flash attention ensure

.gitignore

@@ -84,7 +84,6 @@ Thumbs.db
 *.tmp
 *.temp
 .cache/
-publish.sh
 
 # =============================================================================
 # Old/Archive Folders
@@ -97,3 +96,4 @@ dev/
 # =============================================================================
 !data/config.example.yml
 !.env.example
+publish.sh

core/asr.py

@@ -76,6 +76,7 @@ def load_asr_model():
         ASR_MODEL_NAME,
         device_map=DEVICE,
         dtype=DTYPE,
+        attn_implementation="flash_attention_2",
     )
 
     return QwenASRPipelineWrapper(model)

core/tts.py

@@ -50,6 +50,7 @@ model = Qwen3TTSModel.from_pretrained(
     "Qwen/Qwen3-TTS-12Hz-1.7B-Base",
     torch_dtype=torch.bfloat16,
     device_map=DEVICE,
+    attn_implementation="flash_attention_2",
 )
 
 # Enable optimizations (recommended)
@@ -79,6 +80,7 @@ def speak_stream(text: str, voice: str = "cori", speed: float = 1.0):
         text=text,
         language="english",
         voice_clone_prompt=prompt,
+        overlap_samples=512,
         # Phase 2 settings (stable)
         emit_every_frames=12,
         decode_window_frames=80,

data/config.example.yml

@@ -11,7 +11,7 @@ general:
   wakeword: "computer" 
   # Use a small language model for extra capabilities
   use_ai: True
-  # Use the Moonshine Tiny ASR if running on very small edge devices
+  # Use the Moonshine Tiny ASR for faster speech transcription (English only)
   use_tiny_asr: False
   # Use the Kokoro TTS for faster voice streaming (voice cloning not available)
   use_tiny_tts: False