MotionSpeech: A Co-speech Gesture and Speech Dataset

MotionSpeech is a aligned multimodal dataset of 35,227 paired clips (~12.9 s each) of synchronized speech audio and full-body SMPL-X motion, with multi-layer textual annotations. It is assembled from two public sources and re-segmented, re-annotated and quality-filtered for training text-driven joint motion+speech generative models.

Source	Clips	Notes
BEAT2 (English)	18,063	scripted monologue gestures, 30 speakers
Seamless Interaction (improvised segments)	17,164	improvised English dyadic interaction

Directory structure

motionspeech_data_opensource/
├── README.md            this file
├── split/               official splits (sample IDs, one per line)
│   ├── train.txt        28,175
│   ├── val.txt           3,526
│   ├── test.txt          3,526
│   └── all.txt          35,227 (= train ∪ val ∪ test)
├── audio_wav/           <sample_id>.wav  mono 16-bit PCM
├── motion_smplx/        <sample_id>.npz  SMPL-X + FLAME pose sequence
└── caption/             <sample_id>.txt  multi-layer annotation (see below)

Every sample ID present in split/ has all three data files (audio, motion, caption) — the four modalities are strictly aligned.

Sample ID convention

Two naming patterns, inherited from the sources:

• BEAT2: {speaker_id}_{name}_{take}_{session}_{seg}_{start_frame}_{end_frame} e.g. 10_kieks_0_10_10_1256_1642
• Seamless: V{video}_S{session}_I{interaction}_P{participant}_{start_frame}_{end_frame} e.g. V03_S1399_I00000495_P4044_7194_7580

All files of one clip share the same ID (only the extension differs).

Audio (audio_wav/)

• WAV, mono, 16-bit PCM.
• BEAT2 clips are 16 kHz; Seamless clips are 48 kHz (native rates of the sources are preserved — resample on load if a uniform rate is needed).
• Each clip is ~12.9 s, aligned 1:1 in time with its motion file.

Motion (motion_smplx/)

Each .npz contains one clip at 30 fps (387 frames ≈ 12.9 s):

Key	Shape	Description
poses	(T, 165)	SMPL-X joint rotations in axis-angle (55 joints × 3)
trans	(T, 3)	global root translation (meters)
expressions	(T, 100)	FLAME facial-expression coefficients
betas	(300,)	SMPL-X body-shape coefficients
model	str	model identifier (smplx+flame)
gender	str	male / female
mocap_frame_rate	int	capture frame rate (30)

Load with numpy.load(path); drive SMPL-X+FLAME with the poses/trans/betas to obtain meshes. (For 6D-rotation training features used in our baselines, poses can be converted 3-axis-angle → 6D per joint.)

Captions (caption/)

Each .txt holds four tagged layers concatenated in one line:

<L1>{one-sentence summary of the gesture}</L1>
<L2>{step-by-step detailed gesture description}</L2>
<S>{verbatim speech transcript, lowercase, no punctuation}</S>
<AUDCAP>{speaker voice description: gender, timbre, tone, pace}</AUDCAP>

Example (truncated):

<L1>The hands stay near the abdomen while the body shifts laterally, then both
palms open slightly outward.</L1>
<L2>First, both elbows remain bent and the hands are held apart in front of the
abdomen with gently curved fingers. Next, the hands move closer together ...
Finally, the hands remain low and open ...</L2>
<S>i would prefer to choose a major that is easy to find a good job in the
future like finance or marketing ...</S>
<AUDCAP>Female voice, clear and steady, speaking in an informative tone.</AUDCAP>

• L1 / L2 describe co-speech body gestures at two granularities.
• S is the speech transcript of the paired audio (already normalized to lowercase ASCII without punctuation).
• AUDCAP is a compact speaker-voice caption (gender / timbre / tone / pace) usable as a conditioning prompt for voice characteristics.

Intended use

Text-driven / phoneme-driven joint generation of co-speech gesture and speech (TTS-style audio + gesture avatars), gesture-speech synchronization research, and captioning of multimodal speech+motion streams.

License & attribution

This dataset is a re-segmentation and re-annotation of two public corpora. Users must comply with the original licenses of BEAT2 and Seamless Interaction, and cite both sources in any derived work.