v2 - works

amazing made a screecast out of this

python/tool-speechtotext/assistant.py

@@ -2,25 +2,38 @@ import sounddevice as sd
 import numpy as np
 import pyperclip
 import requests
+import sys
 from faster_whisper import WhisperModel
 
-# Configuration
-MODEL_SIZE = "base" # "base" is fast, "small" is more accurate
-OLLAMA_URL = "http://localhost:11435/api/generate"
-OLLAMA_MODEL = "llama3"
+import os
+os.environ["CT2_CUDA_ALLOW_FP16"] = "1"
+
+# --- Configuration ---
+MODEL_SIZE = "medium"  # Options: "base", "small", "medium", "large-v3"
+OLLAMA_URL = "http://localhost:11434/api/generate" # Default is 11434
+OLLAMA_MODEL = "qwen3:latest"
 
 # Load Whisper on GPU 
-model = WhisperModel(MODEL_SIZE, device="cuda", compute_type="default")
+# float16 is faster and uses less VRAM on NVIDIA cards
+print("Loading Whisper model...")
+try:
+    model = WhisperModel(MODEL_SIZE, device="cuda", compute_type="float16")
+except Exception as e:
+    print(f"Error loading GPU: {e}")
+    print("Falling back to CPU (Check your CUDA/cuDNN installation)")
+    model = WhisperModel(MODEL_SIZE, device="cuda", compute_type="int16")
 
 def record_audio():
-
-    fs = 16000  # Sample rate
-    print("\n--- Press Enter to START recording ---")
+    fs = 16000 
+    print("\n[READY] Press Enter to START recording...")
     input()
-    print("Recording... Press Enter to STOP.")
+    print("[RECORDING] Press Enter to STOP...")
     
     recording = []
+    
     def callback(indata, frames, time, status):
+        if status:
+            print(status, file=sys.stderr)
         recording.append(indata.copy())
 
     with sd.InputStream(samplerate=fs, channels=1, callback=callback):
@@ -29,17 +42,24 @@ def record_audio():
     return np.concatenate(recording, axis=0)
 
 def main():
+    print(f"System active. Model: {OLLAMA_MODEL}")
     while True:
+        try:
             audio_data = record_audio()
             
-        # Transcribe
+            print("[TRANSCRIBING]...")
             segments, _ = model.transcribe(audio_data.flatten(), beam_size=5)
             text = "".join([segment.text for segment in segments]).strip()
             
+            if not text:
+                print("No speech detected. Try again.")
+                continue
+
             print(f"You said: {text}")
-        pyperclip.copy(text) # Copies to clipboard automatically
+            pyperclip.copy(text)
 
             # Send to Ollama
+            print(f"[OLLAMA] Thinking...")
             response = requests.post(OLLAMA_URL, json={
                 "model": OLLAMA_MODEL,
                 "prompt": text,
@@ -49,5 +69,11 @@ def main():
             result = response.json().get("response", "")
             print(f"\nLLM Response:\n{result}\n")
             
+        except KeyboardInterrupt:
+            print("\nExiting...")
+            break
+        except Exception as e:
+            print(f"An error occurred: {e}")
+
 if __name__ == "__main__":
     main()