@happyvertical/smrt-planner 0.0.0-stage → 0.0.2

package/dist/assistant/voice.svelte.js

@@ -0,0 +1,261 @@
+import { createSttDictationSource, LOCAL_SPEECH_MODELS, probeBrowserSpeech, } from '@happyvertical/smrt-svelte/browser-ai';
+/**
+ * Moonshine base first: it is accurate enough for plain talking, where the
+ * tiny model garbled words. Tiny stays for the quickest, least exact option.
+ */
+export const SPEECH_MODELS = [
+    {
+        id: 'moonshine-base',
+        label: 'Moonshine base',
+        note: 'Accurate and still quick, best for talking as you go (default)',
+        adapter: 'moonshine',
+        bytes: LOCAL_SPEECH_MODELS['moonshine-base']?.bytes ?? 0,
+    },
+    {
+        id: 'moonshine-tiny',
+        label: 'Moonshine tiny',
+        note: 'Fastest, less accurate',
+        adapter: 'moonshine',
+        bytes: LOCAL_SPEECH_MODELS['moonshine-tiny']?.bytes ?? 0,
+    },
+    {
+        id: 'whisper-tiny.en',
+        label: 'Whisper tiny',
+        note: 'English, the original choice',
+        adapter: 'whisper-local',
+        bytes: LOCAL_SPEECH_MODELS['whisper-tiny.en']?.bytes ?? 0,
+    },
+];
+export const DEFAULT_SPEECH_MODEL = 'moonshine-base';
+export function speechModelChoice(id) {
+    return (SPEECH_MODELS.find((choice) => choice.id === id) ??
+        SPEECH_MODELS[0]);
+}
+export const VOICE_PREFS_KEY = 'smrt-planner:voice:v1';
+function loadVoicePrefs(storage) {
+    try {
+        const raw = storage?.getItem(VOICE_PREFS_KEY);
+        if (raw) {
+            const value = JSON.parse(raw);
+            const local = value.local === true;
+            const known = SPEECH_MODELS.find((m) => m.id === value.model)?.id;
+            return {
+                local,
+                preferLocal: value.preferLocal === true,
+                // Someone who downloaded before the choice existed has Whisper tiny.
+                model: known ?? (local ? 'whisper-tiny.en' : DEFAULT_SPEECH_MODEL),
+            };
+        }
+    }
+    catch {
+        // Unreadable or unavailable: the defaults.
+    }
+    return { local: false, preferLocal: false, model: DEFAULT_SPEECH_MODEL };
+}
+function saveVoicePrefs(storage, prefs) {
+    try {
+        storage?.setItem(VOICE_PREFS_KEY, JSON.stringify(prefs));
+    }
+    catch {
+        // Not saved; the choice still holds for this visit.
+    }
+}
+/** Whether the visitor has downloaded (and so accepted) the local model. */
+export function loadVoiceConsent(storage) {
+    return loadVoicePrefs(storage).local;
+}
+export function saveVoiceConsent(storage, local) {
+    saveVoicePrefs(storage, { ...loadVoicePrefs(storage), local });
+}
+/** Whether the visitor chose the downloadable model over the browser's own. */
+export function loadPreferLocal(storage) {
+    return loadVoicePrefs(storage).preferLocal;
+}
+/** The speech model the visitor chose (Moonshine base until they say). */
+export function loadSpeechModel(storage) {
+    return loadVoicePrefs(storage).model;
+}
+export function planVoice(support, consented, cached) {
+    if (support === 'works')
+        return 'browser';
+    // They already took the download: turn it on without asking again.
+    return consented && cached ? 'enable-cached' : 'offer';
+}
+export class VoiceSession {
+    status = $state('checking');
+    /** Whether the model files are already on this device. */
+    cached = $state(false);
+    /** Download progress, 0 to 1. */
+    progress = $state(0);
+    error = $state('');
+    /** Whether the browser's own recogniser works here (whatever is in use). */
+    browserWorks = $state(false);
+    /** The visitor chose the downloadable model even where the browser's works. */
+    preferLocal = $state(false);
+    /** The size the offer shows, in bytes. */
+    size = $state(0);
+    /** The chosen downloadable model. */
+    model = $state(DEFAULT_SPEECH_MODEL);
+    /** Every byte is in and the model is being prepared ("Getting ready"). */
+    preparing = $state(false);
+    /** The dictation source for the composer, or null when voice is off. */
+    dictation = $state.raw(null);
+    handle = null;
+    abort = null;
+    options;
+    /**
+     * Bumped by every init, enable and restart. Async work captures the value
+     * it started under and drops its result when a newer one has begun, so a
+     * superseded probe or download cannot overwrite the current status or
+     * dictation source.
+     */
+    generation = 0;
+    constructor(options) {
+        this.options = options;
+    }
+    /** Probe the browser and decide what to offer. Call once. */
+    async init() {
+        const generation = ++this.generation;
+        try {
+            await this.probeAndOffer(generation);
+        }
+        catch (cause) {
+            if (generation !== this.generation)
+                return;
+            this.status = 'error';
+            this.error = cause instanceof Error ? cause.message : String(cause);
+        }
+    }
+    async probeAndOffer(generation) {
+        const probed = await (this.options.probe ?? probeBrowserSpeech)();
+        if (generation !== this.generation)
+            return;
+        this.browserWorks = probed === 'works';
+        this.preferLocal = loadPreferLocal(this.options.storage);
+        this.model = loadSpeechModel(this.options.storage);
+        // Choosing the download treats the browser's recogniser as unavailable.
+        const support = this.preferLocal && probed === 'works' ? 'unreliable' : probed;
+        if (support === 'works') {
+            this.dictation = (this.options.browserSource ??
+                (() => createSttDictationSource({ type: 'browser-speech' })))();
+            this.status = 'browser';
+            return;
+        }
+        const model = this.options.createModel(this.model);
+        this.handle?.dispose();
+        this.handle = model;
+        this.size = model.estimateSize();
+        const cached = await model.isCached().catch(() => false);
+        // Superseded while checking the cache: the restart that superseded this
+        // already disposed the model; the newer session owns the state.
+        if (generation !== this.generation)
+            return;
+        this.cached = cached;
+        const plan = planVoice(support, loadVoiceConsent(this.options.storage), this.cached);
+        if (plan === 'enable-cached')
+            await this.enable();
+        else
+            this.status = 'offer';
+    }
+    /** Download (or read from the cache) and turn voice typing on. */
+    async enable() {
+        const model = this.handle;
+        if (!model || this.status === 'downloading')
+            return;
+        const generation = ++this.generation;
+        this.status = 'downloading';
+        this.error = '';
+        this.progress = 0;
+        this.preparing = false;
+        const abort = new AbortController();
+        this.abort = abort;
+        try {
+            await model.load({
+                onProgress: (p) => {
+                    if (generation !== this.generation)
+                        return;
+                    this.progress = p.bytesTotal > 0 ? p.bytesLoaded / p.bytesTotal : 0;
+                    // The bytes are in; compiling the model for this device takes a moment.
+                    this.preparing = p.state === 'extracting';
+                },
+                signal: abort.signal,
+            });
+            if (generation !== this.generation) {
+                // Superseded (restart, dispose or a newer enable): drop this result.
+                // Whoever superseded it owns the model's disposal.
+                return;
+            }
+            // Consent is kept only once the download completed.
+            saveVoiceConsent(this.options.storage, true);
+            this.cached = true;
+            const choice = speechModelChoice(this.model);
+            this.dictation = (this.options.localSource ??
+                ((m, c) => createSttDictationSource({
+                    type: c.adapter,
+                    // The consent flow and the dictation share one loaded model.
+                    modelHandle: m,
+                })))(model, choice);
+            this.status = 'ready';
+        }
+        catch (error) {
+            if (generation !== this.generation)
+                return;
+            if (abort.signal.aborted) {
+                this.status = 'offer';
+            }
+            else {
+                this.error = error instanceof Error ? error.message : String(error);
+                this.status = 'error';
+            }
+        }
+        finally {
+            if (generation === this.generation) {
+                this.abort = null;
+                this.preparing = false;
+            }
+        }
+    }
+    /** Switch between the browser's recogniser and the downloadable model. */
+    async setPreferLocal(on) {
+        if (!this.browserWorks || on === this.preferLocal)
+            return;
+        await this.restart({ preferLocal: on });
+    }
+    /** Pick the downloadable model; the new one is offered (or turned on if cached). */
+    async setModel(id) {
+        if (id === this.model || !SPEECH_MODELS.some((m) => m.id === id))
+            return;
+        await this.restart({ model: id });
+    }
+    /** Save a changed choice, let go of the current model, and decide again. */
+    async restart(change) {
+        this.generation++;
+        this.abort?.abort();
+        this.abort = null;
+        this.preparing = false;
+        this.handle?.dispose();
+        this.handle = null;
+        this.dictation = null;
+        saveVoicePrefs(this.options.storage, {
+            ...loadVoicePrefs(this.options.storage),
+            ...change,
+        });
+        this.status = 'checking';
+        await this.init();
+    }
+    /** Stop a download in progress. */
+    cancel() {
+        this.abort?.abort();
+    }
+    /** "Not now": hide the offer for this visit. */
+    dismiss() {
+        this.status = 'dismissed';
+    }
+    dispose() {
+        this.generation++;
+        this.abort?.abort();
+        this.handle?.dispose();
+        this.handle = null;
+        this.dictation = null;
+    }
+}