Open live demo · Implementation and reproduction · Test protocol
Say “Hey Robby” and your request. Your voice can continue without another wake phrase for ten seconds; the window is renewed after Robby finishes an answer. A different or uncertain voice needs its own wake phrase. Recognition is approximate; it is not authentication.
Speech-pause default: 300 ms, adjustable. That is not total response latency and may split a sentence. Automatic sound reactions default off. Explicit sound questions still use the short RAM buffer and an after-question window; captions can be wrong.
582 offline tests passed. The latest HTTPS digital-audio browser run passed. It uses authored synthetic voices, actual Parakeet, ECAPA and TTS with PCM playback; LLM: local Nemotron Bolt E4B. This is not a physical-room accuracy evaluation.
| Case | Result | Recognized / captioned | Spoken text | Ignored reason |
|---|---|---|---|---|
| before-wake | PASS | [] | [] | [{'source': 'wake_mode', 'reason': 'Say Hey Robby to start, or speak clearly enough to identify your voice.'}] |
| christoph-wake | PASS | [{'text': 'Hey, Robby, ich bin Christoph.', 'speaker': 'Christoph'}] | ['Hallo Christoph! Ich bin Robby, dein kleiner, warmer Roboterfreund.\nWas für ein toller Name, Christoph! Sollen wir heute etwas Lustiges erleben?'] | [] |
| owner-followup | PASS | [{'text': 'Und wer bin ich?', 'speaker': 'Christoph'}] | ['Du heißt Christoph.'] | [] |
| foreign-no-wake | PASS | [] | [] | [{'source': 'wake_mode', 'reason': 'Another or uncertain voice; say Hey Robby to start your own turn.'}] |
| expired-owner | PASS | [] | [] | [{'source': 'wake_mode', 'reason': 'Say Hey Robby to start, or speak clearly enough to identify your voice.'}] |
| second-speaker-wake | PASS | [{'text': 'Hey Rabbi, ich bin Selma.', 'speaker': 'Selma'}, {'text': 'Schön, dass du da bist.', 'speaker': 'Selma'}] | ['Es freut mich sehr, dass du da bist, Selma!\nLass uns gemeinsam ein imaginäres Abenteuer erleben, wie zum Beispiel eine Weltreise mit einem fliegenden Kätzchen!'] | [] |
| previous-owner-without-wake | PASS | [] | [] | [{'source': 'wake_mode', 'reason': 'Another or uncertain voice; say Hey Robby to start your own turn.'}] |
| open-mode | PASS | [{'text': 'Und wer bin ich?', 'speaker': 'Speaker 3'}] | ['Ich bin mir nicht sicher. Wie heißt du?'] | [] · speaker label correct: False |
| asked-sound-automatic-off | PASS | [{'kind': 'voice', 'caption': 'The audio contains a single utterance of speech. The speaker is perceived as male, with a medium pitch.', 'requested': True}, {'kind': 'music', 'caption': 'The audio features a fast-paced electronic track with a prominent, repetitive synth melody. A driving beat is present, with a strong bassline and a simple, repetitive drum pattern.', 'requested': True}] | Das ist Musik, die schnell und elektronisch klingt, mit einem treibenden Beat und einem starken Bass. Klingt das nach etwas, das dir gefällt? | [] |
Source fingerprint: f2080cbee4b13375eb89d8187fa3e4053bf0b2e1790cce8dda0c69231c502e29.
These tests use the same wake-enrolled synthetic voice, an authored music fixture and the real captioner. LLM: local Nemotron Bolt E4B. Before and after windows are evaluated independently; results are uncertain and captions can mistake speech for sound.
| Sound position | Result | Sound captions | Spoken response |
|---|---|---|---|
| after | PASS | [{'kind': 'sound', 'caption': 'The audio contains a single, distinct sound event. It is a short, sharp, and relatively high-pitched yelp or yelp.'}, {'kind': 'music', 'caption': 'The audio features a short musical phrase played on what sounds like a synthesized keyboard or electronic keyboard. The melody is simple and repetitive, with a slightly melancholic tone.'}] | Ich glaube, es klingt nach Musik Ein bisschen melancholisch, wie ein elektronisches Klavier Möchtest du, dass ich mehr darüber finde? |
| before | PASS | [{'kind': 'music', 'caption': 'The audio features a fast-paced electronic track with a prominent, repetitive synth melody. A driving beat is present, with a strong bassline and a consistent rhythmic pulse.'}, {'kind': 'sound', 'caption': 'The audio contains a single, short, and sharp sound. It is a high-pitched, short burst of noise.'}] | Ich höre eine Musik, Christoph Es klingt nach einem schnellen elektronischen Track Ein treibender Beat mit einer starken Basslinie, findest du nicht? |
Latest two-case sound test passed. The configured Qwen provider returned HTTP 429 during repeated tests; the test can explicitly select the local model without changing the demo default.
Earlier attempts are preserved below. A 300-ms pause split an introduction into two answers, and an earlier version let the ten-second window expire while Robby was still replying. The current version permits the same voice during playback and renews the window after playback ends. Earlier tests also exposed a sound buffer that was too short and a voice embedding distorted by preceding music; the bounded buffer and tail-window verification were revised. Compare source fingerprints rather than treating earlier failures as current results.
The second-speaker introduction may also contain an ignored short continuation even when its first wake turn succeeds. Open mode can admit the returning speaker under an unnamed label; a direct identity question now asks for the name instead of guessing a previous person's name. A case labelled PASS for mode switching does not prove every speaker label or every fragment was correct. These few clips do not establish performance with real users, simultaneous voices, a television, echoed playback, short Stop commands or similar voices. False accepts and missed matches remain possible; the mode is not a biometric security boundary. If recognition is uncertain, repeat the wake phrase or disable the mode.