Zero-training trick lifts EmbeddingGemma-2 to MAEB top

A researcher projected intermediate Conformer layers into EmbeddingGemma-2's 256-dim null space via PCA, raising the MAEB audio macro score from 49.3% to 54.4% with zero training and moving from rank 26 to 1. The work argues text alignment filters paralinguistic traits like emotion and timbre.

• EmbeddingGemma 2 Turned Into On-Device Speech-to-Text Model2026-10-13
• Google launches local-first AI meeting note-taker Foresight2026-10-08
• Google Releases Open Multimodal Embedding Model EmbeddingGemma 22026-10-06
• Google opens Gemini and Search to third-party apps2026-10-28
• Google AI Mode gives evasive answers on 2020 election2026-10-28