headmic/BINAURAL_ROADMAP.md

# Binaural Hearing Roadmap
## What two mic arrays make possible

Ranked by impact × feasibility. All build on the existing dual XVF3800 + `/doa` endpoint.

---

### Tier 1 — High impact, ready to build now

**1. Triangulated sound localization + eye gaze**
- Combine DoA angles from both arrays → compute (x, y) position of sound source
- Post gaze coordinates to eye service → eyes track the speaker spatially
- Front/back disambiguation (single array can't tell 30° front from 30° rear)
- *Prereqs:* Known array positions (measured once), basic trig
- *Complexity:* Low — ~100 lines of math + a gaze-push thread
- *Impact:* Huge — eyes actually follow the person, not just shift left/right

**2. Active speaker tracking with smooth gaze**
- Continuously track the dominant sound source as it moves
- Smooth the gaze updates (low-pass filter) so eyes don't jitter
- When VAD drops, eyes drift back to center (natural idle behavior)
- *Prereqs:* #1
- *Complexity:* Low — Kalman filter or exponential smoothing on top of #1
- *Impact:* Makes her feel present and attentive

**3. Left/right speaker awareness**
- Know which side each speaker is on, combine with speaker ID
- "Alex is on my left" vs "unknown person on my right"
- Feed into LYRA context so responses can reference spatial relationships
- *Prereqs:* #1 + existing speaker ID
- *Complexity:* Medium — associate speaker embeddings with spatial positions
- *Impact:* Multi-person conversations become spatially grounded

---

### Tier 2 — High impact, moderate effort

**4. Distance estimation (near/far)**
- Interaural Level Difference (ILD): close sources have bigger volume gap between ears
- Interaural Time Difference (ITD): measurable with raw mic data (would need 6-channel firmware)
- Rough bins: intimate (<0.5m), conversational (0.5-2m), across room (2m+)
- *Prereqs:* #1, calibration with known distances
- *Complexity:* Medium — ILD from processed channels is easy, ITD needs raw mics
- *Impact:* Interaction style adapts to proximity (whisper vs. room voice)

**5. Multi-speaker separation + selective attention**
- Lock each array's beam to a different speaker simultaneously
- Active speaker gets primary audio feed (wake word, transcription)
- Secondary speaker monitored for interruptions or wake word
- Switch attention on cue ("Hey Vivi" from the other side)
- *Prereqs:* #3, understanding of XVF3800 beam steering commands
- *Complexity:* Medium-high — need to control beamformer direction per-array
- *Impact:* Natural multi-person conversations, not just one-at-a-time

**6. Spatial audio scene mapping**
- Build a persistent map: "TV at 270°, door at 90°, kitchen at 180°"
- Learn from repeated sound sources over hours/days
- Detect anomalies: "sound from an unusual direction"
- *Prereqs:* #1, persistent storage, classification by direction
- *Complexity:* Medium — accumulate (direction, category) pairs, cluster over time
- *Impact:* Environmental awareness, contextual anomaly detection

---

### Tier 3 — Cool, needs more infrastructure

**7. Cocktail party spatial filtering**
- When multiple sound sources active, use both arrays to null out interference
- Focus beam on target speaker, suppress others spatially
- *Prereqs:* #5, possibly raw mic access (6-channel firmware)
- *Complexity:* High — adaptive beamforming, may need custom DSP
- *Impact:* Works in noisy environments (music playing, multiple people)

**8. Sound event localization (what + where)**
- Combine YAMNet classification with triangulated position
- "Dog bark from the backyard direction" not just "dog bark"
- Spatial history: timeline of what happened where
- *Prereqs:* #1, #6
- *Complexity:* Medium — merge classification results with position data
- *Impact:* Rich environmental narrative for LYRA context

**9. Head orientation inference**
- If a known sound source is at a fixed position, infer which way the head is "facing"
- Useful if the skull ever gets a rotating mount
- *Prereqs:* #6 (known spatial map)
- *Complexity:* Low math, but needs stable reference points
- *Impact:* Low for now (head doesn't turn), future-proofing

**10. Binaural recording for training data**
- Record stereo audio preserving spatial information (left ear / right ear)
- Training corpus for spatial audio models, being0 sensor data
- *Prereqs:* Just dual streams saved to stereo WAV
- *Complexity:* Low — already have both streams
- *Impact:* Long-term value for L-Vixy-5 training

---

### Tier 4 — Research / future

**11. Learned spatial attention**
- Train a model to decide where to attend based on context
- Input: both DoA angles, VAD states, current emotional state, conversation history
- Output: beam steering + gaze direction
- *Prereqs:* #5, #6, training data from #10
- *Complexity:* High — ML training pipeline
- *Impact:* Autonomous attention that feels natural, not rule-based

**12. Interaural time difference (ITD) processing**
- Raw mic access (6-channel firmware) enables sub-sample timing analysis
- More precise localization than DoA alone, especially at low frequencies
- *Prereqs:* 6-channel firmware (need to verify LED control works with it first)
- *Complexity:* High — signal processing, cross-correlation
- *Impact:* Lab-grade localization accuracy

---

## Implementation order

```
✅ #1  Triangulation + gaze          — done (spatial.py, auto-select beam DoA)
✅ #2  Smooth tracking               — done (exponential smoothing + idle drift)
✅ #3  Speaker-side awareness        — done (Resemblyzer loaded, ready for enrollment)
✅ #4  Distance estimation           — done (ILD + triangulation fusion, proximity zones)
✅ #6  Spatial scene mapping         — done (spatial_scene.py, persistent, anomaly detection)
✅ #8  Sound event localization      — done (what + where + when via /scene/events)
✅ #10 Binaural recording            — done (opt-in via BINAURAL_RECORD=1)
   #5  Multi-speaker separation
   #7  Cocktail party filtering
#7 Cocktail party filtering
#11 Learned attention
```

## Notes

- Items #1-3 can be built in a single session
- The eye service already accepts gaze via `POST /gaze {"x": N, "y": N}`
- DoA is already polled at 10Hz via `/doa` endpoint
- Array separation distance needs to be measured once and stored in config
- All of this feeds into the being0 "shaped by experience" philosophy