diff --git a/src/app/(app)/ToggleRealtime.tsx b/src/app/(app)/ToggleRealtime.tsx
index e4fdbd7..fc32ad7 100644
--- a/src/app/(app)/ToggleRealtime.tsx
+++ b/src/app/(app)/ToggleRealtime.tsx
@@ -67,8 +67,16 @@ const saveButtonClasses = [
 type ConnectionState = 'idle' | 'requesting' | 'ready' | 'error'
 
 export default function ToggleRealtime() {
-	const { start, stop, remoteStream, transcripts, updateInstructions, updateTurnDelaySeconds } =
-		useRealtimeVoiceSession()
+	const {
+		start,
+		stop,
+		remoteStream,
+		transcripts,
+		updateInstructions,
+		updateTurnDelaySeconds,
+		updateSpeechEnabled,
+		sendText
+	} = useRealtimeVoiceSession()
 	const [connectionState, setConnectionState] = useState<ConnectionState>('idle')
 	const [errorMessage, setErrorMessage] = useState<string | null>(null)
 	const [languageOrder, setLanguageOrder] = useState(languagePhrases)
@@ -78,6 +86,8 @@ export default function ToggleRealtime() {
 	const [saveConfirmation, setSaveConfirmation] = useState('')
 	const [turnDelaySeconds, setTurnDelaySeconds] = useState(1.2)
 	const [turnDelayDraft, setTurnDelayDraft] = useState('1.2')
+	const [speechEnabled, setSpeechEnabled] = useState(true)
+	const [textDraft, setTextDraft] = useState('')
 	const audioContextRef = useRef<AudioContext | null>(null)
 	const sourceRef = useRef<MediaStreamAudioSourceNode | null>(null)
 	const startedRef = useRef(false)
@@ -118,6 +128,13 @@ export default function ToggleRealtime() {
 		setTurnDelayDraft(String(normalized))
 	}, [normalizeTurnDelaySeconds])
 
+	useEffect(() => {
+		if (typeof window === 'undefined') return
+		const stored = window.localStorage.getItem('lilac.speechEnabled')
+		if (stored === null) return
+		setSpeechEnabled(stored !== 'false')
+	}, [])
+
 	useEffect(() => {
 		if (!saveConfirmation) return
 		const timer = window.setTimeout(() => setSaveConfirmation(''), 1800)
@@ -138,6 +155,12 @@ export default function ToggleRealtime() {
 		updateTurnDelaySeconds(turnDelaySeconds)
 	}, [turnDelaySeconds, updateTurnDelaySeconds])
 
+	useEffect(() => {
+		if (typeof window === 'undefined') return
+		window.localStorage.setItem('lilac.speechEnabled', speechEnabled ? 'true' : 'false')
+		updateSpeechEnabled(speechEnabled)
+	}, [speechEnabled, updateSpeechEnabled])
+
 	useEffect(() => {
 		if (typeof navigator === 'undefined') return
 		const navLanguages = navigator.languages ?? [navigator.language]
@@ -206,6 +229,18 @@ export default function ToggleRealtime() {
 				tracks: remoteStream?.getTracks().length
 			})
 
+			if (!speechEnabled) {
+				try {
+					sourceRef.current?.disconnect()
+				} catch {}
+				sourceRef.current = null
+				try {
+					void audioContextRef.current?.close()
+				} catch {}
+				audioContextRef.current = null
+				return
+			}
+
 			if (!remoteStream.getAudioTracks().length) {
 				const onAddTrack = () => {
 					remoteStream.removeEventListener('addtrack', onAddTrack as EventListener)
@@ -240,7 +275,7 @@ export default function ToggleRealtime() {
 			} catch {}
 			audioContextRef.current = null
 		}
-	}, [remoteStream, ensureAudioContext])
+	}, [remoteStream, ensureAudioContext, speechEnabled])
 
 	const beginSession = useCallback(async () => {
 		if (startedRef.current) return
@@ -250,7 +285,10 @@ export default function ToggleRealtime() {
 
 		try {
 			ensureAudioContext()
-			await start({ instructions: instructionsText, voice: 'verse' })
+			await start({
+				instructions: instructionsText,
+				...(speechEnabled ? { voice: 'verse' } : {})
+			})
 			if (cancelInitRef.current) {
 				startedRef.current = false
 				return
@@ -265,7 +303,7 @@ export default function ToggleRealtime() {
 				error instanceof Error ? error.message : 'Something went wrong while starting Lilac.'
 			setErrorMessage(message)
 		}
-	}, [ensureAudioContext, start, instructionsText])
+	}, [ensureAudioContext, start, instructionsText, speechEnabled])
 
 	useEffect(() => {
 		cancelInitRef.current = false
@@ -319,6 +357,7 @@ export default function ToggleRealtime() {
 
 	const phrase = languageOrder[activeIndex] ?? languageOrder[0]
 	const footerText = tab === 'session' ? statusText : ''
+	const canSendText = textDraft.trim().length > 0
 
 	useEffect(() => {
 		const el = transcriptListRef.current
@@ -346,9 +385,37 @@ export default function ToggleRealtime() {
 	const content =
 		tab === 'session' ? (
 			<div className="flex w-full max-w-xl flex-col gap-4">
+				<div className="flex items-center justify-between gap-4 rounded-3xl border border-white/25 bg-[var(--lilac-elevated)]/70 px-4 py-3 text-[var(--lilac-ink)] shadow-sm backdrop-blur">
+					<div className="flex flex-col">
+						<span className="font-semibold text-[10px] text-[var(--lilac-ink-muted)] uppercase tracking-[0.18em]">
+							Speech output
+						</span>
+						<span className="text-[var(--lilac-ink)] text-sm">
+							{speechEnabled ? 'Spoken replies enabled' : 'Text-only replies'}
+						</span>
+					</div>
+					<button
+						type="button"
+						role="switch"
+						aria-checked={speechEnabled}
+						onClick={() => setSpeechEnabled(enabled => !enabled)}
+						className={`relative h-8 w-14 rounded-full border transition ${
+							speechEnabled
+								? 'border-transparent bg-[var(--lilac-ink)]'
+								: 'border-white/40 bg-white/70 dark:bg-white/10'
+						}`}
+					>
+						<span
+							className={`-translate-y-1/2 absolute top-1/2 h-6 w-6 rounded-full bg-white shadow transition ${
+								speechEnabled ? 'translate-x-7' : 'translate-x-1'
+							}`}
+						/>
+					</button>
+				</div>
 				<div
 					ref={transcriptListRef}
-					className="h-[62dvh] overflow-y-auto rounded-3xl border border-white/30 bg-[var(--lilac-elevated)]/70 p-4 shadow-xl backdrop-blur"
+					className="h-[56dvh] overflow-y-auto overscroll-contain rounded-3xl border border-white/30 bg-[var(--lilac-elevated)]/70 p-4 shadow-xl backdrop-blur"
+					style={{ scrollbarGutter: 'stable' }}
 				>
 					{transcripts.length ? (
 						<div className="flex flex-col gap-3">
@@ -399,6 +466,33 @@ export default function ToggleRealtime() {
 						</div>
 					)}
 				</div>
+				<div className="rounded-3xl border border-white/25 bg-[var(--lilac-elevated)]/80 p-4 text-[var(--lilac-ink)] shadow-lg backdrop-blur">
+					<div className="mb-2 font-semibold text-[10px] text-[var(--lilac-ink-muted)] uppercase tracking-[0.18em]">
+						Text input
+					</div>
+					<div className="flex flex-col gap-3 sm:flex-row">
+						<textarea
+							className="min-h-[68px] w-full flex-1 resize-none rounded-2xl border border-white/30 bg-white/70 px-4 py-3 text-[var(--lilac-ink)] text-sm outline-none transition focus:border-white/70 focus:bg-white dark:bg-white/10 dark:focus:border-white/30 dark:focus:bg-white/20"
+							onChange={event => setTextDraft(event.target.value)}
+							placeholder="Type to translate or speak back."
+							rows={2}
+							value={textDraft}
+						/>
+						<button
+							type="button"
+							className={`${saveButtonClasses} h-12 w-full cursor-pointer text-sm sm:w-28 ${
+								canSendText ? '' : 'cursor-not-allowed opacity-60'
+							}`}
+							disabled={!canSendText}
+							onClick={() => {
+								if (!sendText(textDraft)) return
+								setTextDraft('')
+							}}
+						>
+							Send
+						</button>
+					</div>
+				</div>
 			</div>
 		) : (
 			<div className="flex w-full max-w-xl flex-col gap-4 text-left">
diff --git a/src/realtime/provider.tsx b/src/realtime/provider.tsx
index 49df684..3ec855b 100644
--- a/src/realtime/provider.tsx
+++ b/src/realtime/provider.tsx
@@ -16,7 +16,7 @@ export type RealtimeTranscriptMessage = {
 	 * - response_text: assistant text stream
 	 * - response_audio_transcript: assistant audio transcript stream
 	 */
-	source: 'input_transcription' | 'response_text' | 'response_audio_transcript'
+	source: 'input_transcription' | 'input_text' | 'response_text' | 'response_audio_transcript'
 }
 
 type RealtimeContextValue = {
@@ -27,6 +27,8 @@ type RealtimeContextValue = {
 	updateInstructions: (text: string) => void
 	updateVoice: (voice: string) => void
 	updateTurnDelaySeconds: (seconds: number) => void
+	updateSpeechEnabled: (enabled: boolean) => void
+	sendText: (text: string) => boolean
 	clearTranscripts: () => void
 	stop: () => void
 }
@@ -148,6 +150,7 @@ export function RealtimeProvider({ children }: { children: React.ReactNode }) {
 	const pendingAssistantDeltasByResponseIdRef = useRef<
 		Map<string, { text: string; audioTranscript: string }>
 	>(new Map())
+	const speechEnabledRef = useRef(true)
 
 	const cleanup = useCallback(() => {
 		peerRef.current?.close()
@@ -352,7 +355,7 @@ export function RealtimeProvider({ children }: { children: React.ReactNode }) {
 						dc.send(
 							JSON.stringify({
 								session: {
-									...(opts?.voice ? { voice: opts.voice } : {}),
+									...(speechEnabledRef.current && opts?.voice ? { voice: opts.voice } : {}),
 									...(opts?.instructions ? { instructions: opts.instructions } : {}),
 									audio: {
 										input: {
@@ -363,7 +366,7 @@ export function RealtimeProvider({ children }: { children: React.ReactNode }) {
 										}
 									},
 									// Request both audio and text so we can display a high-quality text timeline.
-									modalities: ['audio', 'text'],
+									modalities: speechEnabledRef.current ? ['audio', 'text'] : ['text'],
 									turn_detection: {
 										silence_duration_ms: silenceDurationMs,
 										type: 'server_vad'
@@ -448,6 +451,9 @@ export function RealtimeProvider({ children }: { children: React.ReactNode }) {
 							if (type === 'conversation.item.created') {
 								const itemId = typeof msg?.item?.id === 'string' ? msg.item.id : undefined
 								const role = msg?.item?.role
+								const content = Array.isArray(msg?.item?.content) ? msg.item.content : []
+								const inputText = content.find((entry: { type?: string }) => entry?.type === 'input_text')
+								const text = typeof inputText?.text === 'string' ? inputText.text : null
 								const previousItemId =
 									typeof msg?.previous_item_id === 'string'
 										? msg.previous_item_id
@@ -457,7 +463,18 @@ export function RealtimeProvider({ children }: { children: React.ReactNode }) {
 
 								if (typeof itemId === 'string') {
 									previousItemIdByIdRef.current.set(itemId, previousItemId)
-									if (role === 'user') latestCommittedInputItemIdRef.current = itemId
+									if (role === 'user') {
+										latestCommittedInputItemIdRef.current = itemId
+										if (text !== null) {
+											upsertTranscript({
+												id: itemId,
+												previousItemId,
+												replaceText: text,
+												role: 'user',
+												source: 'input_text'
+											})
+										}
+									}
 									reorderTranscripts()
 								}
 								return
@@ -799,15 +816,79 @@ export function RealtimeProvider({ children }: { children: React.ReactNode }) {
 		[dataChannel]
 	)
 
+	const updateSpeechEnabled = useCallback(
+		(enabled: boolean) => {
+			speechEnabledRef.current = enabled
+			if (!dataChannel) return
+			try {
+				dataChannel.send(
+					JSON.stringify({
+						session: {
+							...(enabled ? { voice: 'verse' } : {}),
+							modalities: enabled ? ['audio', 'text'] : ['text']
+						},
+						type: 'session.update'
+					})
+				)
+			} catch {
+				// ignore
+			}
+		},
+		[dataChannel]
+	)
+
+	const sendText = useCallback(
+		(text: string) => {
+			const trimmed = text.trim()
+			if (!trimmed || !dataChannel) return false
+			const id = crypto.randomUUID()
+			const previousItemId = latestCommittedInputItemIdRef.current
+			previousItemIdByIdRef.current.set(id, previousItemId ?? null)
+			latestCommittedInputItemIdRef.current = id
+			upsertTranscript({
+				id,
+				previousItemId,
+				replaceText: trimmed,
+				role: 'user',
+				source: 'input_text'
+			})
+			try {
+				dataChannel.send(
+					JSON.stringify({
+						item: {
+							content: [{ text: trimmed, type: 'input_text' }],
+							id,
+							...(previousItemId ? { previous_item_id: previousItemId } : {}),
+							role: 'user',
+							type: 'message'
+						},
+						type: 'conversation.item.create'
+					})
+				)
+				dataChannel.send(
+					JSON.stringify({
+						type: 'response.create'
+					})
+				)
+				return true
+			} catch {
+				return false
+			}
+		},
+		[dataChannel, upsertTranscript]
+	)
+
 	const value = useMemo<RealtimeContextValue>(
 		() => ({
 			clearTranscripts,
 			dataChannel,
 			remoteStream,
+			sendText,
 			start,
 			stop,
 			transcripts,
 			updateInstructions,
+			updateSpeechEnabled,
 			updateTurnDelaySeconds,
 			updateVoice
 		}),
@@ -816,9 +897,11 @@ export function RealtimeProvider({ children }: { children: React.ReactNode }) {
 			remoteStream,
 			transcripts,
 			start,
+			updateSpeechEnabled,
 			updateInstructions,
 			updateVoice,
 			updateTurnDelaySeconds,
+			sendText,
 			clearTranscripts,
 			stop
 		]