Qwen3 ASR as default with tiny ASR edge as option

app.py

@@ -39,10 +39,11 @@ from core.assistant import Assistant
 # Configuration
 WAKEWORD = config['general']['wakeword']
 USE_AI = config['general']['use_ai']
+USE_TINY_ASR = config['general'].get('use_tiny_asr', False)
 
 def main():
     """Main entry point for the voice assistant."""
-    assistant = Assistant(wakeword=WAKEWORD, use_ai=USE_AI)
+    assistant = Assistant(wakeword=WAKEWORD, use_ai=USE_AI, use_tiny_asr=USE_TINY_ASR)
     assistant.run()
 
 

core/assistant.py

@@ -11,7 +11,6 @@ import threading
 from typing import Optional
 
 from .audio import AudioCapture
-from .asr import load_asr_model, stream_generator
 from .tts import speak_stream, remove_emoji
 from .slm import load_slm, generate_slm
 
@@ -34,20 +33,23 @@ class Assistant:
         grammar: JSON grammar for structured output
     """
 
-    def __init__(self, wakeword: str, use_ai: bool):
+    def __init__(self, wakeword: str, use_ai: bool, use_tiny_asr: bool = False):
         """
         Initialize the assistant.
 
         Args:
             wakeword: Activation phrase to listen for
-            slm_model_path: Path to SLM model (empty to disable AI)
+            use_ai: Whether to use the SLM for intent detection
+            use_tiny_asr: Whether to use Moonshine Tiny ASR instead of Qwen ASR
         """
         self.wakeword = wakeword.lower()
         self.use_ai = use_ai
+        self.use_tiny_asr = use_tiny_asr
         self.audio_capture = AudioCapture()
 
         # Models loaded lazily in transcriber thread
         self.asr_pipe = None
+        self.asr_stream_generator = None
         self.slm_model = None
         self.grammar = None
         self.intent_prompt = None
@@ -55,7 +57,15 @@ class Assistant:
 
     def _load_models(self):
         """Load ASR and optionally SLM models."""
+        if self.use_tiny_asr:
+            from .asr_tiny import load_asr_model, stream_generator
+            logger.info("Using Moonshine Tiny ASR")
+        else:
+            from .asr import load_asr_model, stream_generator
+            logger.info("Using Qwen ASR")
+
         self.asr_pipe = load_asr_model()
+        self.asr_stream_generator = stream_generator
 
         if self.use_ai:
             self.grammar, self.slm_model = load_slm()
@@ -151,7 +161,7 @@ class Assistant:
         logger.info("Transcriber started")
 
         for result in self.asr_pipe(
-            stream_generator(self.audio_capture.audio_queue),
+            self.asr_stream_generator(self.audio_capture.audio_queue),
             batch_size=1,
             generate_kwargs={"max_new_tokens": 256}
         ):

readme.md

@@ -28,7 +28,7 @@ A privacy-focused voice assistant that runs speech recognition, text-to-speech,
                            +--------+---------+
                                     |
                            +--------v---------+
-                           |  Moonshine ASR   |
+                           |    Qwen3 ASR     |
                            |  (Speech→Text)   |
                            +--------+---------+
                                     |
@@ -92,7 +92,7 @@ The launch script handles model downloads automatically:
 
 Or manually download:
 - [Qwen3-4B-Instruct GGUF](https://huggingface.co/Qwen) → `data/models/`
-- Moonshine and Kokoro download automatically on first run
+- Qwen3-ASR, Moonshine Tiny, and Kokoro download automatically on first run
 
 ### 4. Run
 
@@ -120,8 +120,14 @@ The launch script:
 ```yaml
 general:
   wakeword: "computer"       # Activation phrase
+  use_ai: true               # Enable SLM for intent detection
+  use_tiny_asr: false        # Use Moonshine Tiny ASR for edge devices
 ```
 
+**ASR Options:**
+- `use_tiny_asr: false` (default) — Uses Qwen3-ASR-0.6B for higher accuracy
+- `use_tiny_asr: true` — Uses Moonshine Tiny for low-resource edge devices
+
 ### Spotify
 
 1. Create an app at [Spotify Developer Dashboard](https://developer.spotify.com/dashboard)
@@ -226,7 +232,8 @@ fulloch/
 ├── app.py              # Entry point
 ├── core/               # Core modules
 │   ├── audio.py        # Audio capture and silence detection
-│   ├── asr.py          # Moonshine speech recognition
+│   ├── asr.py          # Qwen3 ASR (default)
+│   ├── asr_tiny.py     # Moonshine Tiny ASR (edge devices)
 │   ├── tts.py          # Kokoro text-to-speech
 │   ├── slm.py          # Qwen language model
 │   └── assistant.py    # Main orchestration