HRTF-guided Binaural Target Speaker Extraction with Real-World Validation
New AI uses your unique ear shape to extract a single voice from a noisy room without losing spatial audio cues.
Researchers Yoav Ellinson and Sharon Gannot have published a new paper proposing a novel AI framework for isolating a single speaker's voice in noisy, multi-speaker environments. The key innovation is using the listener's Head-Related Transfer Function (HRTF)—the acoustic fingerprint of how their unique head and ear shape filter sound—as an explicit spatial guide for the AI. This is a departure from conventional Target Speaker Extraction (TSE) methods that rely on estimating a speaker's Direction of Arrival (DOA) or a separate enrollment sample, approaches which often degrade the perceived 3D spatial location of the sound.
The framework is built on a multi-channel deep blind source separation model, adapted for the binaural (two-ear) setting. Crucially, it was trained on a dataset of measured HRTFs from a diverse population, allowing it to generalize to new listeners without requiring subject-specific tuning. By conditioning the audio extraction process on this HRTF-derived spatial information, the system can pull a target voice from a mixture while meticulously preserving the natural binaural cues that tell our brain where a sound is coming from in space.
The team validated their approach rigorously, not just in simulations but with real-world recordings captured using an anthropomorphic head and torso simulator (HATS). The results demonstrate significant improvements in preserving spatial audio perception while simultaneously enhancing speech quality and intelligibility. This work, submitted to Interspeech 2026, represents a meaningful step toward more natural and effective hearing aids, augmented reality audio, and teleconferencing systems that can replicate the clarity and spatial realism of a one-on-one conversation, even in a crowd.
- Uses listener's HRTF (ear shape signature) as a spatial guide, unlike DOA-based methods that distort sound location.
- Trained on diverse HRTF data for cross-listener generalization, avoiding need for personal calibration.
- Validated with real HATS recordings, showing preserved binaural cues alongside improved speech clarity.
Why It Matters
Enables next-gen hearing aids and AR audio that isolate voices in crowds without losing crucial 3D spatial sound cues.