natural voice

.env.example

@@ -25,6 +25,21 @@ CDN_DIR=./cdn
 OCR_LANG=eng
 OCR_MAX_CHARS=100000
 
+# --- platform tools: neural TTS (first configured wins, TTS_PROVIDER forces one) ---
+# ElevenLabs (most natural): https://elevenlabs.io/app/settings/api-keys
+#ELEVENLABS_API_KEY=
+#ELEVENLABS_VOICE=
+# OpenAI (reuses the LLM key when set, tts-1 + alloy/echo/fable/onyx/nova/shimmer)
+#OPENAI_API_KEY=
+#TTS_VOICE=alloy
+# Local piper (offline, like tesseract): PIPER_BIN + onnx model
+#PIPER_BIN=piper
+#PIPER_MODEL=/var/lib/kisassistant/piper/en_US-lessac-medium.onnx
+# Test/development overrides (base URLs, default = api.elevenlabs.io / api.openai.com)
+#ELEVENLABS_BASE_URL=
+#OPENAI_BASE_URL=
+#TTS_PROVIDER=auto
+
 # Per-driver debug logging to stdout ([driver:name] lines).
 DRIVER_DEBUG=true
 

README.md

@@ -68,6 +68,8 @@ from the **Accounts** panel (each gets a one-time access key to hand over).
 | GET | `/api/admin/ollama/nodes/:id/status`, `…/stats?window=` | admin | live probe (version, running, VRAM) / load charts 1h–30d |
 | GET | `/api/admin/ollama/probe?host=` | admin | test a host before registering it |
 | GET | `/api/admin/platform` | admin | this host facts: hostname, kernel, uptime, load, memory, IPs |
+| GET | `/api/tools/tts/info` | Bearer | neural TTS provider + voices (null = browser fallback) |
+| POST | `/api/tools/tts` | Bearer | synthesize `{text, voice?}` → audio bytes (cached, 30/hour) |
 | GET | `/api/admin/ollama/status` | admin | Ollama connectivity probe: reachability, version, on-disk models |
 | POST | `/api/access/request` | public | reserve username+email with a motivation → ticket (`/review/:id`) |
 | GET/POST | `/api/access/ticket/:id` (+`/message`) | ticket bearer | follow + reply on an access request |

docs/FRONTEND.md

@@ -92,7 +92,9 @@ No router library — a tiny history wrapper (`lib/route.ts`: `navigate()`,
   also repaints the SVG favicon (data URL) so the tab icon matches the dress.
 - Voice (`lib/voice.ts`, no dependency): `Chat.tsx` composer mic with live
   waveform + speech-recognition transcript (editable before send), per-bubble
-  TTS playback with glow, `Settings → Voice` prefs (voice/lang/speed/pitch,
+  playback with glow through a unified player (`playText`: neural server
+  audio via `POST /api/tools/tts` with browser-synthesis fallback),
+  `Settings → Voice` prefs (engine/lang/voice/speed/pitch/neural voice,
   localStorage), admin-editable preview lines.
 
 ## Theming

docs/OPERATIONS.md

@@ -193,15 +193,31 @@ Disk meters are intentionally absent (models live on the Ollama hosts).
   it away. Unsupported browsers get a disabled mic with an explanatory
   tooltip. No audio leaves the browser (no server STT endpoint yet —
   whisper.cpp via `WHISPER_BIN` is the documented next step).
-- **Message playback**: every assistant bubble has a Listen button
+- **Message playback**: every assistant bubble has a Listen button with a
-  (speechSynthesis, Siri-style breathing glow while speaking, click again to
+  Siri-style breathing glow while speaking (click again to stop; playback
-  stop; playback stops on send / new message / conversation switch).
+  stops on send / new message / conversation switch). Engine per user
-- **Settings → Voice** (per browser, localStorage): language (auto-detect
+  (Settings → Voice): **auto** (neural server voice when available, browser
-  per message via `detectLang`, or forced), voice picker grouped by locale,
+  fallback), **browser**, or **server**.
-  speed, pitch, and one-tap previews of the admin-set lines.
+- **Neural server TTS** (`POST /api/tools/tts` `{text, voice?}` → audio
+  bytes, `GET /api/tools/tts/info` → provider + voices): ElevenLabs
+  (any voice_id, most natural) → OpenAI `tts-1` (6 stock voices, reuses the
+  LLM key) → local piper (`PIPER_BIN` + `PIPER_MODEL`, offline WAV).
+  `TTS_PROVIDER` forces one; `TTS_VOICE` / `ELEVENLABS_VOICE` set defaults.
+  Output is cached on disk (sha of provider|voice|text, 200 files cap),
+  rate-limited to 30 jobs/hour/account (429), kill-switched via
+  `tools_tts_enabled` (Admin Center → Features). Test overrides:
+  `ELEVENLABS_BASE_URL` / `OPENAI_BASE_URL`.
+- **Settings → Voice** (per browser, localStorage): engine, language
+  (auto-detect per message or forced), browser voice/rate/pitch, neural
+  voice id, and one-tap previews of the admin-set lines.
 - **Admin Center → Features → Voice previews**: the default playback lines
-  (`voice_previews` setting, 1–6 lines; empty list resets to the built-ins),
+  (`voice_previews` setting, 1–6 lines `{text, lang, rate, pitch, timbre,
-  also served to users via `GET /api/tools`.
+  voice}`, each with its OWN voice character — masculine/feminine/any
+  timbre hint, speed, pitch, optional server voice id — so previews sound
+  different from each other; timbre matching is best-effort over the
+  browser's voice names, rate/pitch always apply; empty list resets to the
+  built-ins), also served to users via `GET /api/tools`. Users can adopt
+  any preview's character as their own playback voice from Settings → Voice.
 
 ## Personal providers (BYOK)
 

frontend/src/components/AdminCenter.tsx

@@ -321,6 +321,13 @@ export function AdminCenter({ initialTab }: { initialTab?: string | null }) {
               </div>
               <Switch label={t("center.usernameChangeTitle")} checked={settings?.usernameChangeEnabled ?? false} onChange={(v) => toggle({ usernameChangeEnabled: v })} />
             </div>
+            <div className="flex items-center gap-4 rounded-3xl border border-stone-200/70 bg-white p-5 shadow-sm dark:border-zinc-800 dark:bg-zinc-900">
+              <div className="min-w-0 flex-1">
+                <p className="text-sm font-medium">{t("center.ttsTitle")}</p>
+                <p className="text-sm opacity-60">{t("center.ttsDesc")}</p>
+              </div>
+              <Switch label={t("center.ttsTitle")} checked={settings?.ttsEnabled ?? false} onChange={(v) => toggle({ ttsEnabled: v })} />
+            </div>
             {saving && <p className="flex items-center gap-2 text-sm opacity-60"><Spinner size={14} /> {t("common.saving")}</p>}
             {!settings && !error && <p className="flex items-center gap-2 text-sm opacity-60"><Spinner size={14} /> {t("common.loading")}</p>}
             <VoicePreviewsCard />
@@ -444,6 +451,20 @@ function VoicePreviewsCard() {
                   />
                 </Field>
               </span>
+              <span className="w-40">
+                <Field label={t("vfeat.voice")} hint={t("vfeat.voiceHint")}>
+                  <Input
+                    value={line.voice}
+                    disabled={busy}
+                    maxLength={64}
+                    placeholder={t("vfeat.voicePh")}
+                    aria-label={t("vfeat.voice")}
+                    onChange={(e) => patch(i, { voice: e.target.value })}
+                    onBlur={() => commit(i)}
+                    onKeyDown={(e) => { if (e.key === "Enter") (e.target as HTMLInputElement).blur(); }}
+                  />
+                </Field>
+              </span>
               <label className="min-w-32 flex-1 text-xs opacity-80">
                 <span className="mb-1 block font-medium">{t("voice.rate")} · {line.rate.toFixed(2)}×</span>
                 <input
@@ -478,7 +499,7 @@ function VoicePreviewsCard() {
             variant="secondary"
             size="sm"
             disabled={busy}
-            onClick={() => setLines((prev) => [...prev, { text: "", lang: "auto", rate: 1, pitch: 1, timbre: "any" as const }])}
+            onClick={() => setLines((prev) => [...prev, { text: "", lang: "auto", rate: 1, pitch: 1, timbre: "any" as const, voice: "" }])}
           >
             <Plus size={14} /> {t("vfeat.add")}
           </Button>

frontend/src/components/Chat.tsx

@@ -7,7 +7,7 @@ import { Markdown } from "./Markdown.tsx";
 import { UserMessageContent } from "./MessageContent.tsx";
 import { cn } from "../lib/cn.ts";
 import { useFeatures } from "../lib/features.ts";
-import { fmtDuration, getVoicePrefs, listTtsVoices, recognizerCtor, speak, stopSpeak, sttSupported, type Recognizer } from "../lib/voice.ts";
+import { fmtDuration, getVoicePrefs, listTtsVoices, playText, recognizerCtor, stopPlayback, sttSupported, type Recognizer } from "../lib/voice.ts";
 import { useT, type StringKey } from "../lib/i18n.ts";
 
 const REPORT_REASONS: ReportReason[] = ["copyright", "gore", "falseinfo", "bug"];
@@ -279,6 +279,7 @@ export function Chat({
   // TTS playback: which assistant bubble is speaking (null = silent).
   const [speakingKey, setSpeakingKey] = useState<number | null>(null);
   const voicesRef = useRef<SpeechSynthesisVoice[]>([]);
+  const ttsServerRef = useRef(false);
   // Voice input: live recording session (null = composer idle).
   const [rec, setRec] = useState<{
     stream: MediaStream;
@@ -296,12 +297,13 @@ export function Chat({
 
   useEffect(() => {
     listTtsVoices().then((v) => { voicesRef.current = v; });
-    return () => stopSpeak();
+    api.ttsInfo().then((r) => { ttsServerRef.current = r.available; }).catch(() => {});
+    return () => stopPlayback();
   }, []);
 
   // New message / conversation: stop any playback (never leak speech).
   useEffect(() => {
-    stopSpeak();
+    stopPlayback();
     setSpeakingKey(null);
     // eslint-disable-next-line react-hooks/exhaustive-deps
   }, [conv?.id]);
@@ -309,22 +311,29 @@ export function Chat({
   useEffect(() => {
     return () => {
       cancelRec(false);
-      stopSpeak();
+      stopPlayback();
     };
     // eslint-disable-next-line react-hooks/exhaustive-deps
   }, []);
 
   function toggleSpeak(index: number, text: string) {
     if (speakingKey === index) {
-      stopSpeak();
+      stopPlayback();
       setSpeakingKey(null);
       return;
     }
-    const ok = speak(text, getVoicePrefs(), voicesRef.current, {
+    const prefs = getVoicePrefs();
+    playText({
+      text,
+      engine: prefs.engine,
+      browser: { voices: voicesRef.current, lang: prefs.lang, rate: prefs.rate, pitch: prefs.pitch, voiceURI: prefs.voiceURI },
+      serverVoice: prefs.serverVoice ?? undefined,
+      serverAvailable: ttsServerRef.current,
+      fetchAudio: (t, v) => api.ttsSpeak(t, v),
       onend: () => setSpeakingKey(null),
       onerror: () => setSpeakingKey(null),
     });
-    if (ok) setSpeakingKey(index);
+    setSpeakingKey(index);
   }
 
   function cleanupRec(r: NonNullable<typeof rec>) {
@@ -446,7 +455,7 @@ export function Chat({
   // A fresh message ends any playback (the glow never sticks to a stale row).
   const msgCount = messages.length;
   useEffect(() => {
-    stopSpeak();
+    stopPlayback();
     setSpeakingKey(null);
     // eslint-disable-next-line react-hooks/exhaustive-deps
   }, [msgCount]);
@@ -464,7 +473,7 @@ export function Chat({
     if (readOnly) return;
     const text = draft.trim();
     if (!text || sending) return;
-    stopSpeak();
+    stopPlayback();
     setSpeakingKey(null);
     setDraft("");
     requestAnimationFrame(autoGrow);

frontend/src/components/dialogs.tsx

@@ -4,7 +4,7 @@ import { api, type UserProvider, type VoicePreview } from "../lib/api.ts";
 import { pushToast } from "../lib/toasts.ts";
 import { cn } from "../lib/cn.ts";
 import { ACCENT_PRESETS, applyAccent, currentAccent, previewAccent, type AccentState } from "../lib/accent.ts";
-import { detectLang, getVoicePrefs, listTtsVoices, pickVoice, setVoicePrefs, speakPreview, stopSpeak, ttsSupported, type VoicePrefs } from "../lib/voice.ts";
+import { detectLang, getVoicePrefs, listTtsVoices, pickVoice, playText, setVoicePrefs, stopPlayback, ttsSupported, type VoicePrefs } from "../lib/voice.ts";
 import { FEATURES, setFeature, useFeatures } from "../lib/features.ts";
 import type { Conversation, FilePreview, User } from "../lib/types.ts";
 import { useT, type StringKey } from "../lib/i18n.ts";
@@ -877,12 +877,14 @@ function VoiceSection() {
   const [voices, setVoices] = useState<SpeechSynthesisVoice[]>([]);
   const [previews, setPreviews] = useState<VoicePreview[]>([]);
   const [playing, setPlaying] = useState<number | null>(null);
+  const [ttsServer, setTtsServer] = useState<{ available: boolean; provider: string | null; voices: string[] } | null>(null);
 
   useEffect(() => {
     let alive = true;
     listTtsVoices().then((v) => { if (alive) setVoices(v); });
     api.tools().then((r) => { if (alive) setPreviews(r.voicePreviews); }).catch(() => {});
-    return () => { alive = false; stopSpeak(); };
+    api.ttsInfo().then((r) => { if (alive) setTtsServer({ available: r.available, provider: r.provider, voices: r.voices }); }).catch(() => {});
+    return () => { alive = false; stopPlayback(); };
   }, []);
 
   function update(p: Partial<VoicePrefs>) {
@@ -893,17 +895,23 @@ function VoiceSection() {
 
   function preview(i: number, line: VoicePreview) {
     if (playing === i) {
-      stopSpeak();
+      stopPlayback();
       setPlaying(null);
       return;
     }
-    // Each line plays with ITS OWN character (timbre/rate/pitch/lang) —
+    // Each line plays with ITS OWN character; the neural server engine is
-    // never the global playback prefs.
+    // used when selected and available (line.server voice or user default).
-    const pick = speakPreview(line.text, line, voices, {
+    playText({
+      text: line.text,
+      engine: prefs.engine,
+      browser: { voices, lang: line.lang, rate: line.rate, pitch: line.pitch, timbre: line.timbre },
+      serverVoice: line.voice || prefs.serverVoice || undefined,
+      serverAvailable: ttsServer?.available ?? false,
+      fetchAudio: (text, voice) => api.ttsSpeak(text, voice),
       onend: () => setPlaying(null),
       onerror: () => setPlaying(null),
     });
-    if (pick || ttsSupported()) setPlaying(i);
+    setPlaying(i);
   }
 
   /** Adopt a preview's character as the global playback voice. */
@@ -911,6 +919,7 @@ function VoiceSection() {
     const lang = line.lang === "auto" ? detectLang(line.text) : line.lang;
     const pick = pickVoice(voices, { lang, timbre: line.timbre });
     const next: VoicePrefs = {
+      ...prefs,
       voiceURI: pick?.voiceURI ?? prefs.voiceURI,
       rate: line.rate,
       pitch: line.pitch,
@@ -954,6 +963,49 @@ function VoiceSection() {
           options={VOICE_LANGS.map((l) => ({ value: l, label: l === "auto" ? t("voice.auto") : l }))}
         />
       </Field>
+      <Field label={t("voice.engine")} hint={ttsServer?.available ? t("voice.engineHintOn", { provider: ttsServer.provider ?? "?" }) : t("voice.engineHintOff")}>
+        <Picker
+          ariaLabel={t("voice.engine")}
+          icon={Volume2}
+          value={prefs.engine}
+          onChange={(v) => update({ engine: v as VoicePrefs["engine"] })}
+          align="left"
+          options={[
+            { value: "auto", label: t("voice.engineAuto") },
+            { value: "browser", label: t("voice.engineBrowser") },
+            { value: "server", label: t("voice.engineServer") },
+          ]}
+        />
+      </Field>
+      {(prefs.engine !== "browser") && (
+        ttsServer?.available ? (
+          ttsServer.provider === "openai" ? (
+            <Field label={t("voice.serverVoice")}>
+              <Picker
+                ariaLabel={t("voice.serverVoice")}
+                icon={Volume2}
+                value={prefs.serverVoice ?? ""}
+                onChange={(v) => update({ serverVoice: v || null })}
+                align="left"
+                options={[{ value: "", label: t("voice.defaultVoice") }, ...ttsServer.voices.map((v) => ({ value: v, label: v }))]}
+              />
+            </Field>
+          ) : ttsServer.provider === "elevenlabs" ? (
+            <Field label={t("voice.serverVoice")} hint={t("voice.serverVoiceHint")}>
+              <Input
+                value={prefs.serverVoice ?? ""}
+                onChange={(e) => update({ serverVoice: e.target.value.trim() || null })}
+                placeholder="pNInz6obpgDQGcFmaJgB"
+                maxLength={64}
+              />
+            </Field>
+          ) : (
+            <p className="rounded-2xl bg-stone-100 px-4 py-2.5 text-sm opacity-70 dark:bg-zinc-800">{t("voice.piperNote")}</p>
+          )
+        ) : (
+          <p className="rounded-2xl bg-amber-500/10 px-4 py-2.5 text-sm text-amber-700 dark:text-amber-400">{t("voice.noServer")}</p>
+        )
+      )}
       <Field label={t("voice.voice")}>
         <Picker
           ariaLabel={t("voice.voice")}

frontend/src/lib/api.ts

@@ -74,6 +74,7 @@ export interface VoicePreview {
   rate: number;
   pitch: number;
   timbre: "masculine" | "feminine" | "any";
+  voice: string;
 }
 
 export interface PlatformInfo {
@@ -102,6 +103,7 @@ export interface AdminSettings {
   recoveryLimitMax: number;
   recoveryLimitWindowMin: number;
   voicePreviews: VoicePreview[];
+  ttsEnabled: boolean;
 }
 
 export interface AdminSettingsPatch {
@@ -115,6 +117,7 @@ export interface AdminSettingsPatch {
   recoveryLimitMax?: number;
   recoveryLimitWindowMin?: number;
   voicePreviews?: VoicePreview[];
+  ttsEnabled?: boolean;
 }
 
 export interface OllamaNode {
@@ -442,6 +445,27 @@ export const api = {
 
   tools: () => req<{ tools: { name: string; description: string; available: boolean; reason: string | null }[]; reportsEnabled: boolean; usernameChangeEnabled: boolean; voicePreviews: VoicePreview[] }>("/api/tools"),
 
+  /** Neural TTS provider status (null = browser speech fallback). */
+  ttsInfo: () => req<{ provider: "elevenlabs" | "openai" | "piper" | null; voices: string[]; available: boolean; reason: string | null }>("/api/tools/tts/info"),
+
+  /** Synthesize text server-side (neural voice). Returns audio bytes. */
+  ttsSpeak: async (text: string, voice?: string): Promise<Blob> => {
+    const res = await fetch("/api/tools/tts", {
+      method: "POST",
+      headers: { "Content-Type": "application/json", Authorization: `Bearer ${getToken()}` },
+      body: JSON.stringify({ text, ...(voice ? { voice } : {}) }),
+    });
+    if (res.status === 401) {
+      clearToken();
+      window.location.reload();
+      throw new ApiError(401, "Session expired.");
+    }
+    if (!res.ok) {
+      const data = (await res.json().catch(() => ({}))) as { error?: string };
+      throw new ApiError(res.status, data.error || `TTS failed (${res.status})`);
+    }
+    return await res.blob();
+  },
   /** Send an image to the platform OCR tool. Returns editable text (never the raw image). */
   ocrImage: async (file: File): Promise<{ text: string; truncated: boolean; chars: number }> => {
     const res = await fetch("/api/tools/ocr", {

frontend/src/lib/i18n.ts

@@ -605,6 +605,22 @@ const en = {
   "vfeat.timbre": "Timbre",
   "vfeat.rate": "Speed",
   "vfeat.pitch": "Pitch",
+  "vfeat.voice": "Server voice",
+  "vfeat.voiceHint": "ElevenLabs id / OpenAI name, empty = default",
+  "vfeat.voicePh": "auto",
+  "center.ttsTitle": "Neural voice playback",
+  "center.ttsDesc": "Server text-to-speech for message playback (ElevenLabs / OpenAI / piper). Off = browser speech only.",
+  "voice.engine": "Playback engine",
+  "voice.engineAuto": "Auto (neural if available)",
+  "voice.engineBrowser": "Browser speech",
+  "voice.engineServer": "Neural server voice",
+  "voice.engineHintOn": "Neural TTS via {provider}",
+  "voice.engineHintOff": "No neural provider configured — browser speech only",
+  "voice.serverVoice": "Neural voice",
+  "voice.serverVoiceHint": "ElevenLabs voice id",
+  "voice.piperNote": "Local piper voice (single configured model).",
+  "voice.noServer": "Configure ELEVENLABS_API_KEY, OPENAI_API_KEY or piper to unlock neural voices.",
+
 
 
   "msg.copy": "Copy message",
@@ -1297,6 +1313,22 @@ const es: Dict = {
   "vfeat.timbre": "Timbre",
   "vfeat.rate": "Speed",
   "vfeat.pitch": "Pitch",
+  "vfeat.voice": "Server voice",
+  "vfeat.voiceHint": "ElevenLabs id / OpenAI name, empty = default",
+  "vfeat.voicePh": "auto",
+  "center.ttsTitle": "Neural voice playback",
+  "center.ttsDesc": "Server text-to-speech for message playback (ElevenLabs / OpenAI / piper). Off = browser speech only.",
+  "voice.engine": "Playback engine",
+  "voice.engineAuto": "Auto (neural if available)",
+  "voice.engineBrowser": "Browser speech",
+  "voice.engineServer": "Neural server voice",
+  "voice.engineHintOn": "Neural TTS via {provider}",
+  "voice.engineHintOff": "No neural provider configured — browser speech only",
+  "voice.serverVoice": "Neural voice",
+  "voice.serverVoiceHint": "ElevenLabs voice id",
+  "voice.piperNote": "Local piper voice (single configured model).",
+  "voice.noServer": "Configure ELEVENLABS_API_KEY, OPENAI_API_KEY or piper to unlock neural voices.",
+
 
 
   "msg.copy": "Copiar mensaje",
@@ -1984,6 +2016,22 @@ const fr: Dict = {
   "vfeat.timbre": "Timbre",
   "vfeat.rate": "Speed",
   "vfeat.pitch": "Pitch",
+  "vfeat.voice": "Server voice",
+  "vfeat.voiceHint": "ElevenLabs id / OpenAI name, empty = default",
+  "vfeat.voicePh": "auto",
+  "center.ttsTitle": "Neural voice playback",
+  "center.ttsDesc": "Server text-to-speech for message playback (ElevenLabs / OpenAI / piper). Off = browser speech only.",
+  "voice.engine": "Playback engine",
+  "voice.engineAuto": "Auto (neural if available)",
+  "voice.engineBrowser": "Browser speech",
+  "voice.engineServer": "Neural server voice",
+  "voice.engineHintOn": "Neural TTS via {provider}",
+  "voice.engineHintOff": "No neural provider configured — browser speech only",
+  "voice.serverVoice": "Neural voice",
+  "voice.serverVoiceHint": "ElevenLabs voice id",
+  "voice.piperNote": "Local piper voice (single configured model).",
+  "voice.noServer": "Configure ELEVENLABS_API_KEY, OPENAI_API_KEY or piper to unlock neural voices.",
+
 
 
   "msg.copy": "Copier le message",
@@ -2659,6 +2707,22 @@ const ru: Dict = {
   "vfeat.timbre": "Timbre",
   "vfeat.rate": "Speed",
   "vfeat.pitch": "Pitch",
+  "vfeat.voice": "Server voice",
+  "vfeat.voiceHint": "ElevenLabs id / OpenAI name, empty = default",
+  "vfeat.voicePh": "auto",
+  "center.ttsTitle": "Neural voice playback",
+  "center.ttsDesc": "Server text-to-speech for message playback (ElevenLabs / OpenAI / piper). Off = browser speech only.",
+  "voice.engine": "Playback engine",
+  "voice.engineAuto": "Auto (neural if available)",
+  "voice.engineBrowser": "Browser speech",
+  "voice.engineServer": "Neural server voice",
+  "voice.engineHintOn": "Neural TTS via {provider}",
+  "voice.engineHintOff": "No neural provider configured — browser speech only",
+  "voice.serverVoice": "Neural voice",
+  "voice.serverVoiceHint": "ElevenLabs voice id",
+  "voice.piperNote": "Local piper voice (single configured model).",
+  "voice.noServer": "Configure ELEVENLABS_API_KEY, OPENAI_API_KEY or piper to unlock neural voices.",
+
 
 
   "msg.copy": "Скопировать сообщение",
@@ -3339,6 +3403,22 @@ const it: Dict = {
   "vfeat.timbre": "Timbre",
   "vfeat.rate": "Speed",
   "vfeat.pitch": "Pitch",
+  "vfeat.voice": "Server voice",
+  "vfeat.voiceHint": "ElevenLabs id / OpenAI name, empty = default",
+  "vfeat.voicePh": "auto",
+  "center.ttsTitle": "Neural voice playback",
+  "center.ttsDesc": "Server text-to-speech for message playback (ElevenLabs / OpenAI / piper). Off = browser speech only.",
+  "voice.engine": "Playback engine",
+  "voice.engineAuto": "Auto (neural if available)",
+  "voice.engineBrowser": "Browser speech",
+  "voice.engineServer": "Neural server voice",
+  "voice.engineHintOn": "Neural TTS via {provider}",
+  "voice.engineHintOff": "No neural provider configured — browser speech only",
+  "voice.serverVoice": "Neural voice",
+  "voice.serverVoiceHint": "ElevenLabs voice id",
+  "voice.piperNote": "Local piper voice (single configured model).",
+  "voice.noServer": "Configure ELEVENLABS_API_KEY, OPENAI_API_KEY or piper to unlock neural voices.",
+
 
 
   "msg.copy": "Copia messaggio",
@@ -4010,6 +4090,22 @@ const pt: Dict = {
   "vfeat.timbre": "Timbre",
   "vfeat.rate": "Speed",
   "vfeat.pitch": "Pitch",
+  "vfeat.voice": "Server voice",
+  "vfeat.voiceHint": "ElevenLabs id / OpenAI name, empty = default",
+  "vfeat.voicePh": "auto",
+  "center.ttsTitle": "Neural voice playback",
+  "center.ttsDesc": "Server text-to-speech for message playback (ElevenLabs / OpenAI / piper). Off = browser speech only.",
+  "voice.engine": "Playback engine",
+  "voice.engineAuto": "Auto (neural if available)",
+  "voice.engineBrowser": "Browser speech",
+  "voice.engineServer": "Neural server voice",
+  "voice.engineHintOn": "Neural TTS via {provider}",
+  "voice.engineHintOff": "No neural provider configured — browser speech only",
+  "voice.serverVoice": "Neural voice",
+  "voice.serverVoiceHint": "ElevenLabs voice id",
+  "voice.piperNote": "Local piper voice (single configured model).",
+  "voice.noServer": "Configure ELEVENLABS_API_KEY, OPENAI_API_KEY or piper to unlock neural voices.",
+
 
 
   "msg.copy": "Copiar mensagem",

frontend/src/lib/voice.ts

@@ -15,21 +15,28 @@ export interface VoicePrefs {
   pitch: number;
   /** "auto" detects per message, otherwise forces a BCP-47 prefix. */
   lang: string;
+  /** Playback engine: neural server TTS when available, else browser. */
+  engine: "auto" | "browser" | "server";
+  /** Server neural voice id (ElevenLabs voice_id / OpenAI voice name). */
+  serverVoice: string | null;
 }
 
 const PREFS_KEY = "ka:voice";
 
-export const DEFAULT_VOICE_PREFS: VoicePrefs = { voiceURI: null, rate: 1, pitch: 1, lang: "auto" };
+export const DEFAULT_VOICE_PREFS: VoicePrefs = { voiceURI: null, rate: 1, pitch: 1, lang: "auto", engine: "auto", serverVoice: null };
 
 export function getVoicePrefs(): VoicePrefs {
   try {
     const raw = JSON.parse(localStorage.getItem(PREFS_KEY) ?? "null") as Partial<VoicePrefs> | null;
     if (!raw || typeof raw !== "object") return { ...DEFAULT_VOICE_PREFS };
+    const engine = raw.engine === "browser" || raw.engine === "server" ? raw.engine : "auto";
     return {
       voiceURI: typeof raw.voiceURI === "string" ? raw.voiceURI : null,
       rate: typeof raw.rate === "number" && raw.rate >= 0.5 && raw.rate <= 2 ? raw.rate : 1,
       pitch: typeof raw.pitch === "number" && raw.pitch >= 0.5 && raw.pitch <= 2 ? raw.pitch : 1,
       lang: typeof raw.lang === "string" && raw.lang.length <= 12 ? raw.lang : "auto",
+      engine,
+      serverVoice: typeof raw.serverVoice === "string" && raw.serverVoice.length > 0 ? raw.serverVoice : null,
     };
   } catch {
     return { ...DEFAULT_VOICE_PREFS };
@@ -83,13 +90,6 @@ export interface SpeakOpts {
 
 export type VoiceTimbre = "masculine" | "feminine" | "any";
 
-export interface PreviewVoice {
-  lang: string;
-  rate: number;
-  pitch: number;
-  timbre: VoiceTimbre;
-}
-
 // Name fragments hinting a masculine / feminine voice (Chrome, Edge,
 // Safari and Android voices carry such markers; absence falls back to
 // any same-language voice — never an error).
@@ -158,43 +158,163 @@ export function stopSpeak(): void {
   }
 }
 
-/**
+// --- unified playback (neural server audio with browser fallback) ---
- * Play a preview line with ITS OWN voice character (timbre + rate + pitch),
+
- * never the user's global playback prefs — previews must sound different
+export interface PlayHandle {
- * from each other. Returns the resolved voice (null when unsupported) so
+  /** Engine that ended up playing (may differ from the request in auto). */
- * callers can offer "use this voice" adoption.
+  engine: "server" | "browser";
- */
+  stop: () => void;
-export function speakPreview(
+  done: Promise<void>;
-  text: string,
+}
-  preview: PreviewVoice,
+
-  voices: SpeechSynthesisVoice[],
+let currentAudio: HTMLAudioElement | null = null;
-  opts: SpeakOpts = {},
+
-): SpeechSynthesisVoice | null {
+function stopCurrentAudio(): void {
-  if (!ttsSupported()) return null;
+  if (currentAudio) {
-  const synth = window.speechSynthesis;
     try {
-    synth.cancel();
+      currentAudio.pause();
+      const url = currentAudio.src;
+      currentAudio.removeAttribute("src");
+      currentAudio.load();
+      if (url.startsWith("blob:")) URL.revokeObjectURL(url);
     } catch {
       // ignore
     }
-  const lang = preview.lang === "auto" ? detectLang(text) : preview.lang;
+    currentAudio = null;
-  const utter = new SpeechSynthesisUtterance(text.slice(0, 2000));
+  }
-  utter.lang = lang;
-  utter.rate = preview.rate;
-  utter.pitch = preview.pitch;
-  const pick = pickVoice(voices, { lang, timbre: preview.timbre });
-  if (pick) utter.voice = pick;
-  utter.onend = () => opts.onend?.();
-  utter.onerror = () => opts.onerror?.("error");
   try {
-    synth.speak(utter);
+    window.speechSynthesis?.cancel();
   } catch {
-    return null;
+    // ignore
   }
-  return pick;
 }
 
-// --- speech recognition (STT) ---
+/** Stop whatever is playing (server audio or browser speech). */
+export function stopPlayback(): void {
+  stopCurrentAudio();
+}
+
+export interface PlayRequest {
+  text: string;
+  /** "server" forces neural, "browser" forces synthesis, "auto" tries neural first. */
+  engine: "auto" | "browser" | "server";
+  browser: {
+    voices: SpeechSynthesisVoice[];
+    /** "auto" detects per text, otherwise a BCP-47 prefix. */
+    lang: string;
+    rate: number;
+    pitch: number;
+    /** Explicit voice URI (user pref); otherwise timbre-matched. */
+    voiceURI?: string | null;
+    /** Preview character timbre (ignored when voiceURI is set). */
+    timbre?: VoiceTimbre;
+  };
+  /** Server neural voice id (preview line voice or user serverVoice). */
+  serverVoice?: string | null;
+  /** True when the platform TTS tool is available. */
+  serverAvailable: boolean;
+  fetchAudio: (text: string, voice?: string) => Promise<Blob>;
+  onstart?: () => void;
+  onend?: () => void;
+  onerror?: (msg: string) => void;
+}
+
+/**
+ * Play one text with the best engine: neural server audio when requested
+ * and available (shared <audio> element, user speed), otherwise the
+ * browser character voice. Single global playback — a new play stops the
+ * previous one. `done` resolves when playback fully ends (fallback
+ * included); it never rejects (errors go to onend/onerror like speech).
+ */
+export function playText(req: PlayRequest): PlayHandle {
+  stopPlayback();
+  const wantServer = req.serverAvailable && req.engine !== "browser";
+  if (wantServer) return playServer(req, req.engine === "auto");
+  return playBrowser(req);
+}
+
+function playServer(req: PlayRequest, allowFallback: boolean): PlayHandle {
+  let settled = false;
+  const stop = () => {
+    settled = true;
+    stopCurrentAudio();
+  };
+  const done = (async () => {
+    try {
+      const blob = await req.fetchAudio(req.text.slice(0, 2000), req.serverVoice ?? undefined);
+      if (settled) return;
+      await playAudioBlob(blob, req, () => settled);
+    } catch (e) {
+      if (settled) return;
+      if (allowFallback) {
+        await playBrowserAudio(req, () => settled);
+        return;
+      }
+      req.onerror?.(e instanceof Error ? e.message : "playback failed");
+    }
+  })();
+  return { engine: "server", stop, done };
+}
+
+function playAudioBlob(blob: Blob, req: PlayRequest, cancelled: () => boolean): Promise<void> {
+  return new Promise((resolve) => {
+    const url = URL.createObjectURL(blob);
+    const audio = new Audio(url);
+    audio.playbackRate = req.browser.rate;
+    currentAudio = audio;
+    const finish = (ok: boolean) => {
+      if (currentAudio === audio) stopCurrentAudio();
+      else {
+        try {
+          URL.revokeObjectURL(url);
+        } catch {
+          // ignore
+        }
+      }
+      if (!cancelled() && ok) req.onend?.();
+      resolve();
+    };
+    audio.onended = () => finish(true);
+    audio.onerror = () => finish(false);
+    req.onstart?.();
+    audio.play().catch(() => finish(false));
+  });
+}
+
+function playBrowser(req: PlayRequest): PlayHandle {
+  let settled = false;
+  const stop = () => {
+    settled = true;
+    stopSpeak();
+  };
+  const done = playBrowserAudio(req, () => settled);
+  return { engine: "browser", stop, done };
+}
+
+function playBrowserAudio(req: PlayRequest, cancelled: () => boolean): Promise<void> {
+  return new Promise((resolve) => {
+    const b = req.browser;
+    const lang = b.lang === "auto" ? detectLang(req.text) : b.lang;
+    const explicit = b.voiceURI ? b.voices.find((v) => v.voiceURI === b.voiceURI) : undefined;
+    const pick = explicit ?? pickVoice(b.voices, { lang, timbre: b.timbre ?? "any" });
+    const finish = (ok: boolean, msg?: string) => {
+      if (cancelled()) {
+        resolve();
+        return;
+      }
+      if (ok) req.onend?.();
+      else req.onerror?.(msg ?? "error");
+      resolve();
+    };
+    const ok = speak(
+      req.text,
+      { voiceURI: pick?.voiceURI ?? null, rate: b.rate, pitch: b.pitch, lang, engine: "browser", serverVoice: null },
+      b.voices,
+      { onstart: () => req.onstart?.(), onend: () => finish(true), onerror: (m) => finish(false, m) },
+    );
+    if (!ok) finish(false, "unsupported");
+  });
+}
 
 export interface Recognizer {
   lang: string;

prompt_v2

@@ -40,3 +40,13 @@ Le modèle doit détecter la langue par le texte/etc. l'user peut customiser l'a
 
 Ces exemples sont les default voix preview . dans l'admin center , l'utilisateur peut les modifier /revenir à défaut
 L'utilisateur doit pouvoir configurée ces ai,vocaux,voix,langue dans une sous-section dédier dans ces paramètres.
+
+--- actually the favicon have an background svg as the same color of the accenthique theme. should be indepandant .
+white = day theme
+black / night theme
+witht the flower itself following thr accentic color
+
+
+-- in specify
+specify the model used in each footer of assistant message
+like <model> <button>

src/config.ts

@@ -80,6 +80,14 @@ export const config = {
   ocrLang: env("OCR_LANG", "eng"),
   ocrMaxChars: envInt("OCR_MAX_CHARS", 100_000),
 
+  // Neural TTS tool (ElevenLabs / OpenAI / local piper, first configured
+  // wins unless TTS_PROVIDER forces one). Without any of these the tool
+  // reports unavailable and clients fall back to browser speech.
+  elevenlabsApiKey: env("ELEVENLABS_API_KEY", ""),
+  elevenlabsVoice: env("ELEVENLABS_VOICE", ""),
+  piperModel: env("PIPER_MODEL", ""),
+  ttsVoice: env("TTS_VOICE", "alloy"),
+
   // Key recovery via email (optional — empty SMTP_HOST disables it).
   smtpHost: env("SMTP_HOST", ""),
   smtpPort: envInt("SMTP_PORT", 587),

src/db.ts

@@ -311,6 +311,11 @@ export async function isOcrToolEnabled(): Promise<boolean> {
   return (await getSetting("tools_ocr_enabled", "1")) === "1";
 }
 
+/** Admin kill-switch for the neural TTS platform tool (default on). */
+export async function isTtsToolEnabled(): Promise<boolean> {
+  return (await getSetting("tools_tts_enabled", "1")) === "1";
+}
+
 /** Admin kill-switch for message reports / flag button (default on). */
 export async function isReportsEnabled(): Promise<boolean> {
   return (await getSetting("reports_enabled", "1")) === "1";

src/index.ts

@@ -143,6 +143,7 @@ import {
   writeCdnFile,
 } from "./cdn.ts";
 import { OCR_MAX_BYTES, ocrAllowed, recordOcrJob } from "./tools/ocr.ts";
+import { TTS_MAX_CHARS, recordTtsJob, ttsAllowed } from "./tools/tts.ts";
 import { tools } from "./tools/registry.ts";
 import { initOllamaNodes } from "./ollamaNodes.ts";
 
@@ -236,13 +237,15 @@ export interface VoicePreview {
   rate: number;
   pitch: number;
   timbre: "masculine" | "feminine" | "any";
+  /** Server neural voice id (ElevenLabs voice_id / OpenAI voice name). Empty = provider default. */
+  voice: string;
 }
 
 const DEFAULT_VOICE_PREVIEWS: VoicePreview[] = [
-  { text: "Salut, euhm, comment je peux t'aider aujourd'hui ?", lang: "fr", rate: 0.85, pitch: 0.75, timbre: "masculine" },
+  { text: "Salut, euhm, comment je peux t'aider aujourd'hui ?", lang: "fr", rate: 0.85, pitch: 0.75, timbre: "masculine", voice: "" },
-  { text: "5 sur 5 je recois ! Que veux-tu ?", lang: "fr", rate: 1, pitch: 0.8, timbre: "masculine" },
+  { text: "5 sur 5 je recois ! Que veux-tu ?", lang: "fr", rate: 1, pitch: 0.8, timbre: "masculine", voice: "" },
-  { text: "Owww, ce chat est trop mignon, oops, pardon. Je me concentre, que puis-je faire pour toi ?", lang: "fr", rate: 1.05, pitch: 1.35, timbre: "feminine" },
+  { text: "Owww, ce chat est trop mignon, oops, pardon. Je me concentre, que puis-je faire pour toi ?", lang: "fr", rate: 1.05, pitch: 1.35, timbre: "feminine", voice: "" },
-  { text: "Hey, que puis-je faire pour toi ?", lang: "fr", rate: 1, pitch: 1.1, timbre: "feminine" },
+  { text: "Hey, que puis-je faire pour toi ?", lang: "fr", rate: 1, pitch: 1.1, timbre: "feminine", voice: "" },
 ];
 
 function sanitizePreviewLine(v: unknown): VoicePreview | null {
@@ -250,7 +253,7 @@ function sanitizePreviewLine(v: unknown): VoicePreview | null {
   if (typeof v === "string") {
     const text = v.trim();
     if (text.length === 0 || text.length > 500) return null;
-    return { text, lang: "auto", rate: 1, pitch: 1, timbre: "any" };
+    return { text, lang: "auto", rate: 1, pitch: 1, timbre: "any", voice: "" };
   }
   if (!v || typeof v !== "object") return null;
   const o = v as Record<string, unknown>;
@@ -261,7 +264,8 @@ function sanitizePreviewLine(v: unknown): VoicePreview | null {
   const rate = typeof o.rate === "number" && o.rate >= 0.5 && o.rate <= 2 ? o.rate : 1;
   const pitch = typeof o.pitch === "number" && o.pitch >= 0.5 && o.pitch <= 2 ? o.pitch : 1;
   const timbre = o.timbre === "masculine" || o.timbre === "feminine" ? o.timbre : "any";
-  return { text, lang, rate, pitch, timbre };
+  const voice = typeof o.voice === "string" && /^[A-Za-z0-9_-]{1,64}$/.test(o.voice.trim()) ? o.voice.trim() : "";
+  return { text, lang, rate, pitch, timbre, voice };
 }
 
 async function getVoicePreviews(): Promise<VoicePreview[]> {
@@ -1031,6 +1035,7 @@ const server = Bun.serve<{ ticketId: string | null; isAdmin: boolean }>({
             recoveryLimitMax: await recoveryLimitMax(),
             recoveryLimitWindowMin: Math.round(await recoveryLimitWindowMs() / 60_000),
             voicePreviews: await getVoicePreviews(),
+            ttsEnabled: (await getSetting("tools_tts_enabled", "1")) === "1",
           },
         });
       }
@@ -1096,10 +1101,14 @@ const server = Bun.serve<{ ticketId: string | null; isAdmin: boolean }>({
             await setSetting("voice_previews", "");
           } else {
             const clean = sanitizeVoicePreviews(body.voicePreviews);
-            if (!clean) return json({ error: "voicePreviews must be 1-6 lines {text, lang?, rate?, pitch?, timbre?}, or [] to reset" }, 400);
+            if (!clean) return json({ error: "voicePreviews must be 1-6 lines {text, lang?, rate?, pitch?, timbre?, voice?}, or [] to reset" }, 400);
             await setSetting("voice_previews", JSON.stringify(clean));
           }
         }
+        if (body.ttsEnabled !== undefined) {
+          if (typeof body.ttsEnabled !== "boolean") return json({ error: "ttsEnabled must be boolean" }, 400);
+          await setSetting("tools_tts_enabled", body.ttsEnabled ? "1" : "0");
+        }
         return json({
           settings: {
             registrationEnabled: await isRegistrationEnabled(),
@@ -1112,6 +1121,7 @@ const server = Bun.serve<{ ticketId: string | null; isAdmin: boolean }>({
             recoveryLimitMax: await recoveryLimitMax(),
             recoveryLimitWindowMin: Math.round(await recoveryLimitWindowMs() / 60_000),
             voicePreviews: await getVoicePreviews(),
+            ttsEnabled: (await getSetting("tools_tts_enabled", "1")) === "1",
           },
         });
       }
@@ -1815,6 +1825,39 @@ const server = Bun.serve<{ ticketId: string | null; isAdmin: boolean }>({
         }
       }
 
+      // --- platform tools: neural TTS (natural voices, cached) ---
+      if (path === "/api/tools/tts/info" && req.method === "GET") {
+        const tts = tools.tts;
+        return json({ ...(await tts.providerInfo()), available: await tts.isAvailable(), reason: await tts.unavailableReason() });
+      }
+
+      if (path === "/api/tools/tts" && req.method === "POST") {
+        const tts = tools.tts;
+        if (!await tts.isAvailable()) return json({ error: (await tts.unavailableReason()) ?? "tts unavailable" }, 501);
+        const { ok, body } = await readJson(req);
+        if (!ok || typeof body.text !== "string") return json({ error: "expected { text, voice? }" }, 400);
+        const text = body.text.trim();
+        if (text.length === 0 || text.length > TTS_MAX_CHARS) {
+          return json({ error: `text: 1-${TTS_MAX_CHARS} chars` }, 400);
+        }
+        const rl = ttsAllowed(user.id);
+        if (!rl.ok) {
+          return json({ error: "tts rate limited: 30 jobs per hour", retryAfterSec: rl.retryAfterSec }, 429);
+        }
+        recordTtsJob(user.id);
+        try {
+          const { audio, contentType } = await tts.synthesize(text, typeof body.voice === "string" ? body.voice : undefined);
+          const bytes = audio.buffer.slice(audio.byteOffset, audio.byteOffset + audio.byteLength) as ArrayBuffer;
+          return new Response(bytes, {
+            headers: { "Content-Type": contentType, "Cache-Control": "private, max-age=86400" },
+          });
+        } catch (e) {
+          const msg = (e as Error).message;
+          if (/not available|disabled|failed to start/i.test(msg)) return json({ error: msg }, 501);
+          return json({ error: msg }, 502);
+        }
+      }
+
       if (path === "/api/chat" && req.method === "POST") {
         const { ok, body } = await readJson(req);
         if (!ok) return json({ error: "invalid JSON" }, 400);

src/tools/registry.ts

@@ -1,13 +1,16 @@
 import { OcrTool } from "./ocr.ts";
+import { TtsTool } from "./tts.ts";
 import type { PlatformTool } from "./types.ts";
 
 /** Platform tools registry (upload flows always go through tools). */
 class ToolRegistry {
   readonly ocr = new OcrTool();
-  private tools: PlatformTool[] = [this.ocr];
+  readonly tts = new TtsTool();
+  private tools: PlatformTool[] = [this.ocr, this.tts];
 
   async init(): Promise<void> {
     await this.ocr.init();
+    await this.tts.init();
   }
 
   async list(): Promise<{ name: string; description: string; available: boolean; reason: string | null }[]> {

src/tools/tts.ts

@@ -0,0 +1,279 @@
+import { createHash } from "node:crypto";
+import { mkdirSync, readdirSync, rmSync, statSync } from "node:fs";
+import { join } from "node:path";
+import { config } from "../config.ts";
+import { isTtsToolEnabled } from "../db.ts";
+import { driverLog, fmtBytes, fmtMs } from "../drivers/log.ts";
+import type { PlatformTool } from "./types.ts";
+
+/**
+ * Neural TTS tool: turns assistant replies / preview lines into natural
+ * speech (far beyond the browser's robotic speechSynthesis).
+ *
+ * Providers, tried in order (first configured wins; `TTS_PROVIDER` forces
+ * one explicitly):
+ * - elevenlabs (ELEVENLABS_API_KEY [+ ELEVENLABS_VOICE]) — most natural,
+ *   any voice_id (20-char ids, clones included).
+ * - openai (OPENAI_API_KEY, reuses the LLM provider key) — tts-1 with the
+ *   6 stock voices (alloy/echo/fable/onyx/nova/shimmer).
+ * - piper (PIPER_BIN + PIPER_MODEL, local binary like tesseract) — offline,
+ *   single configured voice, WAV output.
+ *
+ * Output is cached on disk (DATA_DIR/tts-cache, sha of provider|voice|text,
+ * 200 files cap) so repeated lines cost nothing. Without any provider the
+ * tool reports unavailable and the frontend falls back to browser speech.
+ */
+
+export type TtsProvider = "elevenlabs" | "openai" | "piper";
+
+export const OPENAI_VOICES = ["alloy", "echo", "fable", "onyx", "nova", "shimmer"] as const;
+
+const TTS_MAX_CHARS = 2000;
+const TTS_TIMEOUT_MS = 60_000;
+const TTS_CACHE_CAP = 200;
+
+// Abuse protection: neural TTS costs money per byte.
+const TTS_WINDOW_MS = 3600_000;
+const TTS_MAX_JOBS = 30;
+const ttsJobs = new Map<string, number[]>();
+
+export function ttsAllowed(userId: string): { ok: boolean; retryAfterSec: number } {
+  const now = Date.now();
+  const list = (ttsJobs.get(userId) ?? []).filter((t) => t > now - TTS_WINDOW_MS);
+  ttsJobs.set(userId, list);
+  if (list.length >= TTS_MAX_JOBS) {
+    return { ok: false, retryAfterSec: Math.max(1, Math.ceil((list[0]! + TTS_WINDOW_MS - now) / 1000)) };
+  }
+  return { ok: true, retryAfterSec: 0 };
+}
+
+export function recordTtsJob(userId: string): void {
+  const list = ttsJobs.get(userId) ?? [];
+  list.push(Date.now());
+  ttsJobs.set(userId, list);
+}
+
+function elevenlabsBase(): string {
+  return (process.env["ELEVENLABS_BASE_URL"] || "https://api.elevenlabs.io").replace(/\/$/, "");
+}
+
+function openaiBase(): string {
+  return (process.env["OPENAI_BASE_URL"] || "https://api.openai.com").replace(/\/$/, "");
+}
+
+function piperBin(): string {
+  return process.env["PIPER_BIN"] || "piper";
+}
+
+async function piperAvailable(): Promise<boolean> {
+  if (!config.piperModel) return false;
+  try {
+    const proc = Bun.spawn([piperBin(), "--version"], { stdout: "pipe", stderr: "pipe" });
+    const code = await proc.exited;
+    return code === 0;
+  } catch {
+    return false;
+  }
+}
+
+/** Which provider would serve right now (null = none configured). */
+export async function activeTtsProvider(): Promise<TtsProvider | null> {
+  const forced = (process.env["TTS_PROVIDER"] || "auto").toLowerCase();
+  const eleven = config.elevenlabsApiKey.length > 0;
+  const openai = config.openaiApiKey.length > 0;
+  const piper = await piperAvailable();
+  const pick = (p: TtsProvider): TtsProvider | null =>
+    p === "elevenlabs" ? (eleven ? p : null) : p === "openai" ? (openai ? p : null) : piper ? p : null;
+  if (forced === "elevenlabs" || forced === "openai" || forced === "piper") return pick(forced);
+  return pick("elevenlabs") ?? pick("openai") ?? pick("piper");
+}
+
+export function sanitizeTtsVoice(provider: TtsProvider, raw: unknown): string | null {
+  if (raw === undefined || raw === null || raw === "") return null;
+  if (typeof raw !== "string") return null;
+  const v = raw.trim();
+  if (provider === "openai") {
+    return (OPENAI_VOICES as readonly string[]).includes(v.toLowerCase()) ? v.toLowerCase() : null;
+  }
+  if (provider === "elevenlabs") {
+    return /^[A-Za-z0-9_-]{1,64}$/.test(v) ? v : null;
+  }
+  return null; // piper: single configured model voice
+}
+
+function cacheDir(): string {
+  const dir = join(config.dataDir, "tts-cache");
+  mkdirSync(dir, { recursive: true });
+  return dir;
+}
+
+function cacheKey(provider: TtsProvider, voice: string, text: string): string {
+  return createHash("sha256").update(`${provider}|${voice}|${text}`).digest("hex");
+}
+
+function cachedPath(key: string, ext: string): string {
+  return join(cacheDir(), `${key}.${ext}`);
+}
+
+function pruneCache(): void {
+  try {
+    const dir = cacheDir();
+    const files = readdirSync(dir)
+      .map((f) => {
+        try {
+          return { f, mtime: statSync(join(dir, f)).mtimeMs };
+        } catch {
+          return null;
+        }
+      })
+      .filter((x): x is { f: string; mtime: number } => x !== null)
+      .sort((a, b) => a.mtime - b.mtime);
+    for (const old of files.slice(0, Math.max(0, files.length - TTS_CACHE_CAP))) {
+      try {
+        rmSync(join(dir, old.f));
+      } catch {
+        // ignore
+      }
+    }
+  } catch {
+    // cache must never break synthesis
+  }
+}
+
+async function synthElevenlabs(text: string, voiceId: string): Promise<{ audio: Uint8Array; contentType: string }> {
+  const res = await fetch(`${elevenlabsBase()}/v1/text-to-speech/${encodeURIComponent(voiceId)}`, {
+    method: "POST",
+    headers: {
+      "Content-Type": "application/json",
+      Accept: "audio/mpeg",
+      "xi-api-key": config.elevenlabsApiKey,
+    },
+    body: JSON.stringify({ text, model_id: "eleven_turbo_v2_5" }),
+    signal: AbortSignal.timeout(TTS_TIMEOUT_MS),
+  });
+  if (!res.ok) {
+    const err = await res.text().catch(() => "");
+    throw new Error(`elevenlabs tts failed (${res.status}): ${err.slice(0, 200)}`);
+  }
+  return { audio: new Uint8Array(await res.arrayBuffer()), contentType: "audio/mpeg" };
+}
+
+async function synthOpenai(text: string, voice: string): Promise<{ audio: Uint8Array; contentType: string }> {
+  const res = await fetch(`${openaiBase()}/v1/audio/speech`, {
+    method: "POST",
+    headers: { "Content-Type": "application/json", Authorization: `Bearer ${config.openaiApiKey}` },
+    body: JSON.stringify({ model: "tts-1", input: text, voice, response_format: "mp3" }),
+    signal: AbortSignal.timeout(TTS_TIMEOUT_MS),
+  });
+  if (!res.ok) {
+    const err = await res.text().catch(() => "");
+    throw new Error(`openai tts failed (${res.status}): ${err.slice(0, 200)}`);
+  }
+  return { audio: new Uint8Array(await res.arrayBuffer()), contentType: "audio/mpeg" };
+}
+
+async function synthPiper(text: string): Promise<{ audio: Uint8Array; contentType: string }> {
+  const out = join(cacheDir(), `piper-${Date.now().toString(36)}-${Math.floor(Math.random() * 1e6).toString(36)}.wav`);
+  let proc;
+  try {
+    proc = Bun.spawn([piperBin(), "--model", config.piperModel, "--output_file", out], {
+      stdin: new TextEncoder().encode(text),
+      stdout: "pipe",
+      stderr: "pipe",
+    });
+  } catch (e) {
+    throw new Error(`tts engine failed to start: ${(e as Error).message}`);
+  }
+  const timeout = setTimeout(() => {
+    try {
+      proc.kill();
+    } catch {
+      // already exited
+    }
+  }, TTS_TIMEOUT_MS);
+  try {
+    const [code, errText] = await Promise.all([proc.exited, new Response(proc.stderr).text()]);
+    if (code !== 0) throw new Error(`piper failed (exit ${code}): ${errText.trim().slice(0, 200)}`);
+    const file = Bun.file(out);
+    if (!(await file.exists())) throw new Error("piper produced no output");
+    return { audio: new Uint8Array(await file.arrayBuffer()), contentType: "audio/wav" };
+  } finally {
+    clearTimeout(timeout);
+    try {
+      rmSync(out);
+    } catch {
+      // ignore
+    }
+  }
+}
+
+export class TtsTool implements PlatformTool {
+  readonly name = "tts";
+  readonly description = "Neural text-to-speech (ElevenLabs / OpenAI / local piper) for natural voice playback.";
+
+  async init(): Promise<void> {
+    const p = await activeTtsProvider();
+    driverLog("tools", `tts engine: ${p ?? "none configured — browser speech fallback only"}`);
+  }
+
+  async isAvailable(): Promise<boolean> {
+    return (await activeTtsProvider()) !== null && (await isTtsToolEnabled());
+  }
+
+  async unavailableReason(): Promise<string | null> {
+    if ((await activeTtsProvider()) === null) {
+      return "no TTS provider configured (ELEVENLABS_API_KEY, OPENAI_API_KEY or PIPER_BIN+PIPER_MODEL). Browser speech is used instead.";
+    }
+    if (!(await isTtsToolEnabled())) return "TTS is disabled by the administrator (Admin Center → Features).";
+    return null;
+  }
+
+  async providerInfo(): Promise<{ provider: TtsProvider | null; voices: string[] }> {
+    const p = await activeTtsProvider();
+    if (!p) return { provider: null, voices: [] };
+    if (p === "openai") return { provider: p, voices: [...OPENAI_VOICES] };
+    if (p === "elevenlabs") return { provider: p, voices: config.elevenlabsVoice ? [config.elevenlabsVoice] : [] };
+    return { provider: p, voices: [] };
+  }
+
+  defaultVoice(provider: TtsProvider): string {
+    if (provider === "openai") {
+      const v = sanitizeTtsVoice("openai", config.ttsVoice);
+      return v ?? "alloy";
+    }
+    if (provider === "elevenlabs") return config.elevenlabsVoice || "pNInz6obpgDQGcFmaJgB";
+    return "";
+  }
+
+  async synthesize(text: string, voice?: string): Promise<{ audio: Uint8Array; contentType: string; provider: TtsProvider; fromCache: boolean }> {
+    const provider = await activeTtsProvider();
+    if (!provider) throw new Error((await this.unavailableReason()) ?? "tts unavailable");
+    if (!(await isTtsToolEnabled())) throw new Error("tts disabled");
+    const clean = text.trim().slice(0, TTS_MAX_CHARS);
+    if (!clean) throw new Error("empty text");
+    const resolved = sanitizeTtsVoice(provider, voice) ?? this.defaultVoice(provider);
+    const ext = provider === "piper" ? "wav" : "mp3";
+    const key = cacheKey(provider, resolved, clean);
+    const hit = Bun.file(cachedPath(key, ext));
+    if (await hit.exists()) {
+      return {
+        audio: new Uint8Array(await hit.arrayBuffer()),
+        contentType: provider === "piper" ? "audio/wav" : "audio/mpeg",
+        provider,
+        fromCache: true,
+      };
+    }
+    const t0 = Date.now();
+    const out = provider === "elevenlabs"
+      ? await synthElevenlabs(clean, resolved)
+      : provider === "openai"
+        ? await synthOpenai(clean, resolved)
+        : await synthPiper(clean);
+    await Bun.write(cachedPath(key, ext), out.audio);
+    pruneCache();
+    driverLog("tools", `tts ${provider} voice=${resolved || "default"}: ${fmtBytes(out.audio.length)} in ${fmtMs(Date.now() - t0)}`);
+    return { ...out, provider, fromCache: false };
+  }
+}
+
+export { TTS_MAX_CHARS };