A perception-first spatial audio engine. Echo is part of the Aura product family.
https://github.com/TheBarmaEffect/echo/releases/download/v1.1.0/echo_launch_1080p.mp4
Wear headphones: everything before the flip at 1:05 is an untouched original mix, and everything after it is Echo's actual output, tested on the live track "Breakdown". Also available on the v1.1.0 release (1080p, plus a 9:16 cut and poster/logo art).
Media rights. The launch film and its soundtrack, the song "Breakdown" (written and performed by Karthik Barma), are not part of this repository and are not covered by the Apache-2.0 license below. All rights to the music and video are reserved by the author. The Apache-2.0 license applies only to the source code and documentation in this repository.
Echo lifts flat media into an explicit, inspectable sound scene graph - the "Sound Scene DNA" - and treats that graph, not any single audio file, as the thing it produces. Every render (binaural, ambisonics, object metadata, device mix) is derived from the scene graph and can be regenerated, re-inspected, and re-scored at any time. Echo is not an enhancement effect layered onto a finished mix; it is a reconstructor that recovers a structured spatial scene and renders it deterministically.
Echo lifts flat media into an explicit, inspectable sound scene graph and renders it as personalized binaural audio - engineered for how spatial sound is perceived, not just how it measures.
The systems contribution is the combination, not any one part: an explicit
intermediate representation, listener-conditioned render policies, and a
perceptual evaluation methodology (EchoBench) that unify a pipeline otherwise
addressed piecewise across the prior art Echo builds on. See
docs/research/ for that prior art. No comparative claims.
| Echo is | Echo is not |
|---|---|
| A typed, inspectable scene-graph representation of a spatial audio scene | A one-click "make it sound spatial" enhancement effect |
| A deterministic binaural renderer - identical inputs give byte-identical output on a pinned platform | A real-time or live plug-in (v0 is offline, file-in / file-out) |
| Engineered for how spatial sound is perceived, with an explicit evaluation harness | A guaranteed producer of any felt outcome - externalization, envelopment, and intimacy are hypotheses it is designed for, not results it claims |
| Offline-first, deterministic core built on NumPy/SciPy | Dependent on any cloud service, account, or specific pair of headphones |
| Honest about its boundaries: separation and visual analysis are opt-in and marked EXPERIMENTAL until verified | A replacement for, or competitor to, any commercial spatial-audio format |
Echo's language discipline is deliberate: felt-experience outcomes are always
written as "designed for" / "targets" / "aims to", never "achieves" or
"delivers". See docs/CLAIMS.md for the per-capability
Allowed / Not-allowed / Limitation table.
Numbers here are populated only from live evidence runs logged in
docs/AUDIT.md. Every cell below is backed by a live run logged
there - nothing is asserted from design intent alone. These are objective render
properties on synthetic scenes; felt-experience outcomes (externalization,
envelopment, intimacy) remain hypotheses until listener testing.
| Metric | Target (from SPEC) | Measured | Evidence |
|---|---|---|---|
| ITD azimuth recovery error (synthetic, θ ∈ {−90…90}) | ≤ ±7.5° | max ~5-14 µs off Woodworth; all 9 angles within tolerance | AUDIT.md python -m echo probe |
| ITD / ILD sign correctness (left source → both positive) | 100% of test angles | 100% (G0 temporal + spectral probes PASS) | AUDIT.md python -m echo probe |
| Per-band ILD sign + monotonicity | correct | correct (net head-shadow growth confirmed) | AUDIT.md python -m echo probe |
| Render determinism (SHA-256 byte-identity across two runs) | identical | identical (WAV file byte-for-byte, threads pinned) | AUDIT.md hero render ×2 |
| True-peak ceiling | ≤ −1 dBTP | ≤ −1 dBTP (single static gain, no compression) | AUDIT.md hero render |
| Scene score / gate verdict on example scenes | score ≥ 70, zero criticals | courtyard 93.4 PASS; flyby PASS | AUDIT.md example renders |
Beyond the scene-graph pipeline, Echo ships a deterministic mastering and
up-mix toolchain (python -m echo master|immerse, scripts/), built around
one rule: every output is self-diagnosed against measured release gates
before it is called done.
- Auto-master (
echo.render.enhance) - bounded, transparent corrections toward a measured professional reference envelope: subsonic hygiene, capped linear-phase balance, staged limiter-delivered loudness with a hard gain-reduction budget. - Immersive up-mix (
echo.render.immersive) - the two-bus venue architecture (dry never delayed; three venue engines on a ducked wet bus), with content-adaptive safeties and a closed loop: the render measures its own output with the release gate's detectors and backs off until the gate rule holds. - Release gates (
echo.verify.qa_gates) - one shared implementation of the introduced-artifact, phase-damage, envelope and true-peak rules, used identically by the corpus QA harness (scripts/corpus_qa.py, currently 12/12 real tracks across 9 genres), the library sweep (scripts/catalog_upgrade.py) and the per-asset QC certificate (scripts/qc_certificate.py). - Listening room (
scripts/listening_room.py) - a single self-contained HTML page that plays every render variant in sample-synchronized lockstep with level-matched, blind-capable switching, so a human verdict is formed fairly. It is an instrument, not a metric: no aggregate perceptual score exists anywhere in Echo.
docs/README.md- pipeline stages, quickstart, and the honest capability table.SPEC.md- the locked engine specification (the contract the code is measured against).docs/ARCHITECTURE.md- the why behind each permanent commitment.docs/CLAIMS.md- what may and may not be claimed, per capability.docs/AUDIT.md- append-only log of live-run evidence.ROADMAP.md- where we are and the P0-P8 phase ladder.DEV_SETUP.md- get a working environment and run the tests.
Apache-2.0 (see LICENSE and NOTICE). The bundled HRTF grid is derived
from the SADIE II database (subject KU100), CC BY 4.0 - attribution in
NOTICE and echo/data/hrtf/sadie2_ku100/SOURCE.txt. This repository is
prepared for publication but not yet published.