The emotion2vec+ large speech-emotion-recognition model converted to MLX format for native inference on Apple Silicon, consumed by the xocialize/emotion2vec-mlx-swift Swift port. Refer to the original model card for details.

Model

Family: emotion2vec / emotion2vec+ (Ma et al., "emotion2vec: Self-Supervised Pre-Training for Speech Emotion Representation," arXiv:2312.15185)
Architecture: Data2Vec 2.0 — conv feature extractor → transformer encoder → 9-class linear head
Output: 9-class categorical emotion (angry, disgusted, fearful, happy, neutral, other, sad, surprised, unknown)
Sample rate: 16000 Hz, mono
Precision: fp16 (233 tensors)

Files

emotion2vec_large.safetensors — the MLX weights (fp16).
emotion2vec_large_config.json — model config consumed by the loader.

Usage (Swift / MLX)

swift

import Emotion2VecMLX
import Hub

let dir = try await HubApi().snapshot(from: "mlx-community/emotion2vec-plus-large-mlx")
let recogniser = try await EmotionRecogniser(weightsDirectory: dir,
                                             config: EmotionRecogniserConfig(models: .categorical))
let result = try await recogniser.classify(audioURL: speechURL)
print(result.categorical.label, result.categorical.confidence)

Source

Original model: https://huggingface.co/emotion2vec/emotion2vec_plus_large
Swift consumer: https://github.com/xocialize/emotion2vec-mlx-swift

License

FunASR's custom MODEL_LICENSE — permits use, copy, modification, and redistribution with attribution and model-name retention (no-denigration clause, no warranty). Non-SPDX but permissive. See the original license.