droid-voice · engine 1.0 · 48 kHz mono · planner 1.1 · rendered 2026-09-08

Droid Voice Audition Sheet

Every clip on this page was rendered by the deterministic engine from a seed and a semantic request. Nothing is sampled. Speech comes from the generative planner and varies with the seed; interjections come from each voice's authored lexicon and are meant to repeat. Tap a play button; the timeline under each golden fixture shows the gesture score the renderer was given, with semitone offsets from the voice's tonal center.

Interjections repeat; speech varies

Each interjection is rendered three times with different seeds. It should sound like the same "word" each time. Below them, the same voice speaking the same intent with three seeds, which should sound like three different sentences.

Say, then interject

Mixed mode: a speech phrase that ends in an exclamation.

Said from text

Each line was handed to the LLM translator (openai/gpt-6-astra via OpenRouter), which chose only intent, emotion, intensity, mode and length; the engine then rendered it deterministically from the derived seed. The italic note is the model's own description of the performance.

Golden fixtures

The five regression cases from spec §34, at intensity 0.8.

Every intent, every voice

Neutral emotion, intensity 0.6, 500 ms. Pick a voice.

Every emotion on one phrase

The courier voice saying greet at intensity 0.8, 700 ms, so only the emotional prosody changes.