Research map / Large language models / Language, safety and interpretability
PhaseCoder: Microphone Geometry-Agnostic Spatial Audio Understanding for Multimodal LLMs
Artem Dementyev, Wazeer Zulfikar, Sinan Hersek, Pascal Getreuer, Anurag Kumar, Vivek Kumar
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune Gemma with PhaseCoder spatial tokens (vs. baseline Gemma) | Gemma 3n · Synthetic Task 2 reasoning | accuracy (%): 48.44 -> 76.76 | Improved |
| fine-tune Gemma with spatial tokens | Gemma 3n · Synthetic Task 1 localization | azimuth MAE (degrees): 74.28 -> 4.75 | Improved |
| fine-tune Gemma with spatial tokens | Gemma 3n · RSL2019 Task 2 reasoning | accuracy (%): 53.91 -> 73.83 | Improved |
| add spatial tokens to Gemma transcription | Gemma 3n · Synthetic Task 4 | WER mean (%): 23.71 -> 8.77 | Improved |
| add spatial tokens to Gemma transcription | Gemma 3n · Synthetic Task 3 | WER mean (%): 6.40 -> 10.63 | Hurt |
| route localization through the LLM instead of standalone PhaseCoder | PhaseCoder · Synthetic Task 1 | azimuth MAE (degrees): 3.08 -> 4.75 | Hurt |
| use PhaseCoder-6M instead of GI-DOAEnet | GI-DOAEnet · LOCATA | Acc@10 (%): 82.48 -> 86.96 | Improved |
| use PhaseCoder-6M instead of GI-DOAEnet | GI-DOAEnet · RSL test | Acc@10 (%): 85.96 -> 82.31 | Hurt |
| scale PhaseCoder from 1.5M to 6M parameters (ablation) | PhaseCoder · LOCATA | Acc@10 (%): 40.00 -> 86.96 | Improved |
| increase microphones from 3 to 4 (ablation) | PhaseCoder · Synthetic held-out batch | azimuth MAE (degrees): 80.24 -> 10.27 | Improved |
| increase microphones from 4 to 8 (ablation) | PhaseCoder · Synthetic held-out batch | azimuth MAE (degrees): 10.27 -> 3.03 | Improved |
| increase microphones from 3 to 8 (ablation) | PhaseCoder · Synthetic held-out batch | distance MAE (m): 1.13 -> 0.71 | Improved |
Closest papers by meaning
- OWL : Geometry-Aware Spatial Reasoning for Audio Large Language Models (ICLR 2026 · similarity 0.89)
- The World is Not Mono: Enabling Spatial Understanding in Large Audio-Language Models (NeurIPS 2026 · similarity 0.87)
- Hear, Localize, and Reason: Spatially Aware Scene Understanding for Audio-visual LLMs (NeurIPS 2026 · similarity 0.87)
- MRSAudio: A Large-Scale Multimodal Recorded Spatial Audio Dataset with Refined Annotations (NeurIPS 2025 · similarity 0.87)
- BAT: Learning to Reason about Spatial Sounds with Large Language Models (ICML 2024 · similarity 0.87)
- SING: Spatial Context in Large Language Model for Next-Gen Wearables (ICML 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
