feat: add Deepgram STT provider and cloud-first config

- New DeepgramSTT class using Deepgram nova-3 via REST API
- Factory function create_stt_engine() for provider switching
- faster-whisper import now optional (graceful fallback)
- Config defaults to cloud providers (deepgram STT + venice TTS)
- .env.example updated with DEEPGRAM_API_KEY and VENICE_API_KEY
- requirements.txt adds deepgram-sdk, marks faster-whisper as optional
- Zero GPU required for default configuration

config.yaml

@@ -108,20 +108,19 @@ pipeline:
     # Using v3.2 GPU model for best performance with RTX 5090
     model_path: "smart-turn-v3.2-gpu.onnx"
 
-  # Speech-to-Text (faster-whisper)
+  # Speech-to-Text
   stt:
-    # Model size: tiny, base, small, medium, large-v3
-    # Using "small" for faster transcription (was "medium")
+    # Provider: "deepgram" (cloud, no GPU) or "local" (faster-whisper, requires GPU)
+    provider: "deepgram"
+
+    # Deepgram settings (used when provider is "deepgram")
+    model: "nova-3"
+    language: "en"
+
+    # Local faster-whisper settings (used when provider is "local")
     model_size: "small"
-
-    # Device: cuda or cpu
     device: "cuda"
-
-    # Compute type: float16, float32, int8
     compute_type: "float16"
-
-    # Beam size for decoding (higher = more accurate, slower)
-    # Optimized for voice chat: beam_size=1 is 3-5x faster with minimal quality loss
     beam_size: 1
 
     # Language hint (null = auto-detect)
@@ -165,10 +164,17 @@ pipeline:
 
   # Text-to-Speech
   tts:
-    # TTS engine: chatterbox, coqui, piper
-    engine: "coqui"
+    # Provider: "venice" (cloud, no GPU) or "local" (chatterbox, requires GPU)
+    provider: "venice"
 
-    # Device: cuda or cpu
+    # Venice settings (used when provider is "venice")
+    venice:
+      voice: "am_liam"
+      base_url: "https://api.venice.ai/api/v1"
+      # API key from env: VENICE_API_KEY
+
+    # Local settings (used when provider is "local")
+    engine: "chatterbox"
     device: "cuda"
 
     # Streaming: generate and play audio in chunks