prateek-0-gupta/beat-audio2blendshape

BEAT Audio-to-Blendshape

Paired 80-bin log-mel speech features and 51-dim ARKit blendshape tracks, derived from BEAT v1 (english v0.2.1). 6,357 speech segments / 55.5 hours / 30 speakers.

This is the exact data our audio-driven face model was trained on, in the exact train/validation partition it used — not a re-cut sample.

Contents

Segments	6,357 (train 5,970 · validation 387)
Duration	55.5 h (train 52.5 h · validation 3.0 h)
Speakers	30 (validation covers 26)
Audio	log-mel, 80 bins @ 100 fps
Face	ARKit blendshapes, 51 dims @ 30 fps, in [0, 1]

Segments come from energy-VAD speaking spans of the source recordings, so silence is largely excluded. Blendshapes are BEAT's own ARKit capture, resampled from their native 60 fps onto the 30 fps grid and reordered to a canonical name order (face_names.json).

The two tracks run at different rates. mel has ~10/3 frames per face frame; both start at the same instant. Use num_frames / num_mel_frames rather than assuming an exact ratio.

Schema

column	type	notes
id	string	e.g. 2_scott_0_18_18_seg000
speaker_id / speaker_name	int16 / string	1–30, e.g. scott
source_clip	string	recording the segment was cut from
segment_index	int16	index within that recording
emotion_id / emotion	int8 / string	8 classes, see below
start_frame	int32	segment offset in the source clip, @30 fps
num_frames / num_mel_frames	int32	face rows / mel rows
duration_sec	float32	num_frames / 30
face	list[float16]	flattened, reshape to (num_frames, 51)
mel	list[float16]	flattened, reshape to (num_mel_frames, 80)

Arrays are stored flattened as float16 — quantisation error is ≤0.004 on log-mel (range ≈ −11.5 … 8.3) and ≤0.0003 on blendshapes.

import numpy as np
from datasets import load_dataset

ds = load_dataset("prateek-0-gupta/beat-audio2blendshape", split="train")
r = ds[0]
mel  = np.asarray(r["mel"],  np.float16).reshape(-1, 80)   # (T_mel, 80) @100fps
face = np.asarray(r["face"], np.float16).reshape(-1, 51)   # (T,     51) @30fps

Audio frontend

Only log-mel is distributed, not waveforms, so reproducing it on your own audio matters. Fixed constants, streamable frame-by-frame:

Sample rate	16 kHz mono
Window	400 samples (25 ms), Hann
Hop	160 samples (10 ms) → 100 fps
FFT	512
Mels	80, Slaney-style triangular, fmin 20 Hz, fmax 8 kHz
Output	log(mel_energy + 1e-5), no per-utterance normalisation

Mel energies are power (magnitude²) projected through the filterbank; the mel scale is 2595·log10(1 + f/700), with bin edges floored to floor((n_fft + 1)·hz / sr).

Splits

Split by source recording (1,769 train / 93 validation clips), seed 1337, 5% of clips held out. No recording contributes segments to both sides, so a model cannot see a validation clip's other half during training.

Emotion labels

From BEAT's per-recording emotion spans, assigned per segment by largest time overlap. Heavily skewed — treat as weak conditioning, not a balanced task:

neutral 76.6% · happiness 3.9% · contempt 3.9% · sadness 3.6% · disgust 3.5% · fear 3.2% · surprise 2.8% · anger 2.5%

Provenance and license

Derived from BEAT v1 (Liu et al., ECCV 2022). All speech and blendshape content originates there; this release contributes only segmentation, feature extraction, and alignment.

⚠️ The upstream license is ambiguous. The official BEAT page states "Licensed under the Non-commercial license", while the H-Liu1997/BEAT mirror is tagged apache-2.0. This derivative is released under the stricter reading, CC BY-NC 4.0 — non-commercial use only. Confirm terms with the BEAT authors before any commercial use, and cite BEAT:

@inproceedings{liu2022beat,
  title={{BEAT}: A Large-Scale Semantic and Emotional Multi-Modal Dataset
         for Conversational Gestures Synthesis},
  author={Liu, Haiyang and Zhu, Zihao and Iwamoto, Naoya and Peng, Yichen and
          Li, Zhengqing and Zhou, You and Bozkurt, Elif and Zheng, Bo},
  booktitle={ECCV},
  year={2022}
}

The recordings are of identifiable speakers. Do not use this data to clone, impersonate, or synthesise likenesses of them without their consent.

Limitations

• No waveforms. Log-mel only, so you cannot substitute your own audio encoder (wav2vec2, HuBERT, Whisper) — you are locked into this frontend. Recover raw audio from BEAT directly if you need it.
• Acted, not spontaneous. BEAT is studio-recorded scripted/acted speech; expect a domain gap against conversational audio.
• Blendshape quality is BEAT's. Capture artefacts, tracking noise, and speaker-specific rest poses pass through unchanged; no cleanup was applied.
• Unnormalised mel — no CMVN or per-speaker normalisation, by design.
• English only.