Contextual quick reactions, not a phrase bank

The existing local Nemotron Bolt E4B races the main Qwen/Cerebras answer to produce at most six empathetic words. The full answer continues independently. The main answer wins ties; invalid or late reactions are skipped. Same reference voice, direct handoff.

Measured audio latency

Three paired real API/TTS trials on a shared dual-RTX-3090 host. PCM arrival at the HTTP client, excluding physical playback. Different stochastic answers: these are small observational measurements, not a controlled quality study.

A faster first sound delays the substantive answer by the time needed for the reaction. This is optional. There is no claim of perfectly seamless voice continuity.

Text-generation experiment

16 actual local/API races; local generation 132–402 ms, median 256 ms. Local was ready first in 13/16 cases. Outputs remain imperfect: factual questions can receive generic acknowledgments. Three longer transition probes added a separate call and redundant speech; direct handoff was selected.

tired: {'text': 'Das stimmt. Wirklich sehr anstrengend. Das ist klar.', 'ms': 369.47, 'words': 8}

meaning: {'text': 'Das ist tiefgründig. Manchmal ist es einfach. So geht es weiter.', 'ms': 238.17, 'words': 11}

joke: {'text': 'Das ist ein lustiger Witz. Er ist kurz. Genau richtig.', 'ms': 450.63, 'words': 10}

Can the reaction start below 900 ms?

8 measured reaction onsets: 707–1311 ms, median 745 ms; 7/8 below 900 ms. These HTTP diagnostics deliberately cancel after the first PCM block and do not evaluate the handoff. The full browser trials below include playback scheduling. An earlier post-restart trial took 1311 ms: this is not an unconditional latency guarantee.

What changed after testing

Initial alignment held back short reactions. Removing that lookahead improved first sound, but a mistakenly long duration condition stretched four words to nearly seven seconds. The final reaction uses a 2–2.4 s duration condition, a 160 ms initial block, then 400 ms blocks and 10 ms fades; main answers retain 1600 ms chunks and existing alignment. A malformed optional acting cue also surfaced; it now falls back to a neutral cue without discarding the answer. Internal server-to-server PCM eliminates MP3 encoding and download round trips. Single live replies skip ranking and final-file export. Same-seed 2.48-second speech took 2862 ms with uniform 160-ms blocks versus 1975 ms with a 160-ms first / 400-ms following schedule. Codec waveforms have block-size-dependent numerical differences; perceptual equivalence is not proven. Earlier measurements and failures are retained.

Browser verification

{
  "passed": true,
  "error": null,
  "scope": "Actual HTTPS browser, typed synthetic turns, real local/API/TTS. No physical microphone or acoustic-stop assertion.",
  "layout": {
    "all_static_controls_retained": true,
    "boxes": {
      "avatar": {
        "x": 140,
        "y": 119.3125,
        "width": 470,
        "height": 385
      },
      "conversation": {
        "x": 699,
        "y": 119.3125,
        "width": 584,
        "height": 346.640625
      },
      "latency": {
        "x": 83,
        "y": 662.3125,
        "width": 1200,
        "height": 156
      },
      "models": {
        "x": 689,
        "y": 1077.28125,
        "width": 598,
        "height": 51.09375
      }
    },
    "camera_requests": 0
  },
  "mobile_controls_accessible": true,
  "off_setting_persists": true,
  "source_unchanged": true
}

Heute war ein wirklich anstrengender Tag. Erzähl mir etwas Aufmunterndes. — {'passed': True, 'first_pcm_ms': 865.4, 'scheduled_playback_ms': 1019.91, 'new_underflows': 0, 'max_gap_ms': 0}

Was ist der Sinn des Lebens? Erzähl mir deine Gedanken dazu. — {'passed': True, 'first_pcm_ms': 979.2, 'scheduled_playback_ms': 1133.71, 'new_underflows': 2, 'max_gap_ms': 538.19}

Schreib ein Gedicht mit vier Zeilen über einen kleinen Roboter. — {'passed': True, 'first_pcm_ms': 1862.4, 'scheduled_playback_ms': 2520.59, 'new_underflows': 0, 'max_gap_ms': 0}

Wie entsteht ein Regenbogen? — {'passed': True, 'first_pcm_ms': 1509.2, 'scheduled_playback_ms': 2167.39, 'new_underflows': 0, 'max_gap_ms': 0}

Playback scheduling gaps still occurred in one of the latest two reaction turns. The earlier start does not establish a gap-free transition on this shared host. Device acoustics were not measured.

Listen to authored test outputs

Only synthetic prompts and generated speech are included. No private microphone audio.

Reproduce and inspect

Setup, design and test commands · Text trial receipt · Audio timing receipt · HTTPS browser receipt.

Manual cancellation is tested separately from spoken interruption. Previously documented noisy acoustic Stop failures are not certified fixed by this change.