Voice-to-text dictation app with local (Nvidia Parakeet/Whisper) and cloud models (BYOK). Privacy-first and available cross-platform.
JavaScript aianthropiccross-platformgeminigroqlinuxmacosnvidiaopen-sourceopenaiparakeetspeech-to-texttranscribewhisperwindows
import { useCallback, useEffect, useRef, useState, type ReactNode } from "react";
import { createPortal } from "react-dom";
import { useTranslation } from "react-i18next";
import confetti from "canvas-confetti";
import { ChevronDown, CornerDownLeft, Mic, RefreshCw, Sparkles } from "../icons";
import { Button } from "../ui/button";
import { toolIcons } from "../chat/toolIcons";
import { VoicePill, type VoicePillState } from "../dictation/VoicePill";
import { useListeningEntrancePhase } from "../../hooks/useListeningEntrancePhase";
import {
resolveListeningEntrancePresentation,
resolveVoiceActivityPresentation,
} from "../../helpers/voicePillPresentation";
import type {
OnboardingDemoEvent,
OnboardingDemoKind,
OnboardingDemoStatus,
} from "../../types/electron";
import founderAvatar from "../../assets/onboarding-founder.webp";
import gmailMark from "../../assets/icons/gmail.svg";
/**
* The dictation success celebration: one soft burst from each side edge, angled
* inward and up so the middle of the screen stays readable.
*
* One volley of ~70 small pieces in the onboarding blues plus one warm accent,
* settling in about a second: an acknowledgement, not a continuous stream.
*/
const CONFETTI_BASE = {
// --color-primary and two lighter tints of it, plus one warm accent.
colors: ["#4577e9", "#7da2f2", "#b9cdf8", "#f5c518"],
shapes: ["circle", "square"] as ("circle" | "square")[],
particleCount: 36,
scalar: 0.85,
spread: 50,
startVelocity: 38,
gravity: 0.9,
decay: 0.91,
ticks: 140,
// canvas-confetti opts out natively; the CSS in index.css cannot reach a canvas.
disableForReducedMotion: true,
};
function ConfettiLayer() {
const canvasRef = useRef<HTMLCanvasElement>(null);
useEffect(() => {
const canvas = canvasRef.current;
if (!canvas) return;
// useWorker: false on purpose. The worker path hands the canvas to an
// OffscreenCanvas, which can only be transferred once, and StrictMode mounts
// effects twice in dev — the second pass throws.
const fire = confetti.create(canvas, { resize: true, useWorker: false });
// 60 from the left edge and 120 from the right both angle inward and up.
void fire({ ...CONFETTI_BASE, angle: 60, origin: { x: 0, y: 0.7 } });
void fire({ ...CONFETTI_BASE, angle: 120, origin: { x: 1, y: 0.7 } });
return () => fire.reset();
}, []);
// Escape the transformed onboarding step so the fixed canvas spans the full
// window and particles can fall naturally past the panel's bottom edge.
return createPortal(
<canvas
ref={canvasRef}
className="pointer-events-none fixed inset-0 z-20 h-full w-full"
aria-hidden="true"
/>,
document.body
);
}
// Figma "Frame 26": 40px round avatar, bottom-aligned with the bubble beside it.
function FounderAvatar() {
return (
<img
src={founderAvatar}
// Decorative: the demo conversation never names the sender, so a
// descriptive alt would announce a person the transcript doesn't mention.
alt=""
aria-hidden="true"
width={40}
height={40}
decoding="async"
draggable={false}
className="size-9 shrink-0 rounded-full object-cover"
/>
);
}
// Figma "Frame 25"/"Frame 27": bubble on surface/brand, 10/20 padding, Inter
// Medium 14/140% in light/surface-primary. Radius 16 rather than the spec's 38:
// the second message wraps, and a full pill radius turned two lines into a blob.
function FounderBubble({ children }: { children: ReactNode }) {
return (
<p
className="w-fit rounded-2xl bg-[var(--onboarding-accent)] px-4 py-2 text-sm font-medium leading-[1.4] text-[var(--onboarding-accent-foreground)]"
style={BUBBLE_IN}
>
{children}
</p>
);
}
// The keyframe the chat surfaces already use for message entry (ChatMessage.tsx,
// MeetingTranscriptChat.tsx), so the demo bubbles land with the app's one bubble
// motion instead of a second dialect. The global prefers-reduced-motion block in
// index.css clamps the duration, and `both` leaves the bubble visible either way.
const BUBBLE_IN = { animation: "agent-message-in 220ms ease-out both" } as const;
// Streaming transcript partials arrive while the microphone is still open.
const isListening = (status: OnboardingDemoStatus | undefined) =>
status === "listening" || status === "partial";
function DemoVoicePill({
status,
agent = false,
getLevel,
stopLabel,
onStop,
}: {
status: OnboardingDemoStatus | undefined;
/** The assistant demo: purple identity and thinking glow. */
agent?: boolean;
getLevel: () => number | null;
stopLabel: string;
onStop: () => void;
}) {
const listening = isListening(status);
const busy = status === "processing" || status === "replying";
const phase = useListeningEntrancePhase(listening);
const entrance = resolveListeningEntrancePresentation({ isRecording: listening, phase });
// Processing collapses to the logo and lights the Signal glow, as the
// dictation window's pill does.
const activity = resolveVoiceActivityPresentation({
isRecording: listening,
isProcessing: busy,
isAssistantVoice: agent,
assistantThinking: agent && status === "replying",
});
// The presentation helpers are untyped JS; their active state is one of ours.
const state = (entrance.activeState ?? activity.activeState ?? "idle") as VoicePillState;
return (
<VoicePill
variant="floating"
state={state}
expanded={listening ? entrance.compactPill : activity.compactPill}
collapseToLogo={entrance.collapseToLogo}
waveformVisible={entrance.waveformVisible}
agentMode={agent && (listening || busy)}
horizontalDirection="left"
getAudioLevel={getLevel}
role={listening ? "button" : "status"}
tabIndex={listening ? 0 : undefined}
aria-label={listening ? stopLabel : undefined}
title={listening ? stopLabel : undefined}
onClick={listening ? onStop : undefined}
onKeyDown={
listening
? (keyEvent) => {
if (keyEvent.key === "Enter" || keyEvent.key === " ") {
keyEvent.preventDefault();
onStop();
}
}
: undefined
}
className="shrink-0"
/>
);
}
function TypingDots() {
// A 40x20 row of 8px dots on 16px centers (so an 8px gap) in light/text-tertiary.
return (
<span className="flex h-5 items-center gap-2">
{[0, 1, 2].map((index) => (
<span
key={index}
className="onboarding-typing-dot size-1.5 rounded-full bg-[var(--onboarding-text-tertiary)]"
style={{ animationDelay: `${index * 160}ms` }}
/>
))}
</span>
);
}
function TypingBubble() {
const { t } = useTranslation();
return (
// Figma "Onboarding / Frame 25": on light/surface-stroke, 10/20 padding,
// hugging the dots row. Radius matches FounderBubble.
<div
className="flex items-center rounded-2xl bg-[var(--onboarding-control-border)] px-4 py-2"
style={BUBBLE_IN}
aria-label={t("onboarding.rehaul.demo.typing")}
>
<TypingDots />
</div>
);
}
interface DemoStepProps {
kind: OnboardingDemoKind;
firstMessage: string;
secondMessage: string;
/** Dictation only; the assistant card uses secondMessage as its placeholder. */
placeholder?: string;
listeningLabel: string;
processingLabel: string;
stopLabel: string;
retryLabel: string;
onSuccessChange: (successful: boolean) => void;
initialSuccessful?: boolean;
}
export default function DemoStep({
kind,
firstMessage,
secondMessage,
placeholder = "",
listeningLabel,
processingLabel,
stopLabel,
retryLabel,
onSuccessChange,
initialSuccessful = false,
}: DemoStepProps) {
const [messageCount, setMessageCount] = useState(0);
const [event, setEvent] = useState<OnboardingDemoEvent | null>(null);
const [draft, setDraft] = useState("");
// Assistant demo only: what the user said, shown above the reply it produced.
const [transcript, setTranscript] = useState("");
const [demoId, setDemoId] = useState(() => crypto.randomUUID());
const [restoredSuccessful, setRestoredSuccessful] = useState(initialSuccessful);
const inputRef = useRef<HTMLTextAreaElement>(null);
// Sampled by the pill's waveform from a rAF loop; a ref keeps the ~12 level
// events a second from re-rendering the step.
const levelRef = useRef(0);
const getLevel = useCallback(() => levelRef.current, []);
useEffect(() => {
const first = window.setTimeout(() => setMessageCount(1), 650);
const second = window.setTimeout(() => setMessageCount(2), 1400);
const input = window.setTimeout(() => inputRef.current?.focus(), 1550);
return () => {
window.clearTimeout(first);
window.clearTimeout(second);
window.clearTimeout(input);
};
}, [demoId]);
useEffect(() => {
void window.electronAPI?.beginOnboardingDemo?.({ id: demoId, kind });
const unsubscribe = window.electronAPI?.onOnboardingDemoEvent?.((payload) => {
if (payload.demoId !== demoId || payload.kind !== kind) return;
if (payload.status === "level") {
levelRef.current = payload.level ?? 0;
return;
}
if (payload.status === "listening") levelRef.current = 0;
setRestoredSuccessful(false);
setEvent(payload);
if (payload.text) {
// The assistant demo hears first and writes second: transcript text
// (live partials, then the final handed to the model) stays out of the
// reply box, which only ever holds the streamed answer.
const heard = payload.status === "partial" || payload.status === "processing";
if (kind === "assistant" && heard) setTranscript(payload.text);
else setDraft(payload.text);
}
if (payload.status === "success") onSuccessChange(true);
});
return () => {
unsubscribe?.();
void window.electronAPI?.endOnboardingDemo?.(demoId);
};
}, [demoId, kind, onSuccessChange]);
const retry = () => {
setRestoredSuccessful(false);
onSuccessChange(false);
setEvent(null);
setDraft("");
setTranscript("");
setMessageCount(0);
setDemoId(crypto.randomUUID());
};
const effectiveEvent: OnboardingDemoEvent | null = restoredSuccessful
? { demoId, kind, status: "success" }
: event;
const status = effectiveEvent?.status;
const successful = status === "success";
const stop = () => void window.electronAPI?.stopOnboardingDemo?.(demoId);
return (
<div
// One compact width for both demos keeps the two steps from stepping in
// and out as the flow moves between them.
className="relative mx-auto mt-5 w-full max-w-lg"
>
{successful && kind === "dictation" && <ConfettiLayer />}
{kind === "dictation" ? (
// Frame 2147258978: 40 between the bubbles and the card. Explicit gaps
// rather than space-y so the card's spacing can't collide with the
// 12 that separates the two bubble rows.
<div className="flex flex-col gap-6">
<div className="flex flex-col gap-3">
{messageCount === 0 ? (
<div className="flex items-end gap-2.5">
<FounderAvatar />
<TypingBubble />
</div>
) : (
<>
{/* Frame 32: the first bubble carries the avatar gutter as padding
(40 avatar + 10 gap) so both messages share one left edge and
the avatar rides the newest row. */}
<div className="flex items-end gap-2.5 ps-11">
<FounderBubble>{firstMessage}</FounderBubble>
</div>
{/* Frame 33 */}
<div className="flex items-end gap-2.5">
<FounderAvatar />
{messageCount === 1 ? (
<TypingBubble />
) : (
<FounderBubble>{secondMessage}</FounderBubble>
)}
</div>
</>
)}
</div>
{messageCount >= 2 && (
<VoiceSurface
inputRef={inputRef}
value={draft}
onChange={setDraft}
placeholder={placeholder}
event={effectiveEvent}
getLevel={getLevel}
listeningLabel={listeningLabel}
processingLabel={processingLabel}
stopLabel={stopLabel}
retryLabel={retryLabel}
onRetry={retry}
onStop={stop}
/>
)}
</div>
) : (
<EmailThread body={firstMessage}>
<VoiceSurface
inputRef={inputRef}
value={draft}
onChange={setDraft}
placeholder={transcript ? "" : secondMessage}
event={effectiveEvent}
transcript={transcript}
getLevel={getLevel}
agent
listeningLabel={listeningLabel}
processingLabel={processingLabel}
stopLabel={stopLabel}
retryLabel={retryLabel}
onRetry={retry}
onStop={stop}
embedded
/>
</EmailThread>
)}
</div>
);
}
/**
* One message in a mail client, with the reply composer as the voice surface.
* Screen context reads this card, so the sender and the ask are plain text.
*/
function EmailThread({ body, children }: { body: string; children: ReactNode }) {
const { t } = useTranslation();
const senderName = t("onboarding.rehaul.assistantDemo.email.senderName");
return (
<article className="overflow-hidden rounded-2xl border border-[var(--onboarding-control-border)] bg-[var(--onboarding-surface)] text-start">
<header className="flex items-center gap-2.5 border-b border-[var(--onboarding-control-border)] px-4 py-2">
<img
src={gmailMark}
alt=""
aria-hidden="true"
width={16}
height={12}
decoding="async"
draggable={false}
className="h-3 w-4 shrink-0 select-none"
/>
<h2 className="truncate text-sm font-medium leading-[1.4] text-[var(--onboarding-text-primary)]">
{t("onboarding.rehaul.assistantDemo.email.subject")}
</h2>
</header>
<div className="flex gap-3 px-4 py-3">
<span
aria-hidden="true"
className="flex size-9 shrink-0 items-center justify-center rounded-full bg-[color-mix(in_srgb,var(--onboarding-accent)_12%,transparent)] text-sm font-semibold text-[var(--onboarding-accent)]"
>
{senderName.trim().charAt(0)}
</span>
<div className="min-w-0 flex-1">
<p className="text-sm font-medium leading-[1.4] text-[var(--onboarding-text-primary)]">
{senderName}
</p>
<p className="flex items-center gap-1 text-xs leading-[1.4] text-[var(--onboarding-text-secondary)]">
{t("onboarding.rehaul.assistantDemo.email.recipient")}
<ChevronDown className="size-3 shrink-0" strokeWidth={1.5} aria-hidden="true" />
</p>
<p className="mt-2 whitespace-pre-line text-sm leading-[1.5] text-[var(--onboarding-text-primary)]">
{body}
</p>
</div>
</div>
<div
role="group"
aria-label={t("onboarding.rehaul.assistantDemo.email.reply")}
className="flex gap-3 border-t border-[var(--onboarding-control-border)] bg-[var(--onboarding-surface-secondary)] px-4 py-3"
>
<span
aria-hidden="true"
className="flex size-9 shrink-0 items-center justify-center rounded-full border border-[var(--onboarding-control-border)] bg-[var(--onboarding-surface)] text-[var(--onboarding-text-secondary)]"
>
<CornerDownLeft className="size-4" strokeWidth={1.8} />
</span>
{children}
</div>
</article>
);
}
function VoiceSurface({
inputRef,
value,
onChange,
placeholder,
event,
transcript,
getLevel,
agent = false,
listeningLabel,
processingLabel,
stopLabel,
retryLabel,
onRetry,
onStop,
embedded = false,
}: {
inputRef: React.RefObject<HTMLTextAreaElement | null>;
value: string;
onChange: (value: string) => void;
placeholder: string;
event: OnboardingDemoEvent | null;
/** What was heard, shown above the reply while the assistant answers it. */
transcript?: string;
getLevel: () => number | null;
agent?: boolean;
listeningLabel: string;
processingLabel: string;
stopLabel: string;
retryLabel: string;
onRetry: () => void;
onStop: () => void;
embedded?: boolean;
}) {
const { t } = useTranslation();
const status = event?.status;
// The transcript has been handed to the model and no token has landed yet.
const awaitingReply =
(status === "processing" || status === "replying") && !!transcript && !value;
const ToolIcon = event?.tool ? (toolIcons[event.tool] ?? Sparkles) : null;
return (
<div
// The assistant variant runs taller so a few-sentence reply fits unscrolled.
className={`relative flex flex-col rounded-[14px] border border-[var(--onboarding-control-border)] bg-[var(--onboarding-surface)] p-3 ${
embedded ? "h-52 min-w-0 flex-1" : "h-36"
}`}
>
{transcript && (
<p className="mb-1.5 flex items-start gap-1.5 pe-10 text-xs italic leading-[1.4] text-[var(--onboarding-text-secondary)]">
<Mic className="mt-0.5 size-3 shrink-0" aria-hidden="true" />
<span className="line-clamp-2">{transcript}</span>
</p>
)}
{awaitingReply && (
<span className="mb-1" aria-label={t("onboarding.rehaul.demo.typing")}>
<TypingDots />
</span>
)}
<textarea
dir="auto"
ref={inputRef}
value={value}
onChange={(inputEvent) => onChange(inputEvent.target.value)}
placeholder={placeholder}
// Placeholder is text-tertiary at 38% — 16/140% in the mail card, 18/140%
// in the dictation one. The caret takes the brand colour, which is what
// Figma draws as the 3x18 bar.
className={`input-inline min-h-0 w-full flex-1 resize-none bg-transparent pe-12 leading-[1.4] text-[var(--onboarding-text-primary)] caret-[var(--onboarding-accent)] outline-none placeholder:text-[color-mix(in_srgb,var(--onboarding-text-tertiary)_38%,transparent)] ${
embedded ? "text-sm" : "text-base"
}`}
/>
{/* Anchored to the corner so the pill grows into the card, leftward. */}
<div className="absolute bottom-2 end-2 flex justify-end">
<DemoVoicePill
status={status}
agent={agent}
getLevel={getLevel}
stopLabel={stopLabel}
onStop={onStop}
/>
</div>
<div
className="absolute bottom-3 start-3 max-w-[15rem] text-xs text-[var(--onboarding-text-secondary)]"
aria-live="polite"
>
{isListening(status) && listeningLabel}
{status === "processing" && !transcript && processingLabel}
{ToolIcon && event?.tool && (
<span className="inline-flex items-center gap-1.5 rounded-full bg-[var(--onboarding-surface-secondary)] px-2 py-1 text-[var(--onboarding-text-primary)]">
<ToolIcon className="size-3.5 shrink-0" aria-hidden="true" />
{t(`agentMode.tools.${event.tool}Status`, {
defaultValue: t(`agentMode.tools.${event.tool}Name`, { defaultValue: event.tool }),
})}
</span>
)}
{status === "error" && (
<span className="inline-flex items-center gap-1 text-[var(--onboarding-danger)]">
{event.message}
<Button type="button" variant="ghost" size="sm" onClick={onRetry}>
<RefreshCw className="size-3" />
{retryLabel}
</Button>
</span>
)}
</div>
</div>
);
}