The all-in-one local AI studio for your desktop: chat, image and video generation and a coding agent in one free, open source app. Windows and Linux. No Docker, no terminal, no cloud required.
TypeScript agentai-chatai-studioall-in-onecoding-agentdesktop-appfluximage-generationlinuxllmlocal-ailocal-llmopen-sourceprivacyself-hostedstable-diffusionuncensoreduncensored-aivideo-generationwindows
import { useEffect, useImperativeHandle, useRef, type Ref } from "react"
import { Mic, MicOff, Loader2 } from "lucide-react"
import { useVoice } from "../../hooks/useVoice"
import { platformText } from "../../lib/platform-text"
import { useVoiceStore } from "../../stores/voiceStore"
import { useSettingsStore } from "../../stores/settingsStore"
/**
* Why the microphone is greyed out, in the user's terms.
*
* The old wording was "Speech-to-text off. Enable it in Settings, Voice &
* Remote", which named a switch as the cause. The Gegenprobe on the Windows
* build (2026-08-30) took the faster_whisper package out of every Python it
* could find and got exactly that sentence, while the real cause was the
* missing package. There is no speech-to-text switch to flip either: the mic
* is gated on `sttAvailable`, a fresh probe of what is installed, and the
* persisted `sttEnabled` flag is read by nobody and offered nowhere in
* Settings. So the text now names the state the app actually found and points
* at the control that fixes it.
*
* Two states can grey the button out, and they need different sentences:
* - Local mode: faster-whisper is not installed. Voice & Remote has the
* Download & Install button right there.
* - Cloud mode without a usable cloud account: dictation is hosted, and the
* Voice tab deliberately shows no install button in cloud mode, so the way
* out is the account, not an install.
*/
export function micUnavailableHint(cloudMode: boolean): string {
return cloudMode ? platformText("micHintCloud") : platformText("micHintLocal")
}
/** What the composer can ask of the microphone. */
export interface VoiceButtonHandle {
/** The message was sent while recording: end the take, keep nothing of it
* back. The words already in the field went out with the message. */
endOnSend: () => void
}
interface Props {
ref?: Ref<VoiceButtonHandle>
onTranscript: (text: string) => void
/** Live interim transcript while recording (streaming dictation). */
onInterim?: (text: string) => void
onRecordingChange?: (isRecording: boolean) => void
disabled?: boolean
}
export function VoiceButton({ ref, onTranscript, onInterim, onRecordingChange, disabled }: Props) {
const { isRecording, isTranscribing, sttSupported, sttError, clearSttError, startRecording, stopRecording, cancelRecording, recheckStt, maxRecordingMs } = useVoice()
// Only for the greyed-out hint: cloud mode and local mode fail for different
// reasons and have different ways out.
const appMode = useSettingsStore((s) => s.settings.appMode)
// Auto-stop timer for cloud dictation — the transcribe route rejects takes
// past ~6.5 min (12 MiB) with a 413 that loses the WHOLE recording, so the
// take is stopped and transcribed just under the cap instead.
const autoStopRef = useRef<ReturnType<typeof setTimeout> | null>(null)
const wasRecordingRef = useRef(false)
useEffect(() => {
// The startup probe (App.tsx) can run before the persistent Whisper server
// has finished loading its model. If STT still reads unavailable when the
// mic mounts, do one fresh probe so a late-ready server lights it up.
if (!sttSupported) void recheckStt()
// eslint-disable-next-line react-hooks/exhaustive-deps
}, [])
// Dictation failures (cloud 401/403/413/429/5xx, dead mic) surface as a
// transient bubble over the mic instead of the take silently vanishing.
useEffect(() => {
if (!sttError) return
const t = setTimeout(() => clearSttError(), 6000)
return () => clearTimeout(t)
}, [sttError, clearSttError])
// Recording can end outside handleClick (close-to-tray teardown, unmount
// recovery) — mirror the transition to the composer so "Recording…" never
// sticks, and drop a pending auto-stop timer so it can't fire on a dead take.
useEffect(() => {
if (wasRecordingRef.current && !isRecording) {
if (autoStopRef.current) { clearTimeout(autoStopRef.current); autoStopRef.current = null }
onRecordingChange?.(false)
}
wasRecordingRef.current = isRecording
// eslint-disable-next-line react-hooks/exhaustive-deps
}, [isRecording])
useEffect(() => () => { if (autoStopRef.current) clearTimeout(autoStopRef.current) }, [])
const finishRecording = async () => {
if (autoStopRef.current) { clearTimeout(autoStopRef.current); autoStopRef.current = null }
onRecordingChange?.(false)
const transcript = await stopRecording()
if (transcript.trim()) {
onTranscript(transcript.trim())
}
}
// Sent while recording (David, 15.10.2026): the take went on after the
// send, and the button that ends it was greyed out with the answer
// running, so only a stop of the answer got the microphone back.
useImperativeHandle(ref, () => ({
endOnSend: () => {
if (autoStopRef.current) { clearTimeout(autoStopRef.current); autoStopRef.current = null }
if (!useVoiceStore.getState().isRecording) return
onRecordingChange?.(false)
void cancelRecording()
},
}))
const handleClick = async () => {
if (disabled || isTranscribing) return
if (isRecording) {
await finishRecording()
} else {
onRecordingChange?.(true)
const ok = await startRecording((interim) => onInterim?.(interim))
// Roll back the composer's "Recording…" state when the mic never
// started (permission denied / no input device) — otherwise Enter-to-
// send stays blocked with no recovery path.
if (!ok) {
onRecordingChange?.(false)
return
}
if (maxRecordingMs) {
autoStopRef.current = setTimeout(() => {
autoStopRef.current = null
if (!useVoiceStore.getState().isRecording) return
useVoiceStore.getState().setSttError("Dictation limit reached, transcribing what was recorded so far")
void finishRecording()
}, maxRecordingMs)
}
}
}
if (!sttSupported) {
const hint = micUnavailableHint(appMode === "cloud")
return (
<div className="relative group/mic shrink-0">
<button
disabled
className="lu-control lu-control--icon"
aria-label="Microphone unavailable"
title={hint}
>
<MicOff size={14} />
</button>
<div className="absolute bottom-full left-1/2 -translate-x-1/2 mb-2 px-2 py-1 w-max max-w-[240px] bg-gray-800 dark:bg-gray-700 !text-white t-micro leading-snug rounded text-center opacity-0 group-hover/mic:opacity-100 transition-opacity pointer-events-none">
{hint}
</div>
</div>
)
}
// Transcribing state — show spinner.
// Der Zustand kommt aus `aria-busy`, wie beim Modellwaehler, der waehrend
// eines Ladevorgangs dieselbe Akzentkante traegt. Vorher war das ein
// eigenes blaues Rezept (`bg-blue-500/20 border-blue-500/40
// text-blue-400`) — dieselbe Farbe, die auch der Fokusring fuehrt.
if (isTranscribing) {
return (
<button
disabled
aria-busy="true"
className="lu-control lu-control--icon"
aria-label="Transcribing audio"
>
<Loader2 size={14} className="animate-spin" />
</button>
)
}
return (
<div className="relative shrink-0">
<button
onClick={handleClick}
disabled={disabled}
// Das Mikrofon war bis hierher die einzige eigene Formsprache der
// Composer-Leiste (Audit Welle 3): `p-1.5 rounded-lg` ergab Radius
// 9,2px neben 8px ueberall sonst, und der Ein-Zustand war ein rotes
// Pill. Jetzt dasselbe Rezept wie Paperclip, Think, Stop und Send —
// der Ein-Zustand kommt aus `aria-pressed`, nicht aus einer zweiten
// Klassenkette. composer-grammar.test.ts zaehlt diese Datei mit.
className="lu-control lu-control--icon"
aria-pressed={isRecording}
data-voice-button
aria-label={isRecording ? "Stop recording" : "Start voice input"}
>
{isRecording && <span className="lu-control__pulse" aria-hidden="true" />}
<Mic size={14} />
</button>
{sttError && (
<div
role="alert"
className="absolute bottom-full left-1/2 -translate-x-1/2 mb-2 px-2 py-1 w-max max-w-[240px] bg-red-600/95 dark:bg-red-500/90 text-white t-micro leading-snug rounded text-center pointer-events-none z-10"
>
{sttError}
</div>
)}
</div>
)
}