Motore voce Voicebox (Qwen3-TTS via MLX, profili clonati, istruzioni di stile)

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

avatar/assistant.py

@@ -234,7 +234,7 @@ class Assistant:
                 self.settings.save()
         else:
             eng = self.settings.get("tts_engine")
-            label = "Sistema" if eng == "system" else ("Sara" if eng in ("chatterbox", "elevenlabs") else
+            label = "Sistema" if eng == "system" else ("Sara" if eng in ("chatterbox", "elevenlabs", "voicebox") else
                 {"if_sara": "Sara", "im_nicola": "Nicola"}.get(self.settings.get("kokoro_voice"), "Sara"))
             try:
                 if cm.get_voice() != label:

avatar/settings.py

@@ -26,6 +26,9 @@ DEFAULTS: dict[str, Any] = {
     "kokoro_voice": "if_sara",
     "chatterbox_exaggeration": 0.6,    # 0.3 sobria … 0.9 molto enfatica
     "chatterbox_cfg": 0.3,             # più basso = più veloce e meno aderente al testo
+    "voicebox_profile_id": "",
+    "voicebox_engine": "qwen",         # qwen | chatterbox | chatterbox_turbo | kokoro | luxtts
+    "voicebox_instruct": "",           # istruzione di stile per Qwen (es. "parla in modo caloroso e calmo")
     "elevenlabs_voice_id": "",
     "elevenlabs_model": "eleven_flash_v2_5",   # rapido; eleven_multilingual_v2 = qualità massima
     "elevenlabs_stability": 0.45,

avatar/settings_dialog.py

@@ -99,7 +99,7 @@ class SettingsDialog(QDialog):
         self.stack.setCurrentIndex(self.provider.currentIndex())
 
         form2 = QFormLayout()
-        self.tts_engine = _combo([("kokoro", "Kokoro, voce neurale in locale"), ("elevenlabs", "ElevenLabs, espressiva nel cloud (chiave API)"), ("chatterbox", "Chatterbox, espressiva in locale (lenta)"), ("system", "Voce di sistema (macOS)")], s.get("tts_engine"))
+        self.tts_engine = _combo([("kokoro", "Kokoro, voce neurale in locale"), ("voicebox", "Voicebox: Qwen3-TTS in locale, espressiva, voce clonata"), ("elevenlabs", "ElevenLabs, espressiva nel cloud (chiave API)"), ("chatterbox", "Chatterbox, espressiva in locale (lenta)"), ("system", "Voce di sistema (macOS)")], s.get("tts_engine"))
         form2.addRow("Motore voce", self.tts_engine)
         self.kokoro_voice = _combo(list(KOKORO_VOICES.items()), s.get("kokoro_voice")); form2.addRow("Voce Kokoro", self.kokoro_voice)
         self.system_voice = _combo([("", "Automatica (Alice)")] + [(v, v) for v in SystemVoice.list_voices()], s.get("system_voice"))
@@ -116,6 +116,16 @@ class SettingsDialog(QDialog):
         self.cb_ref = QLineEdit(cur_ref if cur_ref.startswith("/") or cur_ref.startswith("~") else ""); self.cb_ref.setPlaceholderText("percorso di un wav di 10 secondi con la voce da imitare"); row.addWidget(self.cb_ref, 1)
         form2.addRow("Voce Chatterbox", row)
         row = QHBoxLayout()
+        self.vb_profile = QComboBox(); self.vb_profile.setMinimumWidth(220)
+        self._vb_current = str(s.get("voicebox_profile_id") or "")
+        self.vb_profile.addItem("(carica i profili da Voicebox)" if not self._vb_current else f"profilo salvato: {self._vb_current[:8]}…", self._vb_current)
+        row.addWidget(self.vb_profile, 1)
+        b = QPushButton("Carica profili"); b.clicked.connect(self._vb_profiles); row.addWidget(b)
+        self.vb_engine = _combo([("qwen", "Qwen3-TTS (consigliato)"), ("kokoro", "Kokoro"), ("chatterbox_turbo", "Chatterbox Turbo"), ("chatterbox", "Chatterbox"), ("luxtts", "LuxTTS")], s.get("voicebox_engine")); row.addWidget(self.vb_engine)
+        form2.addRow("Voicebox", row)
+        self.vb_instruct = QLineEdit(s.get("voicebox_instruct") or ""); self.vb_instruct.setPlaceholderText("istruzione di stile per Qwen, es. «parla in modo caloroso e calmo» (facoltativa)")
+        form2.addRow("", self.vb_instruct)
+        row = QHBoxLayout()
         self.el_key = QLineEdit(); self.el_key.setEchoMode(QLineEdit.EchoMode.Password)
         self.el_key.setPlaceholderText("•••••• (salvata)" if s.get_secret("elevenlabs_api_key") else "chiave API da elevenlabs.io"); row.addWidget(self.el_key, 1)
         self.el_voice = QComboBox(); self.el_voice.setMinimumWidth(220); row.addWidget(self.el_voice, 1)
@@ -313,6 +323,15 @@ class SettingsDialog(QDialog):
                 self._async.emit("wa", (f"Errore: {err}", None))
         threading.Thread(target=work, daemon=True).start()
 
+    def _vb_profiles(self) -> None:
+        def work():
+            try:
+                from .tts import VoiceboxVoice
+                self._async.emit("vb", VoiceboxVoice.list_profiles())
+            except Exception as err:
+                self._async.emit("vb_error", str(err)[:120])
+        threading.Thread(target=work, daemon=True).start()
+
     def _el_voices(self) -> None:
         key = self.el_key.text().strip() or self.settings.get_secret("elevenlabs_api_key")
         if not key:
@@ -328,6 +347,15 @@ class SettingsDialog(QDialog):
         threading.Thread(target=work, daemon=True).start()
 
     def _on_async(self, kind: str, payload) -> None:
+        if kind == "vb":
+            self.vb_profile.clear()
+            for pid, label in payload:
+                self.vb_profile.addItem(label, pid)
+            idx = self.vb_profile.findData(self._vb_current)
+            self.vb_profile.setCurrentIndex(idx if idx >= 0 else 0)
+            return
+        if kind == "vb_error":
+            self.vb_profile.clear(); self.vb_profile.addItem(f"errore: {payload}", ""); return
         if kind == "el":
             self.el_voice.clear()
             for vid, label in payload:
@@ -389,6 +417,7 @@ class SettingsDialog(QDialog):
             "claudecode_config_dir": self.cc_config.currentText().strip(), "claudecode_path": self.cc_path.text().strip(),
             "tts_engine": self.tts_engine.currentData(), "kokoro_voice": self.kokoro_voice.currentData(),
             "system_voice": self.system_voice.currentData(), "stt_model": self.stt_model.currentData(),
+            "voicebox_profile_id": self.vb_profile.currentData() or "", "voicebox_engine": self.vb_engine.currentData(), "voicebox_instruct": self.vb_instruct.text().strip(),
             "elevenlabs_voice_id": self.el_voice.currentData() or "", "elevenlabs_model": self.el_model.currentData(),
             "elevenlabs_stability": float(self.el_stab.value()), "elevenlabs_style": float(self.el_style.value()),
             "chatterbox_exaggeration": float(self.cb_exag.value()), "chatterbox_cfg": float(self.cb_cfg.value()),

avatar/tts.py

@@ -266,7 +266,91 @@ class ElevenLabsVoice:
         return np.frombuffer(r.content, dtype=np.int16).astype(np.float32) / 32768.0
 
 
+class VoiceboxVoice:
+    """Voicebox (app locale): Qwen3-TTS, Chatterbox, Kokoro… con accelerazione MLX e profili vocali clonati."""
+
+    API = "http://127.0.0.1:17493"
+
+    def __init__(self, profile_id: str, engine: str = "qwen", instruct: str = "", on_status=None) -> None:
+        self.profile_id, self.engine, self.instruct = profile_id, engine, instruct
+        self.voice = f"voicebox:{profile_id}:{engine}:{instruct}"
+        self._on_status = on_status or (lambda m: None)
+
+    @classmethod
+    def alive(cls) -> bool:
+        import requests
+        try:
+            return requests.get(f"{cls.API}/health", timeout=3).ok
+        except requests.RequestException:
+            return False
+
+    @classmethod
+    def ensure_running(cls, wait: int = 60) -> None:
+        import subprocess
+        import time as _t
+        if cls.alive():
+            return
+        subprocess.run(["open", "-a", "Voicebox"], check=False, timeout=15)
+        for _ in range(wait):
+            _t.sleep(1)
+            if cls.alive():
+                return
+        raise RuntimeError("Voicebox non risponde: installalo da github.com/jamiepine/voicebox e aprilo una volta.")
+
+    @classmethod
+    def list_profiles(cls) -> list[tuple[str, str]]:
+        import requests
+        cls.ensure_running()
+        d = requests.get(f"{cls.API}/profiles", timeout=15).json()
+        d = d if isinstance(d, list) else d.get("profiles", [])
+        return [(p["id"], f"{p.get('name')} ({p.get('language') or '?'}, {p.get('voice_type') or ''})") for p in d]
+
+    def load(self) -> None:
+        if not self.profile_id:
+            raise RuntimeError("Voicebox: scegli un profilo vocale in Motore e Voce (pulsante «Carica profili»).")
+        self._on_status("Avvio Voicebox…")
+        try:
+            self.ensure_running()
+        finally:
+            self._on_status(None)
+
+    def synthesize(self, text: str) -> np.ndarray:
+        import io
+        import time as _t
+        import requests
+        import soundfile as sf
+        body = {"profile_id": self.profile_id, "text": text, "language": "it", "engine": self.engine}
+        if self.engine == "qwen":
+            body["model_size"] = "1.7B"
+        if self.instruct and self.engine in ("qwen", "qwen_custom_voice"):
+            body["instruct"] = self.instruct[:500]
+        r = requests.post(f"{self.API}/generate", json=body, timeout=120)
+        if not r.ok:
+            raise RuntimeError(f"Voicebox {r.status_code}: {r.text[:160]}")
+        g = r.json()
+        gid = g["id"]
+        t0 = _t.time()
+        # /generate/{id}/status è un flusso SSE: per il polling si usa /history/{id} (JSON).
+        while g.get("status") in ("generating", "pending", "queued", "processing") and _t.time() - t0 < 300:
+            _t.sleep(0.25)
+            h = requests.get(f"{self.API}/history/{gid}", timeout=10)
+            if h.ok and h.headers.get("Content-Type", "").startswith("application/json"):
+                g = h.json()
+        if g.get("status") != "completed":
+            raise RuntimeError(f"Voicebox: {g.get('status')} {g.get('error') or ''}".strip())
+        a = requests.get(f"{self.API}/audio/{gid}", timeout=60)
+        data, sr = sf.read(io.BytesIO(a.content), dtype="float32")
+        if data.ndim > 1:
+            data = data[:, 0]
+        if sr != OUT_RATE:
+            data = np.interp(np.arange(0, data.size, sr / OUT_RATE), np.arange(data.size), data).astype(np.float32)
+        return data
+
+
 def make_voice(settings, on_status=None):
+    if settings.get("tts_engine") == "voicebox":
+        return VoiceboxVoice(str(settings.get("voicebox_profile_id") or ""), str(settings.get("voicebox_engine") or "qwen"),
+                             str(settings.get("voicebox_instruct") or ""), on_status=on_status)
     if settings.get("tts_engine") == "elevenlabs":
         return ElevenLabsVoice(settings.get_secret("elevenlabs_api_key"), str(settings.get("elevenlabs_voice_id") or ""),
                                str(settings.get("elevenlabs_model") or "eleven_flash_v2_5"),