Add Hey-Ozwell on-device voice stack — wake word, speaker verify, voice enrollment, dictation, hands-free chat (#288) * AIChat: on-device voice (Whisper) story Wires the existing talkToText/onRecordingComplete seam to a browser-resident Whisper model so the mic actually transcribes — on-device, audio never leaves the page. Additive: a new *.stories.tsx only; the shipped AIChat is unchanged. Co-Authored-By: Claude Opus 4.8 * AIChat voice: drop language/task opts (whisper-small.en is English-only) Co-Authored-By: Claude Opus 4.8 * AIChat voice: quantized model + persistent storage (fixes re-download on every reload) Default fp32 whisper-small exceeded the browser's cache quota (~300MB) so it never cached and re-downloaded each reload. Switch to base.en q8 (~4x smaller, fits) and request persistent storage so reloads are instant after the first load. Co-Authored-By: Claude Opus 4.8 * AIChat voice: turbo default + base.en fallback (multilingual-aware) large-v3-turbo (fp16/q4) on WebGPU for best accuracy, falling back to base.en q8 when WebGPU is unavailable. transcribe() now passes language/task only for the multilingual turbo. Caching of turbo needs adequate browser storage quota. Co-Authored-By: Claude Opus 4.8 * WakeWord: Phase 1 on-device detection component (hook + story + assets) Vendors the framework-agnostic HeyBuddy detector under WakeWord/lib, adds a headless useWakeWord() hook and a Listener story that fires on 'hey ozwell' / 'ozwell i'm done' with a live speech/probability readout. Detection front-end + wake models hosted in .storybook/public/wakeword. Needs onnxruntime-web + the VAD model (see story header). Phase 2 adds the doctor-only speaker gate. Co-Authored-By: Claude Opus 4.8 * WakeWord: drop webpack-only onnxruntime fallback import (Vite can't resolve it) Co-Authored-By: Claude Opus 4.8 * WakeWord: load onnxruntime-web WASM from CDN + single-threaded (fixes 'no backend') Co-Authored-By: Claude Opus 4.8 * WakeWord: add silero-vad v4 model (matches vad.js input/h/c/sr interface) Co-Authored-By: Claude Opus 4.8 * WakeWord: run detection regardless of VAD gate (public VAD reads low) + polish UI Decouples wake detection from the substituted silero-vad (which under-triggers isSpeaking and was suppressing all detections). Restyles the Listener story as a clean card with a pulsing status, gradient meters, and detection chips. Co-Authored-By: Claude Opus 4.8 * WakeWord: re-entrancy guard on wake inference + low VAD threshold Guard prevents the 'Session already started' overlap crash, which lets us gate generously (low positive/negative VAD thresholds) so the low-reading VAD still lets speech through to detection. Co-Authored-By: Claude Opus 4.8 * WakeWord: load onnxruntime-web@1.19.0 as global ort (match the working demo exactly) The demo pins ort 1.19.0 via a script tag; the newer npm version was slower and behaved differently (choppy inference, no detection). Inject the same 1.19.0 build as a global so the detector's onnx.js uses it (its preferred path). Co-Authored-By: Claude Opus 4.8 * WakeWord: add raw mic-level diagnostic story (isolate capture vs detector pipeline) Co-Authored-By: Claude Opus 4.8 * WakeWord: drop redundant permission getUserMedia (two mic streams → detector silent) The mic-test diagnostic confirmed capture works with ONE stream. The hook was opening a second getUserMedia for permission while the detector opened its own; two concurrent streams + AEC/AGC made the detector's stream near-silent (VAD ~1%). Let the detector own the single stream. Co-Authored-By: Claude Opus 4.8 * Hands-Free Chat: compose wake-word + on-device dictation + AIChat (one shared mic) New composition story: 'hey ozwell' starts a MediaRecorder on the wake detector's OWN stream (useWakeWord().getStream() — no second getUserMedia), 'ozwell i'm done' stops + transcribes on-device + posts to AIChat. Extracts whisperTranscribe.ts as a shared util. Primitives stay independent; this is the thin composition layer. Co-Authored-By: Claude Opus 4.8 * Hands-Free Chat: fix stop-phrase leak (trim audio tail + fuller strip) + manual dictate button - transcribeBlob trims the last ~0.8s so Whisper never hears 'ozwell i'm done'; stripStopPhrase restored to the fuller demo regex (catches bare Also/Ozwell/Oswald mishearings). - Adds a manual 'Dictate' button in the status bar that uses the SAME shared mic stream (AIChat's built-in talkToText mic can't be used — it opens a second getUserMedia that goes silent). Co-Authored-By: Claude Opus 4.8 * Hands-Free Chat: use the composer's real mic button (controlled) on the shared stream Replace the separate Dictate button with the actual RecordButton in the composer's trailing slot, in CONTROLLED mode (internal recorder disabled) + onClickCapture driving our shared-stream dictation. Now the built-in mic and 'hey ozwell' do the same thing — either works, one mic, no second-getUserMedia conflict. Co-Authored-By: Claude Opus 4.8 * SpeakerVerify: on-device WHO gate primitive (useSpeakerVerify + enroll/verify story) Vendors the TitaNet speaker-verification runtime (speaker-verify.js → window.SpeakerVerify) under SpeakerVerify/lib, adds a useSpeakerVerify hook and a story to enroll a voice and verify against it (cosine + z-score, on-device). Hosts the ~50MB sherpa WASM/data + the 500-voice cohort in .storybook/public/sv-runtime. Separate primitive; folds into the hands-free composition next (doctor-only). NOTE: ~50MB in git — use LFS/CDN for the real PR. Co-Authored-By: Claude Opus 4.8 * SpeakerVerify: concrete guided enrollment (exact phrase, chime cue, silence retry, confirm chime, progress) Mirrors the product enrollment UX per Doug's asks: shows the exact phrase, chimes 660Hz to cue when to speak, records, checks it wasn't silent (retries that rep), 990Hz confirm chime, per-rep progress, both phrases. One mic stream opened per session (no repeated getUserMedia). Co-Authored-By: Claude Opus 4.8 * SpeakerVerify: phrase-validated enrollment via the wake model (deny if not the phrase) useWakeWord now exposes onUtterance(name, samples) — the captured wake audio + which phrase fired. Enrollment/verify run the detector and only accept a rep when the wake fires for the EXPECTED phrase; wrong phrase or no detection is denied and retried. The model-captured audio builds the voiceprint. Composes useWakeWord + useSpeakerVerify, all on-device. Co-Authored-By: Claude Opus 4.8 * SpeakerVerify: validate enrollment on onWake (not onRecording) — can't enroll junk onRecording fires on any speech segment, so junk got paired with a stale phrase and accepted. Switch validation to onWake (wake-GATED — only fires when the model classifies the phrase), record the audio off the shared stream during the wait. Wrong phrase / no fire = denied + retry. Co-Authored-By: Claude Opus 4.8 * SpeakerVerify: show the refusal prominently (red card 'not counted, try again') + 5s timeout The deny was happening but only flashed a tiny status line while the big card still said 'say it now'. Now denials render as a red card ('❌ … — not counted · try again'), held 1.6s, and the wait is 5s so junk is refused promptly. Co-Authored-By: Claude Opus 4.8 * SpeakerVerify: shorter refuse-timeout (4s) + pulsing 'listening…' so the wait feels active Co-Authored-By: Claude Opus 4.8 * Two-gate + multi-condition: WHAT phrase voiceprint + WHO speaker + Add condition useWakeWord exposes the WHAT gate (getLastEmbedding/setVoiceprint/hasVoiceprint/phraseCosine over the wake model's fire-frame embedding). The enroll/verify story now: one phrase-validated enroll builds BOTH prints (WHO via TitaNet centroid + WHAT via phrase voiceprint), 'Add condition' appends another room (both gates match best-over-conditions), and verify requires BOTH the right voice AND the right phrase. Matches the product's two-gate model. All on-device. Co-Authored-By: Claude Opus 4.8 * SpeakerVerify diagnostic: fluid live gate readout (enroll once, then talk) Replaces the one-shot Verify button with the product-style continuous readout: after enrolling, just talk — every wake shows WHICH phrase + base/WHO/WHAT, each green(pass)/red(fail)/grey(off), with the verdict. One rolling recorder (2nd consumer of the shared stream) feeds both enroll + live scoring. Relabeled the story 'dev diagnostic' since verification is invisible in the product. Co-Authored-By: Claude Opus 4.8 * SpeakerVerify: fix 'base' — verdict is WHO+WHAT, wake is just the trigger (frozen, not stale) The readout was treating the wake probability as a third pass/fail gate AND reading it stale (prob spikes at the fire then drops), so real fires showed base 0.04 and got rejected even with WHO+WHAT passing. Now: freeze the peak fire confidence at the fire (hey-buddy lastWakeProb), expose it via useWakeWord.getLastProb(), show it as 'wake' (always green — a row only exists because it fired), and base the verdict on WHO+WHAT only. Co-Authored-By: Claude Opus 4.8 * WHAT gate: use common-mode-subtracted similarity (raw cosine pinned ~0.9, couldn't separate) phraseCosine was raw cosine, which the shared common-mode pins near 0.9 for everything — so WHAT passed all fires. Switch to hey-buddy's voiceprintSimilarity (subtracts the running embeddingMean so the score reflects CONTENT), and drop WHAT_THRESHOLD to 0.3 (subtracted range is lower; tune from the live readout). WHAT is still the fragile gate by nature, but now it actually discriminates. Co-Authored-By: Claude Opus 4.8 * SpeakerVerify: instrument WHAT capture (console + 'WHAT prints stored' line) to find the null Co-Authored-By: Claude Opus 4.8 * WHAT gate: back to raw cosine + 0.80 threshold (product value, consistent for real phrases) The subtracted similarity scored real phrases erratically (0.40-0.89), so no clean threshold. Raw cosine lands real phrases ~0.85-0.94 consistently, so 0.80 passes them cleanly and rejects unrelated speech — matching the product the demo used. (Still can't separate 'ozwell X' variants; that's the fire-frame-on-shared-word limit — WHO + a better model own that.) Co-Authored-By: Claude Opus 4.8 * SpeakerVerify diagnostic: show the WHO z-score in the live readout (shadow — informational, not gating) Co-Authored-By: Claude Opus 4.8 * WakeWord: raise hey-ozwell threshold 0.5 → 0.8 (product value) — junk fired ~0.55 and enrolled The enrollment check (require the wake to fire for the phrase) was fine, but hey-ozwell's wake threshold was left at 0.5 from detection bring-up, so low-confidence misfires ('yeah bro' ~0.55) counted as the phrase. Back to 0.8 like the product; done stays 0.5 (longer, harder to false-fire). Co-Authored-By: Claude Opus 4.8 * Voice Setup: Apple-style enrollment page with a reactive glowing orb Polished client-facing enrollment (vs the dev diagnostic): a glowing orb that scales/pulses with your voice, guided phrase prompts, chime cue, per-rep progress dots, graceful retry, finish state. One pass builds both prints (WHO+WHAT), phrase-validated, persisted. All on-device. Co-Authored-By: Claude Opus 4.8 * Voice Setup: Ozwell octopus instead of generic orb — tap to start, pulses as you talk Brand-aligned light design around the actual Ozwell octopus (/ozwell/icon.svg, brand blue #0BA0E0): tap the octopus to begin, it glows + scales with your voice, bounces ✓ on each success, shakes on a miss. Floating idle animation, blue pulse rings + glow, progress dots. Same enroll logic underneath. Co-Authored-By: Claude Opus 4.8 * Voice Setup: fix off-center pulse ring (backwards fill-mode) + softer privacy line Co-Authored-By: Claude Opus 4.8 * Voice Setup: drop redundant bottom privacy line, fold 'privately on your device' into intro copy Co-Authored-By: Claude Opus 4.8 * Hands-Free Chat: fold in the invisible doctor-only gate (WHO + WHAT on each wake) On every wake the chat now silently verifies the speaker (TitaNet WHO) + phrase voiceprint (WHAT) using the enrolled prints — only the enrolled doctor saying the phrase starts/stops dictation; everyone else is ignored (logged, not shown). Restores WHAT prints + opens a rolling recorder (a 3rd consumer of the one shared stream) for the wake-utterance audio. Open if not enrolled yet; fails closed if enrolled but can't capture. Banner shows '🔒 your voice only' when enrolled. Co-Authored-By: Claude Opus 4.8 * Hands-Free Chat: don't gate the STOP phrase (only START) — reliable stop fixes the leak Gating 'ozwell i'm done' made it miss stops (the mid-dictation utterance is mixed with the dictation tail → low WHO → fail), so the phrase got repeated and leaked into transcripts past the tail trim. START stays the doctor-only gate; STOP always stops (session is already the doctor's; stopping is harmless). Bumped tail trim 0.8s → 1.0s. Co-Authored-By: Claude Opus 4.8 * Hands-Free Chat: gate STOP on WHO-only (keeps it doctor-only, robust to run-on phrasing) + log gates Re-gate the stop, but on the speaker (WHO) only — the fragile WHAT failed because the run-on 'ozwell i'm done' at the dictation tail embeds unlike the enrolled isolated one. WHO is text- independent so it holds; a coworker still fails WHO. Logs WHO/WHAT per wake so we can see it. Co-Authored-By: Claude Opus 4.8 * Hands-Free Chat: both gates (WHO+WHAT) on stop too — WHAT is what catches the doctor's own false fires; log the numbers Co-Authored-By: Claude Opus 4.8 * Preload Whisper at app/enrollment open so first dictation isn't slow Add warmWhisper() to kick off the (memoized) model load eagerly, and call it on mount in Hands-Free Chat and the Voice Setup page. The dictation model now downloads/initializes in the background while the doctor enrolls or sits idle, instead of lazily on the first "ozwell I'm done". Since loadWhisper() returns the same cached promise, the later transcribeBlob awaits the same in-flight load — no double download, no race. Co-Authored-By: Claude Opus 4.8 * Make octopus "Add another spot" append a condition + load-lifecycle logging VoiceSetup: "Add another spot" now appends a new enrollment condition instead of overwriting. The append pass calls enroll({append:true}) so WHO adds a new centroid (the list is kept; verify = max cosine across them) and WHAT concatenates templates (max over them), capped at VP_CAP. First-time setup is still a fresh enroll. Restores persisted WHAT templates on mount so append works across reloads, and the append pass reads as "Add a spot". The multi-condition engine was intact — the new page just wasn't calling it. Logging: whisper now logs which model actually loaded (turbo/WebGPU vs base.en fallback) + load time and why it fell back; wake and speaker hooks log when their models are ready. Co-Authored-By: Claude Opus 4.8 * Loosen stop-phrase WHAT gate to 0.75 (start stays 0.80) Per-phrase WHAT thresholds: the run-on "…ozwell I'm done" at the dictation tail embeds slightly differently from the isolated enrolled phrase, landing a bit lower than the clean start phrase, so legit stops were being gated out. 0.75 catches them without admitting the doctor's own non-phrase speech. Log now prints the threshold it checked against. Co-Authored-By: Claude Opus 4.8 * Preload Whisper in AIChat (Voice) story too Warm the model on mount so it loads in the background while the story is open, instead of on the first mic tap — first transcription is fast. Uses this story's own local loadWhisper (separate memoized promise from the shared util). Co-Authored-By: Claude Opus 4.8 * Docs: voice model-hosting plan (PR-readiness for upstream merge) Records the ~57MB of model binaries currently committed under .storybook/public, why they can't ship in mieweb/ui, and the runtime-fetch cutover plan (host on HF/GitHub Release, parameterize asset base, git rm binaries, browser-test). The cutover is deferred until browser-testable; the local demo is unaffected. Co-Authored-By: Claude Opus 4.8 * Wire ozwellai-api into the voice chat stories (stub -> real streamed reply) Replace the canned "Heard: ..." stub in the AIChat voice + hands-free stories with a real OpenAI-compatible call to ozwellai-api, streamed token-by-token into one assistant message. New framework-agnostic helper ozwellChat.ts (plain fetch, no SDK dep) reads the key at runtime from window.__ozwell / localStorage (never committed) and falls back to the stub when no key is set, so the demo still works keyless. Multi-turn (passes history). Co-Authored-By: Claude Opus 4.8 (1M context) * Parameterize model asset base for off-repo hosting (model-hosting cutover, code half) Wake + speaker-verify loaders now resolve their asset base from an assetBase prop / window.__ozwellAssets global, defaulting to the current local /wakeword and /sv-runtime paths so the demo is unchanged. Flipping to a hosted URL (e.g. HuggingFace) becomes a one-line config. Adds the HuggingFace upload runbook; updates MODEL-HOSTING.md. Co-Authored-By: Claude Opus 4.8 (1M context) * Read asset/backend config across Storybook iframe frames (+ prefer localStorage) Storybook runs stories in an iframe, so a console-set window.__ozwell / window.__ozwellAssets lands on the parent frame and the story never sees it. The config readers now check current -> parent -> top (same-origin). localStorage remains the reliable path (shared across frames); doc updated to recommend it. Co-Authored-By: Claude Opus 4.8 (1M context) * Allow forcing the small base.en Whisper model (ozwellConfig.whisper) The turbo model is hundreds of MB and stalls when browser storage is tight/over-quota (it can't cache, so it re-downloads and crawls). Setting { whisper: 'base.en' } in ozwellConfig (or window.__ozwell) forces the ~75 MB English model, which downloads fast and runs on a cramped machine. Default behavior (turbo, base.en fallback) is unchanged. Co-Authored-By: Claude Opus 4.8 (1M context) * Support small.en / tiny.en Whisper override (balance turbo's load vs base.en's errors) Generalize the whisper override to accept small.en (best speed/accuracy balance), base.en, tiny.en, or turbo. turbo is heavy and slow to load on modest machines; base.en hallucinates on unclear audio. small.en (~250 MB) is the recommended middle ground. Co-Authored-By: Claude Opus 4.8 (1M context) * Log Whisper download progress (make the one-time turbo load visible, not an opaque freeze) Adds a progress_callback that logs each model file at 25% steps during download. Cached loads emit no progress events, so silence = a fast cache hit. Lets you see turbo actually loading (and confirm it finished + cached) instead of guessing whether it froze. Co-Authored-By: Claude Opus 4.8 (1M context) * Add localStorage fallback for the model asset base (ozwellAssetBase) Mirror the backend-config fix: read the asset base from localStorage['ozwellAssetBase'] too, the reliable path across the Storybook iframe. HF upload doc updated to recommend it. Co-Authored-By: Claude Opus 4.8 (1M context) * Move voice models off the repo to HuggingFace (model-hosting cutover complete) Remove the ~57 MB of bundled wake/speaker binaries from .storybook/public and default both loaders to the HuggingFace host (jlocala/ozwell-voice-assets), so a fresh clone fetches wake, speaker, and Whisper at runtime with no setup. Verified in-browser loading from HuggingFace (wake fires, enrollment works, both ready lines). Overrides (assetBase prop / __ozwellAssets / localStorage) still repoint to a local copy. Repo is now binary-free — unblocks a clean PR. Co-Authored-By: Claude Opus 4.8 (1M context) * Log Whisper file start/finish (HF omits content-length, so big files had no % progress) The big .onnx model files download in apparent silence because HuggingFace's redirect drops the size header, so transformers emits no numeric progress for them. Also log 'initiate' (fetching X…) and 'done' (✓ X) so the slow files announce themselves instead of looking frozen. Co-Authored-By: Claude Opus 4.8 (1M context) * Cache voice models across app opens via a service worker (fixes turbo 0-byte cache) transformers.js caches the Whisper turbo model as a ZERO-byte entry because HuggingFace omits content-length on its newer storage and the progress-stream drains the response body before it's cached — so every app open re-downloads ~200MB. (base.en sends content-length and caches fine, proving the mechanism works; turbo specifically doesn't.) Closing/reopening the app IS a reload, so this would hit production, not just dev. Add public/ozwell-model-sw.js: a service worker that, for model assets only, fetches the full response itself and serves a complete cached copy on later opens — independent of content-length. Caches turbo, wake, and speaker models; passes all other requests (app JS/HMR) straight through so it can't break the dev server. Registered idempotently from the Whisper + wake load paths (modelCache.ts). Co-Authored-By: Claude Opus 4.8 (1M context) * Fix SW caching of HuggingFace Xet model files (rebuild Response before cache.put) The turbo .onnx files never cached: HF Xet storage returns them as a redirected, no-content-length stream, which a raw cache.put() refuses to store (it silently failed in the swallowed catch). Now, for model assets, the SW reads the full body into a buffer and rebuilds a clean Response — which materializes the stream and strips the redirect flag — then caches that. Logs each cached file + any put failure. Bumps cache to v2 and drops the old (empty-for-turbo) cache on activate. Co-Authored-By: Claude Opus 4.8 (1M context) * Disable transformers.js internal cache (SW is the single layer; stop double-storing ~1GB) cache.put for the big turbo file failed with QuotaExceededError — the ~1GB model was being stored twice (transformers.js's own Cache-API storage + the SW cache). Set env.useBrowserCache=false so the service worker is the only cache, halving storage pressure. Transformers still fetches over the network (the SW intercepts and serves from its cache on later opens). Co-Authored-By: Claude Opus 4.8 (1M context) * Set explicit Content-Length when caching (fixes spurious QuotaExceeded on Xet files) Quota is fine (94 MB used / 10.8 GB) yet cache.put threw QuotaExceeded on the big turbo file. Cause: the rebuilt Response still declared no size (HF Xet sends none), and cache.put refuses an unknown-length body with a misleading QuotaExceeded even with GBs free. Set Content-Length from the buffer we already hold, and use minimal headers to avoid Xet header carryover. Cache bumped to v3. Co-Authored-By: Claude Opus 4.8 (1M context) * Load Whisper turbo from self-hosted LFS mirror (cacheable; onnx-community Xet copy isn't) Point the turbo pipeline at jlocala/whisper-large-v3-turbo-ozwell — a classic-LFS mirror with real size headers — instead of onnx-community/whisper-large-v3-turbo, which is on HF Xet storage that Chrome refuses to cache (re-downloaded ~1GB on every open). LFS files cache cleanly here (proven by the wake/speaker models), so the SW now persists turbo across opens. Co-Authored-By: Claude Opus 4.8 (1M context) * Revert turbo to onnx-community; the cache failure is managed-Chrome, not the model host Self-hosting turbo on jlocala made no difference — the SAME QuotaExceeded fires on jlocala's copy. The real pattern is FILE SIZE: small files cache (configs, wake models, 40 MB speaker), large ones (~250 MB+ turbo weights) fail cache.put regardless of source, with 10 GB free. That's not normal Chrome — it's a managed/enterprise Chrome storage cap (matches the 300 MB quota seen earlier). Point back at the known-complete onnx-community repo (the self-host mirror may be incomplete) and document that turbo caching depends on an unmanaged browser. Co-Authored-By: Claude Opus 4.8 (1M context) * Document the model-cache SW + the managed-Chrome large-entry cap finding Co-Authored-By: Claude Opus 4.8 (1M context) * Cache turbo via transformers.js native streaming cache (LFS host), not the buffered SW The QuotaExceeded was our own doing: we disabled transformers.js's cache and used a SW that buffers the whole ~1.2GB file in memory before cache.put — which chokes on a file that large. transformers.js STREAMS the model to the Cache API (how whisper-web caches this exact turbo model). Re-enable it (useBrowserCache=true), load turbo from the self-hosted LFS mirror (size headers, unlike onnx-community Xet which caches 0 bytes), and narrow the SW to only the wake/speaker files (which it handles fine) so it no longer intercepts — and choke on — the Whisper download. Co-Authored-By: Claude Opus 4.8 (1M context) * Load Whisper turbo from Cloudflare R2 (sends Content-Length → it actually caches) Point transformers.js at our R2 bucket (pub-…r2.dev) for the turbo model instead of HuggingFace. R2 sends a Content-Length header; HF Xet doesn't, which is what broke transformers.js's cache (0-byte entry → re-download every cold open). With the size header, transformers.js streams the model into the Cache API and it loads once per device. Audio still never leaves the browser — R2 only hosts the model file. Resets remoteHost to HF for the base.en fallback. Repoint via R2_HOST for a mieweb-owned bucket. Co-Authored-By: Claude Opus 4.8 (1M context) * Document final caching outcome: turbo cached via R2 (HF Xet was the blocker) Update MODEL-HOSTING.md to the resolved state: turbo is hosted on Cloudflare R2 (sends Content-Length), so transformers.js's streaming cache works — confirmed cold ~85s / cached ~10s. Records the root cause (HF Xet sends no Content-Length → 0-byte cache → re-download), what was ruled out, and honest caveats for the Doug architecture conversation (load-dependent variance, best-effort website cache eviction under disk pressure, the corrupted-dev-profile artifact, and server-side ASR as the uniform-reliability alternative). Co-Authored-By: Claude Opus 4.8 (1M context) * Fixing onnxruntime-web dependency * Address Copilot review: ollama auth, VAD sample rate, async test() catches, stale docs - ozwellChat: don't send `Authorization: Bearer ollama` (the 'ollama' sentinel isn't a real credential) - hey-buddy: pass the local `targetSampleRate` to SileroVAD (this.targetSampleRate wasn't set yet); add .catch() to the async model test() calls so a failed load surfaces as a warning, not an unhandled rejection - WakeWord.stories: correct stale header (onnxruntime-web is now a dep; models fetch from the hosted base) - modelCache: clarify the SW caches wake/speaker only — Whisper turbo caching is handled by transformers.js + R2 Co-Authored-By: Claude Opus 4.8 (1M context) * Hey Ozwell: reorg into AI/HeyOzwell/, #287 docs + Storybook controls, Copilot review fixes - Reorg: move WakeWord/, SpeakerVerify/, HandsFreeChat into src/components/AI/HeyOzwell/; nest the Storybook sidebar under AI/Hey Ozwell. Shared backend (ozwellChat/whisperTranscribe/ modelCache) stays in AI/. - #287: HeyOzwell overview (mdx) + Demo + HeyOzwellToggle — octopus header toggle with volume pulse and right-click/long-press "Ozwell settings" menu. - Controls: Demo (auto-dictate, close-on-done, transcription browser/server), Toggle playground, Wake Word live thresholds (wake + VAD), Speaker Verify WHO gates (cosine/z-score/AS-norm); enable the Code panel in preview. - useWakeWord: release the mic on disable/unmount (stop tracks + close AudioContext) and reset ready; live-tunable wake + VAD thresholds. - Hands-Free: octopus header replaces the navy status banner. - Address Copilot review: audio.js resampler clamp (inputFrameSize), onnx.js init-failure surfacing, modelCache per-URL memoization, AIChatVoice uses the shared whisperTranscribe, doc/typo corrections. Co-Authored-By: Claude Opus 4.8 * Address Copilot re-review: SW cache, a11y, znorm, audio init - ozwell-model-sw.js: serve the buffered response when cache.put fails instead of re-downloading the file; only refetch when buffering itself fails. - VoiceSetup: the enrollment octopus is now a real

Uh oh!

Hey Ozwell Docs page #287

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions