Bundle WhatsApp female speaker embedding

WHATSAPP_HF.md

@@ -30,5 +30,6 @@ If you need 24/7 availability, use an always-on Hugging Face deployment or equiv
 - The planner model is optional and is not auto-downloaded on startup; set `AARA_ENABLE_PLANNER_DOWNLOAD=1` if you want it.
 - WhatsApp replies now include a generated audio attachment, so the reply plays as a voice message instead of plain text.
 - WhatsApp voice replies use an Indian female SpeechT5 checkpoint from Hugging Face for a clearer local voice.
+- The female speaker embedding is bundled in `reference_audio/female_speaker_embedding.npy`, so the Space does not need to fetch the xvector archive at runtime.
 - If you want inbound WhatsApp voice notes to be transcribed, set `TWILIO_ACCOUNT_SID` and `TWILIO_AUTH_TOKEN` so the Space can download Twilio media.
 - If you want the bot to greet first, send a first message like `hello` in WhatsApp after joining the sandbox.

reference_audio/female_speaker_embedding.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5718a29583dfdc17d760f5fb807fdd8ce68c1cf48d6acbec208337474bed178a
+size 2176

voice_agent_standalone.py

@@ -1922,6 +1922,7 @@ class TTSEngine:
     SPEECH_T5_TTS_MODEL = "aryamannningombam/speecht5_finetuned_indian_female_v3"
     SPEECH_T5_PROCESSOR_MODEL = "microsoft/speecht5_tts"
     SPEECH_T5_VOCODER_MODEL = "microsoft/speecht5_hifigan"
+    FEMALE_SPEAKER_EMBEDDING_FILE = "female_speaker_embedding.npy"
     SPEAKER_ARCHIVE_REPO = "Matthijs/cmu-arctic-xvectors"
     FEMALE_SPEAKER_FILE = "spkrec-xvect/cmu_us_slt_arctic-wav-arctic_a0402.npy"
     MALE_SPEAKER_FILE = "spkrec-xvect/cmu_us_bdl_arctic-wav-arctic_a0153.npy"
@@ -2051,6 +2052,16 @@ class TTSEngine:
                 self._speaker_embedding_cache[cache_key] = tensor
                 return tensor
 
+        local_embedding_path = self.cfg.reference_audio_dir / self.FEMALE_SPEAKER_EMBEDDING_FILE
+        if preference != "male" and local_embedding_path.exists():
+            cache_key = str(local_embedding_path.resolve())
+            if cache_key in self._speaker_embedding_cache:
+                return self._speaker_embedding_cache[cache_key]
+            embedding = np.load(local_embedding_path, allow_pickle=False)
+            tensor = torch.from_numpy(np.asarray(embedding, dtype=np.float32).reshape(1, -1))
+            self._speaker_embedding_cache[cache_key] = tensor
+            return tensor
+
         speaker_file = self.MALE_SPEAKER_FILE if preference == "male" else self.FEMALE_SPEAKER_FILE
         if speaker_file in self._speaker_embedding_cache:
             return self._speaker_embedding_cache[speaker_file]