduced from 112 to 108 dimensions by dropping constant axes (see Motion Code Layout). Audio was resampled to 24 kHz mono int16 and trimmed so that every clip satisfies len(audio) == T * 960 exactly. Clips shorter than 300 frames (12 s) were dropped, since the training window plus autoregressive context needs more than that. Per-dimension mean and standard deviation were computed over the full set and stored in talkvid_stats.json.

Quality signals from upstream — DOVER video quality, CoTracker tracking consistency, and head-motion detail scores — are retained per clip in metadata.json so downstream users can filter further.

Who are the source data producers?

The underlying videos are YouTube uploads by their original creators, collected and curated by the TalkVid authors. Neither TalkVid nor this derivative involved recording new footage.

Annotations

No new human annotation was performed for this derivative. All demographic and categorical labels (Person ID, Gender, Age Group, Ethnicity, Language, Video Category), the natural-language descriptions, and the quality scores are inherited from TalkVid. Motion codes are machine-generated by the tracking pipeline and were not manually verified.

Personal and Sensitive Information

This dataset concerns identifiable real people and their data should be treated accordingly:

• info["Video Link"] points to the original YouTube video, which ties every clip to a public identity.
• info["Person ID"] groups clips by speaker, making per-individual aggregation trivial.
• Gender, Age Group and Ethnicity are demographic labels attached to those individuals. They are coarse, third-party inferred, and include an Unknown category — they are not self-reported.
• Audio contains real voices and unredacted speech content; there are no transcripts, so the spoken content has not been reviewed.

No video frames, images, or FLAME shape parameters are included, so the release does not carry facial appearance. Speech and speaker grouping are nevertheless personal data, and the non-commercial license does not by itself make every use appropriate.

Bias, Risks, and Limitations

Attribute distributions, inherited from TalkVid:

Attribute	Distribution
Language	English 10,531 · Chinese 1,689 · Unknown 595 · Spanish 201 · Hindi 57 · Other 15
Gender	Female 9,500 · Male 3,599 · Unknown 12
Age group	31–45 8,774 · 19–30 3,037 · 46–60 950 · 60+ 324 · Unknown 26
Ethnicity	White 8,479 · Asian 2,928 · African 1,692 · Unknown 12
Category	Personal Experience 5,620 · Vlogger/Creator 3,171 · Popular Science 663 · Global Culture 594 · Health Advice 506 · Life Hacks 312 · …

• Skewed composition. 80% English, 72% female, 67% aged 31–45, 65% labeled White. Models trained here will fit those groups best; cross-lingual lip synchrony in particular is likely to degrade on the long tail.
• Content monoculture. Nearly all clips are single-speaker pieces to camera (vlogs, personal stories, explainers). Conversational turn-taking, listening behavior, multi-party dynamics, and non-frontal framing are effectively absent.
• Tracker-limited ceiling. Motion codes are only as good as the tracking pipeline. Errors — jitter, lip closure failures, eye-gaze drift — are baked in and no amount of downstream modeling recovers the ground truth. Anything the 108-d parameterization cannot express (tongue, wrinkles, non-rigid detail) is simply not there.
• Lossy by design. Motion is stored as float16, and 4 tracked dimensions were dropped.
• Overlapping splits. val ⊂ test, and speakers span train/test (see Data Splits).
• Synthesis misuse. The dataset's purpose is generating plausible talking-face motion, which is inherently dual-use for deceptive synthetic media.

Recommendations

Report results on test, not val, and say which you used. For any claim about generalization to new speakers, re-split on Person ID first. When reporting per-language or per-demographic metrics, include group sizes — several groups here are small enough that a single bad clip moves the number. Filter on dover_scores, cotracker_ratio and head_detail if your application is sensitive to tracking noise. Disclose synthetic outputs as synthetic.

How to Use

hf download xg-chu/TalkVid_Fallingwater --repo-type dataset --local-dir talkvid_lmdb

import io
import json

import lmdb
import numpy as np

env = lmdb.open("talkvid_lmdb", readonly=True, lock=False, readahead=False, meminit=True)
split = json.load(open("talkvid_lmdb/split.json"))
stats = json.load(open("talkvid_lmdb/talkvid_stats.json"))
mean = np.array(stats["mean"], dtype=np.float32)
std = np.array(stats["std"], dtype=np.float32)

with env.begin(write=False) as txn:
    payload = np.load(io.BytesIO(txn.get(split["train"][0].encode())), allow_pickle=False)

motion = payload["motion_code"].astype(np.float32)  # (T, 108) @ 25 fps
audio = payload["audio"].astype(np.float32) / 32768.0  # (T * 960,) @ 24 kHz
motion = (motion - mean) / (std + 1e-8)  # normalization used by Fallingwater

Open the environment once per worker and reuse it — readonly=True, lock=False makes the store safe to share across DataLoader workers.

Rendering a motion code to a mesh requires the FLAME model, licensed separately and available from the FLAME project page:

verts = flame_model.get_flame_verts(motion_code)  # (B, T, V, 3), FLAMEModel(n_shape=300, n_exp=100)

Consumption by Fallingwater

Fallingwater is a two-stage speech-driven facial motion model. Both stages read this store through one loader configured with MOTION_FPS: 25, AUDIO_SAMPLE_RATE: 24000, MOTION_DIM: 108, pointed at data.mdb, metadata.json and split.json; talkvid_stats.json is loaded separately as the model's normalization buffer.

Stage 1 — motion codec. A transformer encoder/decoder with a multi-scale BSQ quantizer (scale schedule [1, 25, 50, 100], code dim 32) compresses normalized motion into discrete tokens. Reconstruction is supervised in both parameter and mesh space: expression, head pose, head velocity and smoothness, FLAME vertex, and lip terms. Head pose is weighted well above expression, so dimensions [100:104] carry disproportionate weight at this stage.

Stage 2 — autoregressive generator. A transformer over stage-1 tokens, conditioned on Mimi audio features, the previous motion chunk, and a style template: a 100-frame clip sampled from a different clip of the same Person ID. Each condition is independently dropped 10% of the time to enable classifier-free guidance at inference. This is why speaker grouping is a first-class part of the metadata.

Sampling. Training draws 100-frame (4 s) windows at a 25-frame stride, plus the preceding window as autoregressive context (zero-padded at clip start). Validation and test clips are truncated to 750 frames (30 s).

Citation

Please cite the upstream dataset:

@misc{chen2025talkvidlargescalediversifieddataset,
      title={TalkVid: A Large-Scale Diversified Dataset for Audio-Driven Talking Head Synthesis}, 
      author={Shunian Chen and Hejin Huang and Yexin Liu and Zihan Ye and Pengcheng Chen and Chenghao Zhu and Michael Guan and Rongsheng Wang and Junying Chen and Guanbin Li and Ser-Nam Lim and Harry Yang and Benyou Wang},
      year={2025},
      eprint={2508.13618},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2508.13618}, 
}

@misc{chu2026personalizingcausalaudiodrivenfacial,
      title={Personalizing Causal Audio-Driven Facial Motion via Dynamic Multi-modal Retrieval}, 
      author={Xuangeng Chu and Yu Han and Wei Mao and Shih-En Wei},
      year={2026},
      eprint={2604.23692},
      archivePrefix={arXiv},
      primaryClass={cs.GR},
      url={https://arxiv.org/abs/2604.23692}, 
}

Dataset Card Authors

Xuangeng Chu

Dataset Card Contact

Open a discussion on this dataset's Community tab.