No observation yet.
4 · SpeakingSilent
No speech yet.
Waiting for audio.
Test your speaker without making an AI request.
REHEARSAL · simulated observations, live writing & voice
Ready to begin
Choose Begin for a visitor, or Test without camera to try the writing and voice.
Shape the encounter
Choose what the camera notices, what the character says, and how it sounds. Your settings are separate from other visitors. Changes save automatically and flow into new work. Queued clips finish with their existing settings.
Loading…
Speaker Character
Define who is speaking, then tune the tone. These instructions guide the writer; they do not change what the camera reports.
Tone
0 = little influence · 100 = strong influence. Presets change the sliders only.
Observation
Choose what the camera reports to the writer. A shape is the set of fields and instructions used to analyse each image.
Loading observation settings…
Edit the shape, compare models on the same image, then apply your selection to the encounter.
Image timing & detail
Waiting for timing measurements.
Measured from analysis starting to the first spoken line that uses it, including writing, voice generation, queueing and pauses. Smart uses the 90th percentile of the last 10 samples plus the margin (maximum wait: 60s). While the room is empty, smart checks again within 1s. Longer waits reduce image requests but slow reactions to people leaving.
Only one image is analysed at a time. Both delay modes wait after a response arrives. The analysis timeout excludes this intentional wait. Images are sent to OpenRouter and are not saved by this app.
Writing
Steer what is said and how each line continues the story. Length, bridge and hesitation controls affect future writing.
Greetings & farewells
Greeting
Save a varied greeting pool for the selected voice and delivery. Makes paid speech requests.
Farewell
Save this farewell’s audio for reuse, in your hosted account. Changing its text, voice or delivery needs a matching cached clip. Makes one paid speech request if not already cached.
Checking farewell cache…
Save a varied farewell pool for the selected voice and delivery. Once all 50 are ready, departures use shuffled messages without repeats until the pool is used. The farewell above is the fallback. Caching makes up to 50 paid speech requests and resumes from saved clips. Keep this page open while caching; it resumes when you return.
Checking farewell pool…
Voice
Turn the written line into speech. Voice changes flow into new synthesis; prepared audio finishes without resetting the queue.
Cached greetings & farewells by voice
Pools and fallback clips belong to an exact model, voice, delivery and text. A full pool takes priority. Until a new voice is ready, an already prepared or cached voice can respond immediately. The 50-message pool controls are in Writing.
Checking cached voices…
Pauses
1000 ms = 1 second. Both modes use this minimum and maximum. In “Every N” mode, 0 disables pauses.
One extra LLM call per generated line, using the writing model. It judges whether to pause after that exact line finishes, while voice generation continues. If the decision is late or fails, playback continues without a pause. Greetings, farewells and Skip bypass pauses.
Waiting for a queued line to assess.
Timing & buffering
Balance smooth delivery with reacting to changes in the room. Observation timing is in Observation; dramatic pauses are in Voice.
More prepared lines reduce gaps but take longer to react. Group-size changes may replace future lines using the policy above. When everyone leaves, the current clip finishes and the farewell follows.
Advanced writer instructions. Character identity lives in Speaker Character; story direction lives in Writing. These sections control continuity, output rules and delivery. The live observation and conversation are added automatically.
Placeholders in double braces follow your controls. You can leave an instruction section blank to omit it.
Preview assembled instructions
Uses the text and control values currently in the editor. Preview makes no model calls. Numbered in prompt order: 1 left, 2 right, 3 below 1, and so on.
Pipeline inspector
Follow an image through analysis, writing, speech generation and playback.
Settings changes flow into new work. Prepared speech keeps playing; scene changes can still replace it.
Timers include network and decoding. Discarded lines remain in history.
Vision—
Writing—
Speech—
First voice—
1 · Observe
Camera image + previous reading → structured observation.
StoppedNo work yet.
No completed request yet.Camera offWaiting for a frame—Image width: 640 PX
2 · Write
Observation + character + dialogue history → next line.
StoppedNo work yet.
No completed request yet.3 · Voice
Written line + delivery settings → playable audio.
StoppedNo work yet.
No completed request yet.4 · Play
Prepared audio → speaker, with pauses and skip.
StoppedNo work yet.
No completed request yet.Buffer empty · no audio scheduled
Line history · including discarded speech
0 lines tracked
No generated lines yet.
Event log
What the image model received
The exact instructions and comparison context from the latest completed analysis. The latest accepted result is shown in the observation pane above.
Begin with a camera to inspect an image-analysis request. Test without camera uses a simulated observation.
Writer context
The complete latest request, as sent to the writer.
Waiting for a request
Begin an encounter or rehearsal to see the instructions and context sent with each writing request.
Request history & raw messages
Complete requests, including retries, newest first. JSON context is expanded for reading. Latest 30 requests retained per encounter.
No writer requests yet.
Future clips in playback order. Voice delivery details show any annotations passed to speech generation. Use Skip clip in the speaking pane to advance.
Queue0 queued
Nothing queued.
Transcript5 MIN MEMORY
No speech yet.