feat(conversation): OpenAI gpt-realtime-2.1 라이브 음성 대화 Premium 백엔드

- realtime-token Edge Function: 티어 검증 + realtime_session 쿼터 + ephemeral key 발급
- useRealtimeConversation 훅: WebRTC 직결 (마이크 → OpenAI, 오디오 자동 재생)
- VoiceConversationPage: conversationBackend 분기 + 연결 실패 시 로컬 fallback
- AppConfig.conversationBackend + SettingsModal 음성 대화 엔진 선택
- ErrorCode 799 ConversationRealtimeTokenFailed, i18n ko/en
This commit is contained in:
Yun Chan 2026-07-21 12:10:13 +09:00
parent 983c60cda2
commit 8f7d300b89
15 changed files with 762 additions and 40 deletions

View file

@ -12,6 +12,7 @@ import CancelIcon from '@mui/icons-material/Cancel'
import { MetalCard, PhosphorText, Led, PhysicalButton, ScreenPanel, InstrumentPanel } from '@d3ro/ui/components/ds'
import { PageHeader } from '../components/shared'
import { VoiceRecordingPanel } from '../components/voice-conversation/VoiceRecordingPanel'
import { useRealtimeConversation } from '../hooks/useRealtimeConversation'
import { isImeComposingEvent } from '../utils/keyboard'
import { d3roPalette, d3roFontMono, d3roTypo } from '@d3ro/ui/theme'
import { useI18n } from '@d3ro/i18n'
@ -24,41 +25,68 @@ import type {
export function VoiceConversationPage(): React.ReactElement {
const { t } = useI18n()
const [state, setState] = useState<ConversationState>('idle')
const [messages, setMessages] = useState<ConversationMessage[]>([])
const [isActive, setIsActive] = useState(false)
const [streamingText, setStreamingText] = useState('')
const [streamingMsgId, setStreamingMsgId] = useState<string | null>(null)
const [localState, setLocalState] = useState<ConversationState>('idle')
const [localMessages, setLocalMessages] = useState<ConversationMessage[]>([])
const [localActive, setLocalActive] = useState(false)
const [localStreamingText, setLocalStreamingText] = useState('')
const [localStreamingMsgId, setLocalStreamingMsgId] = useState<string | null>(null)
const [textInput, setTextInput] = useState('')
const [errorBanner, setErrorBanner] = useState<ConversationError | null>(null)
const [backend, setBackend] = useState<'local' | 'realtime'>('local')
const messagesEndRef = useRef<HTMLDivElement>(null)
const realtime = useRealtimeConversation()
const isRealtime = backend === 'realtime'
// 표시용 상태 — 백엔드에 따라 로컬 파이프라인(IPC) vs Realtime 훅
const state = isRealtime ? realtime.conversationState : localState
const messages = isRealtime ? realtime.messages : localMessages
const isActive = isRealtime ? realtime.isActive : localActive
const streamingText = isRealtime ? realtime.streamingText : localStreamingText
const streamingMsgId = isRealtime ? realtime.streamingMsgId : localStreamingMsgId
// 설정에서 대화 백엔드 로드
useEffect(() => {
window.electronAPI.config.getAll().then((r) => {
if (r.success && r.data.conversationBackend === 'realtime') {
setBackend('realtime')
}
})
}, [])
// Realtime 에러 → 배너
useEffect(() => {
if (realtime.error) {
setErrorBanner({ phase: 'llm', message: realtime.error })
}
}, [realtime.error])
const scrollToBottom = useCallback(() => {
messagesEndRef.current?.scrollIntoView({ behavior: 'smooth' })
}, [])
// IPC 이벤트 구독
// IPC 이벤트 구독 (로컬 파이프라인)
useEffect(() => {
const unsubState = window.electronAPI.voiceConversation.onStateChanged((info) => {
setState(info.state)
setMessages(info.messages)
setIsActive(info.isActive)
setLocalState(info.state)
setLocalMessages(info.messages)
setLocalActive(info.isActive)
})
const unsubUser = window.electronAPI.voiceConversation.onUserMessage((msg) => {
setMessages((prev) => [...prev, msg])
setStreamingText('')
setStreamingMsgId(null)
setLocalMessages((prev) => [...prev, msg])
setLocalStreamingText('')
setLocalStreamingMsgId(null)
setTimeout(scrollToBottom, 50)
})
const unsubDelta = window.electronAPI.voiceConversation.onAssistantDelta((data: ConversationAssistantDelta) => {
setStreamingMsgId(data.messageId)
setStreamingText(data.accumulated)
setLocalStreamingMsgId(data.messageId)
setLocalStreamingText(data.accumulated)
setTimeout(scrollToBottom, 50)
})
const unsubComplete = window.electronAPI.voiceConversation.onAssistantMessage((msg) => {
setMessages((prev) => [...prev, { id: msg.messageId, role: 'assistant', content: msg.content, timestamp: Date.now() }])
setStreamingText('')
setStreamingMsgId(null)
setLocalMessages((prev) => [...prev, { id: msg.messageId, role: 'assistant', content: msg.content, timestamp: Date.now() }])
setLocalStreamingText('')
setLocalStreamingMsgId(null)
setTimeout(scrollToBottom, 50)
})
const unsubError = window.electronAPI.voiceConversation.onError((err: ConversationError) => {
@ -69,9 +97,9 @@ export function VoiceConversationPage(): React.ReactElement {
// 초기 상태 로드
window.electronAPI.voiceConversation.getState().then((r) => {
if (r.success) {
setState(r.data.state)
setMessages(r.data.messages)
setIsActive(r.data.isActive)
setLocalState(r.data.state)
setLocalMessages(r.data.messages)
setLocalActive(r.data.isActive)
}
})
@ -84,36 +112,78 @@ export function VoiceConversationPage(): React.ReactElement {
}
}, [scrollToBottom])
// Realtime 메시지 갱신 시 스크롤
useEffect(() => {
if (isRealtime) setTimeout(scrollToBottom, 50)
}, [isRealtime, realtime.messages, realtime.streamingText, scrollToBottom])
const handleStartSession = useCallback(async () => {
if (isRealtime) {
const ok = await realtime.start()
if (!ok) {
// Realtime 연결 실패 → 로컬 파이프라인 자동 fallback
setBackend('local')
setErrorBanner({ phase: 'llm', message: t('conversation.realtime.fallback') })
await window.electronAPI.voiceConversation.startSession()
}
return
}
await window.electronAPI.voiceConversation.startSession()
}, [])
}, [isRealtime, realtime, t])
const handleStopSession = useCallback(async () => {
if (isRealtime) {
realtime.stop()
return
}
await window.electronAPI.voiceConversation.stopSession()
}, [])
}, [isRealtime, realtime])
const handleFinishListening = useCallback(async () => {
if (isRealtime) return // server VAD가 자동 처리
await window.electronAPI.voiceConversation.finishListening()
}, [])
}, [isRealtime])
const handleSendText = useCallback(async () => {
if (!textInput.trim()) return
const text = textInput.trim()
setTextInput('')
if (isRealtime) {
if (!realtime.isActive) {
const ok = await realtime.start()
if (!ok) {
setBackend('local')
setErrorBanner({ phase: 'llm', message: t('conversation.realtime.fallback') })
await window.electronAPI.voiceConversation.startSession()
await window.electronAPI.voiceConversation.sendMessage({ text })
return
}
}
realtime.sendText(text)
return
}
if (!isActive) {
await window.electronAPI.voiceConversation.startSession()
}
await window.electronAPI.voiceConversation.sendMessage({ text })
}, [textInput, isActive])
}, [textInput, isActive, isRealtime, realtime, t])
const handleClearHistory = useCallback(async () => {
if (isRealtime) {
realtime.clearMessages()
return
}
await window.electronAPI.voiceConversation.clearHistory()
setMessages([])
}, [])
setLocalMessages([])
}, [isRealtime, realtime])
const handleCancelResponse = useCallback(async () => {
if (isRealtime) {
realtime.cancelResponse()
return
}
await window.electronAPI.voiceConversation.cancelResponse()
}, [])
}, [isRealtime, realtime])
const handleTextKeyDown = useCallback((e: React.KeyboardEvent) => {
if (isImeComposingEvent(e)) return
@ -130,6 +200,11 @@ export function VoiceConversationPage(): React.ReactElement {
speaking: t('conversation.speaking'),
}
const displayStateLabel =
isRealtime && realtime.connectionState === 'connecting'
? t('conversation.realtime.connecting')
: stateLabel[state]
const formatErrorMessage = useCallback((err: ConversationError): string => {
// Bug 13의 전형적 STT 메시지는 i18n 키로 매핑 — 그 외는 서비스 원문 + phase label.
if (err.phase === 'stt' && err.message.toLowerCase().startsWith('no speech')) {
@ -152,9 +227,23 @@ export function VoiceConversationPage(): React.ReactElement {
title={t('conversation.title').toUpperCase()}
action={
<Box sx={{ display: 'flex', gap: 1, alignItems: 'center' }}>
{isRealtime && (
<PhosphorText
variant="micro"
sx={{
color: d3roPalette.accent.amber,
border: `1px solid ${d3roPalette.accent.amber}`,
borderRadius: '3px',
px: 0.5,
lineHeight: 1.6,
}}
>
{t('conversation.realtime.badge')}
</PhosphorText>
)}
<Led color={stateLedColor[state]} pulse={state === 'listening' || state === 'thinking'} size={8} />
<PhosphorText variant="label" sx={{ color: d3roPalette.text.dimLabel }}>
{stateLabel[state].toUpperCase()}
{displayStateLabel.toUpperCase()}
</PhosphorText>
{messages.length > 0 && (
<Tooltip title={t('conversation.clearHistory')}>
@ -167,8 +256,8 @@ export function VoiceConversationPage(): React.ReactElement {
}
/>
{/* 메시지 영역 — state 분기: listening=몰입 패널, 그 외=메시지 리스트 */}
{state === 'listening' ? (
{/* 메시지 영역 — 로컬 파이프라인의 listening만 몰입 패널 (Realtime은 항상 채팅 뷰) */}
{!isRealtime && state === 'listening' ? (
<Box
sx={{
flex: 1,
@ -292,7 +381,11 @@ export function VoiceConversationPage(): React.ReactElement {
<MicIcon sx={{ fontSize: 20 }} />
</PhysicalButton>
) : state === 'listening' ? (
<PhysicalButton selected onClick={handleFinishListening} sx={{ minWidth: 48, px: 2 }}>
<PhysicalButton
selected
onClick={isRealtime ? handleStopSession : handleFinishListening}
sx={{ minWidth: 48, px: 2 }}
>
<StopIcon sx={{ fontSize: 20 }} />
</PhysicalButton>
) : state === 'thinking' || state === 'speaking' ? (
@ -305,8 +398,8 @@ export function VoiceConversationPage(): React.ReactElement {
</PhysicalButton>
)}
{/* 텍스트 입력 — listening 시 숨김, thinking/speaking 시 disabled */}
{state !== 'listening' && (
{/* 텍스트 입력 — 로컬 listening 시 숨김 (Realtime은 항상 표시), thinking/speaking 시 disabled */}
{(isRealtime || state !== 'listening') && (
<TextField
value={state === 'thinking' || state === 'speaking' ? '' : textInput}
onChange={(e) => setTextInput(e.target.value)}
@ -335,8 +428,8 @@ export function VoiceConversationPage(): React.ReactElement {
/>
)}
{/* 전송 버튼 — listening 시 숨김 */}
{state !== 'listening' && (
{/* 전송 버튼 — 로컬 listening 시 숨김 */}
{(isRealtime || state !== 'listening') && (
<IconButton
onClick={handleSendText}
disabled={!textInput.trim() || state === 'thinking' || state === 'speaking'}