Switch podcast TTS to Gemini Kore

start-hf.sh

@@ -125,6 +125,153 @@ if "project: Optional[str] = None" not in src or "google.auth.default(scopes=sco
     if count != 1:
         raise RuntimeError("Could not patch Esperanto Vertex TTS _get_access_token; source layout changed")
 
+if "_is_gemini_tts_model" not in src:
+    new_generate = r'''    def _is_gemini_tts_model(self) -> bool:
+        """Return True for Gemini text-to-speech models."""
+        model = self.model_name or ""
+        return model.startswith("gemini-") and "tts" in model
+
+    def _get_gemini_tts_language(self, kwargs: Dict) -> str:
+        """Resolve Gemini-TTS language code."""
+        config = getattr(self, "_config", {}) or {}
+        return (
+            kwargs.get("language_code")
+            or config.get("language_code")
+            or os.getenv("VERTEX_TTS_LANGUAGE")
+            or os.getenv("GOOGLE_CLOUD_REGION_LANGUAGE")
+            or "be-BY"
+        )
+
+    def _get_gemini_tts_prompt(self, kwargs: Dict) -> str:
+        """Resolve Gemini-TTS style prompt."""
+        config = getattr(self, "_config", {}) or {}
+        return (
+            kwargs.get("prompt")
+            or config.get("prompt")
+            or os.getenv("VERTEX_TTS_PROMPT")
+            or "Read the following text in Belarusian (Belarus), with a warm, calm, clear, intelligent, natural podcast delivery."
+        )
+
+    def _build_tts_payload(self, text: str, voice: str, kwargs: Dict) -> Dict:
+        """Build Cloud Text-to-Speech payload for legacy and Gemini-TTS models."""
+        if self._is_gemini_tts_model():
+            return {
+                "input": {
+                    "prompt": self._get_gemini_tts_prompt(kwargs),
+                    "text": text,
+                },
+                "voice": {
+                    "languageCode": self._get_gemini_tts_language(kwargs),
+                    "name": voice,
+                    "modelName": self.model_name,
+                },
+                "audioConfig": {
+                    "audioEncoding": kwargs.get("audio_encoding", "MP3"),
+                },
+            }
+
+        language_code = voice.split("-")[0] + "-" + voice.split("-")[1]
+        return {
+            "input": {
+                "text": text
+            },
+            "voice": {
+                "languageCode": language_code,
+                "name": voice
+            },
+            "audioConfig": {
+                "audioEncoding": "MP3"
+            }
+        }
+
+    def _audio_content_type(self, kwargs: Dict) -> str:
+        encoding = kwargs.get("audio_encoding", "MP3").upper()
+        if encoding == "LINEAR16":
+            return "audio/wav"
+        if encoding == "OGG_OPUS":
+            return "audio/ogg"
+        return "audio/mp3"
+
+    def generate_speech(
+        self,
+        text: str,
+        voice: str,
+        output_file: Optional[Union[str, Path]] = None,
+        **kwargs
+    ) -> AudioResponse:
+        """Generate speech from text."""
+        payload = self._build_tts_payload(text, voice, kwargs)
+
+        response = self.client.post(
+            f"{self.base_url}/text:synthesize",
+            headers=self._get_headers(),
+            json=payload
+        )
+        self._handle_error(response)
+
+        response_data = response.json()
+        audio_data = base64.b64decode(response_data["audioContent"])
+
+        response_audio = AudioResponse(
+            audio_data=audio_data,
+            content_type=self._audio_content_type(kwargs),
+            model=self.model_name or self._get_default_model(),
+            voice=voice,
+            provider="vertex"
+        )
+
+        if output_file:
+            output_path = Path(output_file)
+            output_path.parent.mkdir(parents=True, exist_ok=True)
+            output_path.write_bytes(audio_data)
+
+        return response_audio
+
+    async def agenerate_speech(
+        self,
+        text: str,
+        voice: str,
+        output_file: Optional[Union[str, Path]] = None,
+        **kwargs
+    ) -> AudioResponse:
+        """Generate speech from text asynchronously."""
+        payload = self._build_tts_payload(text, voice, kwargs)
+
+        response = await self.async_client.post(
+            f"{self.base_url}/text:synthesize",
+            headers=self._get_headers(),
+            json=payload
+        )
+        self._handle_error(response)
+
+        response_data = response.json()
+        audio_data = base64.b64decode(response_data["audioContent"])
+
+        response_audio = AudioResponse(
+            audio_data=audio_data,
+            content_type=self._audio_content_type(kwargs),
+            model=self.model_name or self._get_default_model(),
+            voice=voice,
+            provider="vertex"
+        )
+
+        if output_file:
+            output_path = Path(output_file)
+            output_path.parent.mkdir(parents=True, exist_ok=True)
+            output_path.write_bytes(audio_data)
+
+        return response_audio
+'''
+
+    src, count = re.subn(
+        r"    def generate_speech\([\s\S]*$",
+        new_generate,
+        src,
+        count=1,
+    )
+    if count != 1:
+        raise RuntimeError("Could not patch Esperanto Vertex TTS generate_speech; source layout changed")
+
     target.write_text(src, encoding="utf-8")
 
 print(f"Patched Vertex TTS provider: {target}")
@@ -212,7 +359,7 @@ async def upsert_speaker_profile(tts_model_id: str) -> str:
     speakers = [
         {
             "name": "Вядоўца",
-            "voice_id": "en-US-Standard-A",
+            "voice_id": "Kore",
             "backstory": "Вядоўца беларускамоўнага падкаста.",
             "personality": "Warm, calm, clear, intelligent, natural delivery.",
             "voice_model": tts_model_id,
@@ -231,7 +378,7 @@ async def upsert_speaker_profile(tts_model_id: str) -> str:
     profile.description = ""
     profile.voice_model = tts_model_id
     profile.tts_provider = "vertex"
-    profile.tts_model = "standard"
+    profile.tts_model = "gemini-2.5-flash-tts"
     profile.speakers = speakers
     await profile.save()
     return str(profile.id)
@@ -288,7 +435,7 @@ async def seed() -> None:
     flash_id = await upsert_model("gemini-2.5-flash", "language", credential_id)
     pro_id = await upsert_model("gemini-2.5-pro", "language", credential_id)
     embedding_id = await upsert_model("text-embedding-005", "embedding", credential_id)
-    tts_id = await upsert_model("standard", "text_to_speech", credential_id)
+    tts_id = await upsert_model("gemini-2.5-flash-tts", "text_to_speech", credential_id)
     await set_default_models(flash_id, pro_id, embedding_id, tts_id)
     await upsert_speaker_profile(tts_id)
     await upsert_episode_profile(flash_id, flash_id)