Audio2Face Emotion Arkit Teacher: 14k Emotional Speech Clips with Blendshape Labels

14,082 emotional-speech clips annotated with facial animation data. Each clip contains two parallel 52-channel ARKit blendshape sequences generated by NVIDIA Audio2Face-3D-v2.3.1-James and LAM_Audio2Expression models, plus a 26-dimensional NVIDIA Audio2Emotion conditioning vector. The dataset was created by myned-ai and last updated on HuggingFace in June 2026.

Use Cases

• Train audio-to-blendshape models based on the parallel 52-channel ARKit sequences.
• Benchmark emotion-conditioned animation models using the 26-dimensional Audio2Emotion vectors.
• Develop multi-model facial animation pipelines by comparing outputs from NVIDIA Audio2Face and LAM_Audio2Expression.
• Create expressive virtual avatars, like the Nyx avatar mentioned, driven by emotional speech.

Strengths

• Contains 14,082 annotated emotional-speech clips.
• Provides two parallel 52-channel ARKit blendshape sequences per clip for model comparison.
• Includes a 26-dimensional emotion conditioning vector for each clip.

Limitations

• The original audio is not shipped, limiting use to the provided blendshape labels.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is known, but other scale metrics like file size and format are unknown.

Provenance

Source

myned-ai on HuggingFace

Collection Method

Blendshape labels derived from emotional-speech clips using NVIDIA Audio2Face-3D-v2.3.1-James and LAM_Audio2Expression models.

Freshness

Last updated 2026-06-09 09:59:34; freshness should be verified.