vs a measured HRTF
A dummy head, or one listener measured in a lab: one head, captured once, at one distance, and played to everyone. BSE computes the cues for each head, at any distance, and every one of them stays adjustable while it plays.
Research · Spatialization algorithm
Binaural Sound Engine
Binaural, made to measure.
BSE models every mechanism of spatial hearing as its own set of parametric filters: the delay and the shadow of the head, the notches of the pinna, the shoulder, the floor. Each one is a value you can set, so the render fits the listener’s own head and the product’s own voicing.
The result holds up against the best of the market, with a few samples of latency and a fraction of a core. Do not take our word for it: listen.
Drag the source · the cues follow it
Listen
The binaural engine from RIPL, compiled for your browser: the head model that places the sound, and the room it plays in. Place a mono sound or a stereo track around your head, with our examples or with up to 90 seconds of your own audio.
Put on headphones for a better experience. On speakers the two ears mix together and the effect collapses.
The problem
Binaural reproduces, on two headphone channels, what reaches your two eardrums from a sound out in the world. The brain places that sound from a handful of cues: the time difference between the ears, the level difference the head’s shadow creates, and the spectral fingerprint the outer ear, the shoulders and the torso stamp on everything that reaches the drum.
Get those cues right and the sound stands outside your head, in front, behind, above. They are usually captured as an HRTF: a set of filters measured for every direction around one head. There are two ways to get one, and neither is satisfying.
Further down the market, generic models play one dummy head, or an average of many, to every listener. We set BSE against the two approaches above, not against those.
The third way
BSE does not borrow a head. It models your hearing, cue by cue.
Every psychoacoustic cue is its own parameter: the delay and the shadow of the head, the notches of the pinna, the reflections of the shoulders and the floor. Two measurements set the geometry, then each cue can be adjusted until the render matches how that person really hears.
| Measured in a lab | Matched from a scan | BSE | |
|---|---|---|---|
| Whose cues you hear | yours | the closest head in a database | yours, computed |
| What it takes | an anechoic room and an hour or more | a scan and a database | two measurements |
| Adjustable afterwards | no, a new session | no, another entry | every cue, while it plays |
| Distance | the one it was measured at | the one it was measured at | continuous, 0.3 to 6 m |
How it works
Made to measure
Two measurements set the geometry. Every other cue stays a parameter, so the render can match one listener exactly, or be voiced once for a whole product line.
What the model makes possible not built yet
The best personalised solutions already capture the geometry of the head and ears with a phone, then use it to look up the closest head in a database of measured HRTFs. Because BSE is parametric, the same scan could do more: the 3D model itself could set the psychoacoustic parameters, head width, ear height and the shape of the pinna, for a render made to measure with nothing borrowed. This is a direction the model opens, not a product today.
Measured
Other models start from an existing head. A dummy head is built to be a good microphone: symmetrical, stable, repeatable. It does not carry all the subtleties of a real human head, and every listener inherits its shape.
BSE starts from the scientific literature on spatial hearing, and reproduces each psychoacoustic cue as it was evaluated on real listeners. Because every cue is a parameter, a product can then choose to accentuate some of them: a clearer front, a deeper pinna notch, a more open rear.
Real listeners: 60 people from the ARI HRTF database (Acoustics Research Institute, Austrian Academy of Sciences, CC BY-SA 3.0; Majdak, Goupell and Laback, 2010), analysed with the same code as BSE and the KU 100. Across the four bands, BSE sits 1.2 dB from the average listener, the KU 100 1.4 dB.
Under the hood
No FFT, no impulse response, no block. Each cue is a set of filters: a complex response built from simple elements, arranged methodically to match the literature. The cost per source stays small, and the answer to a head turn arrives in samples, inside the window where the brain fuses what it hears with how it moved.
A single core renders dozens of sources, and the engine reaches platforms convolution cannot: earbuds with onboard DSP, automotive head units, phones that still have to last the day.
Compared to
A dummy head, or one listener measured in a lab: one head, captured once, at one distance, and played to everyone. BSE computes the cues for each head, at any distance, and every one of them stays adjustable while it plays.
Personalised solutions scan your head and hand you the closest entry in their database: somebody else’s response. BSE takes your dimensions and sets the cues from them directly.
Delays and phase tricks make a mix sound a little wider, and fall apart in mono. BSE recreates what really happens when a sound stands somewhere around you: delay, shadow, pinna, room.
An FFT per source, blocks of latency, and a load many embedded DSPs cannot carry. Each cue in BSE is a set of filters built from simple elements, arranged to match the literature: a few samples of latency and a fraction of a core.
Applications
At a glance
| Personalization | Head width and ear height, from a tape measure, applied live |
| Parameters | 57 scalar parameters across 36 controls; the online demo exposes head width |
| Input | Mono per source, with no practical limit on the source count |
| Output | Binaural stereo |
| Latency | A few samples, at any sample rate |
| Implementation | Parametric filter sets built from simple elements: no FFT, no convolution, no block processing |
| Control | Azimuth, elevation, distance · yaw, pitch, roll · room amount |
| Formats | 44.1 to 192 kHz · 16, 24 and 32-bit float |
| Availability | Ships in RIPL · OEM licensing |
Integration
| Delivery | C · C++ · MATLAB · .dsp, full source code |
| Platforms | macOS · Windows · Linux · iOS · Android · embedded ARM |
| DSP platforms | Flow DSP · Audio Weaver, in preparation |
| Documentation | Technical documentation under NDA |
OEM licensing