Acoustic embeddings via deep models (wav2vec2, HuBERT)

[Hugo-W]

Follow-up from #23.

Description

The stimulus feature extraction module (pyeeg/features/) currently supports classical acoustic features (envelope, filterbank, gammatone). Issue #23 originally called for deep-learning-based acoustic embeddings via wav2vec2 and HuBERT.

Tasks

• Add DeepAcousticFeatureExtractor class to pyeeg/features/acoustic.py
• Support wav2vec2 (HuggingFace facebook/wav2vec2-base-960h or similar)
• Support HuBERT (HuggingFace facebook/hubert-base-ls960 or similar)
• Extract hidden-layer embeddings as per-frame feature vectors
• Integrate into FeaturePipeline via extractor_type='acoustic_deep' (or extend existing "acoustic" type)
• Add StimulusEncoder.add_deep_acoustic_features() convenience method
• Handle model download from ModelScope mirror (HF Hub may be firewalled)
• Test suite (unit tests with synthetic audio, integration test with real model if available)

Context

• The AcousticFeatureExtractor (envelope/filterbank/gammatone) is already implemented and tested
• The pipeline architecture (FeatureSpec, FeaturePipeline, StimulusEncoder) supports adding new extractor types
• Models can be downloaded from ModelScope (https://modelscope.ai) if HuggingFace Hub is unavailable

Priority: Medium