The second10⁰ s
Claude
It reads the cue list, interprets stage events, picks the next cue and prepares trajectories. It anticipates: when the sound is needed, the decision is already made.
Technology
What the room sees, the sound knows.
A complete chain, from camera to speaker, built on open standards. Every cue can combine what the cameras see, what the mics hear and where each sound sits: the combinations are endless. Briquemault is in development: this page describes its architecture.
The chain
Three inputs, two brains, and the venue’s hardware at the output.
Seconds · decisions
Cue list
show cues
Claude
cue-level decisions
Milliseconds · execution
Cameras
visible and infrared
Ceiling mics
steerable beam
Local perception
markerless tracking
Voice processing
isolation, feedback
Real-time engine
millisecond scale
Immersive system
ADM-OSC · L-ISA…
PosiStageNet
positions, sound and light
PTZ cameras
VISCA · NDI
Two clocks, never mixed up.
Sound rendering cannot wait for an AI. So Briquemault separates two time scales.
The millisecond10⁻³ s
The local engine
Position tracking, trajectory rendering, ADM-OSC and PosiStageNet commands, microphone control. It runs on site, independent of the network.
The five functions in detail
01See
Cameras that understand the stage.
Fixed or PTZ cameras, in infrared when it is dark, track performers with nothing to wear, locate the audience as anonymous positions and detect key moments: an object raised, a gesture, an entrance.
- Inputs
- Video from fixed or PTZ cameras, visible or infrared.
- Processing
- Markerless detection and tracking, pose estimation and stage events, on a local computer.
- Outputs
- Performer positions over PosiStageNet, anonymous audience positions, timestamped events.
CAM 01 · VISIBLESimulation
02Hear
The audience speaks, no microphone needed.
Briquemault locates in 3D the volunteer the camera designates and sends that point to a steerable ceiling mic array, which aims its beam at them. They speak from their seat, and the whole room hears them.
- Inputs
- Beamforming ceiling microphone arrays, steerable by command.
- Processing
- The volunteer’s 3D position, computed from the cameras and sent to the array as its aiming point, then voice isolation, noise reduction and feedback control, on site.
- Outputs
- Isolated voice, tied to the audience member’s position, ready for the immersive system.
Simulation
03Place
Every sound in its real place.
Briquemault drives your immersive sound system over ADM-OSC. An audience member’s voice comes from their seat, an effect crosses the room, the sound follows the performer into the audience.
- Inputs
- Positions, voices, effects and trajectories.
- Processing
- Parameters computed for every sound object: position, distance, elevation, width.
- Outputs
- ADM-OSC and OSC commands to the venue’s immersive system.
Simulation
04Understand
An AI that knows the cue list.
Claude, Anthropic’s AI, reads the show’s cue list and connects what the cameras see to the right sound cues. When the performer improvises, it adapts. The sound operator talks to it like to a stage manager.
- Inputs
- Show cue list, stage events, operator instructions.
- Processing
- Claude picks the cue, adapts to the unexpected and explains its choices.
- Outputs
- Cues and rules sent to the real-time engine, never the audio signal itself.
Simulation
05Direct
A control room that follows the action.
Lights follow the performers and PTZ cameras frame the action on their own. The same installation films the show from several angles.
- Inputs
- Tracked positions, cues, lighting and framing instructions.
- Processing
- Lights driven over DMX, Art-Net or sACN, and PTZ cameras over VISCA or NDI.
- Outputs
- Lights that follow the action, automatic framing and multi-angle capture.
REC · AUTO FRAMINGSimulation
Hardware
Briquemault relies on off-the-shelf hardware. A venue survey settles the rest.
- 01Fixed or PTZ cameras, with infrared illumination
- 02A local computer with GPU acceleration
- 03Beamforming ceiling microphone arrays, steerable by command
- 04A sound system, ideally object-based immersive
- 05The venue’s fixtures and lighting console, over DMX, Art-Net or sACN
- 06Optionally, ultra-directional speakers
- 07A local network dedicated to the show
From venue to show
- 1
Venue survey
Floor plan, existing hardware, acoustics and capacity.
- 2
Calibration
Cameras, microphones and speakers share one frame of reference, in metres.
- 3
The cue list
We import the show and its cues; Claude prepares the logic.
- 4
Rehearsals
The operator adjusts in plain language: “slower”, “closer to the front row”.
- 5
The show
The local engine executes, the operator stays in control.
What Claude does, and what it does not
It does
- Read and understand the cue list
- Connect a stage event to the right cue
- Adapt when the show goes off script
- Answer the operator in plain language
It does not
- Touch the audio signal
- Decide at millisecond scale
- See camera images: it only receives anonymous events
- Identify anyone in the audience
- Stop the operator from taking over
RGPD
Privacy by design
- Images processed on site, no video sent out
- Audience reduced to anonymous positions
- No facial recognition
- No audience recording by default
- Audience informed at the entrance
Claude, wired to the cue list
Briquemault is built on Anthropic’s Claude API. Claude never produces sound: it reasons about the show and acts through tools that the real-time engine controls.
The cue list in memory
The cue list, the venue map and the operator’s rules are loaded once and cached: every decision starts from the whole show.
Tools, not text
Claude acts through typed tool calls: arm a cue, prepare a trajectory, designate a seat, write to the log. The engine checks each call before running it.
Three autonomy levels
Suggestion, operator approval, or automatic on authorised cues. The venue decides, act by act.
Decisions explained
Every choice is written to the show log with its reason, ready to review after the performance.
Plain-language direction
"Slower next time" becomes a rule in the cue list, applied from the next performance.
Fallback without AI
With no network or no answer in time, the local engine follows the planned cue list. The show never depends on a request.
- SEEobject_raised · ART-01 · confidence 0.94
- CLAUDEarm_cue(16, trigger: "object_raised ART-01", trajectory: "T16 → R8·S14")
- ENGINEcue 16 armed · level: automatic on authorised cue
- ENGINEfired on the event · trajectory 2.2 s
- CLAUDElog: "card raised, vanish sent towards R8·S14"