diff --git a/main.py b/main.py
index a05c07e..a96a9f7 100644
--- a/main.py
+++ b/main.py
@@ -10,14 +10,66 @@ import time
 import webbrowser
 from ctypes import POINTER, cast
 
+# Print a startup banner BEFORE the heavy imports — otherwise the user
+# sees an empty terminal for 10-60 seconds while torch / vosk / silero
+# load. This used to be the #1 "did it crash?" support issue.
+print("=" * 60, flush=True)
+print("  J.A.R.V.I.S.  loading (Python edition)...", flush=True)
+print("  Heavy modules (vosk / torch / pvrecorder) take a few seconds.", flush=True)
+print("=" * 60, flush=True)
+
+# Helpful diagnostic: if a critical dep is missing, tell the user EXACTLY
+# what to install instead of dumping a raw ModuleNotFoundError. Pulls
+# every required package by name and reports them in one go.
+_MISSING_DEPS = []
+for _modname, _pkg in [
+    ("numpy", "numpy"),
+    ("openai", "openai>=1.0"),
+    ("vosk", "vosk"),
+    ("webrtcvad", "webrtcvad"),
+    ("yaml", "PyYAML"),
+    ("comtypes", "comtypes"),
+    ("fuzzywuzzy", "fuzzywuzzy"),
+    ("pvrecorder", "pvrecorder"),
+    ("pycaw", "pycaw"),
+    ("rich", "rich"),
+]:
+    try:
+        __import__(_modname)
+    except ImportError:
+        _MISSING_DEPS.append(_pkg)
+if _MISSING_DEPS:
+    print("\n[ERROR] Missing Python packages:", flush=True)
+    for _p in _MISSING_DEPS:
+        print(f"  - {_p}", flush=True)
+    print("\nRun:  pip install -r requirements.txt", flush=True)
+    print("Or use the bundled venv: launch via run.bat.\n", flush=True)
+    sys.exit(2)
+
 import numpy as np
 import openai
 from openai import OpenAI
-import simpleaudio as sa
 import vosk
 import webrtcvad
 import yaml
 
+# simpleaudio needs MSVC to build from source; gracefully fall back to the
+# stdlib's `winsound` on Windows when it isn't installed.
+try:
+    import simpleaudio as sa
+    _HAS_SIMPLEAUDIO = True
+except ImportError:
+    sa = None
+    _HAS_SIMPLEAUDIO = False
+    try:
+        import winsound as _winsound
+    except ImportError:
+        _winsound = None
+    if _winsound is None:
+        print("[WARN] Neither simpleaudio nor winsound available — sound cues disabled.", flush=True)
+    else:
+        print("[INFO] simpleaudio not installed; falling back to winsound for sound cues.", flush=True)
+
 import extensions  # new-command handlers ported from rust fork
 import plugins_store  # user plugin packs under APP_CONFIG_DIR/plugins/
 from comtypes import CLSCTX_ALL
@@ -154,14 +206,24 @@ def play(phrase, wait_done=True):
     if wait_done:
         recorder.stop()
 
-    wave_obj = sa.WaveObject.from_wave_file(filename)
-    play_obj = wave_obj.play()
+    # Prefer simpleaudio (lets us wait_done precisely); fall back to winsound
+    # on installs that couldn't build the simpleaudio C extension.
+    if _HAS_SIMPLEAUDIO and sa is not None:
+        wave_obj = sa.WaveObject.from_wave_file(filename)
+        play_obj = wave_obj.play()
+        if wait_done:
+            play_obj.wait_done()
+    elif _winsound is not None:
+        flags = _winsound.SND_FILENAME
+        if not wait_done:
+            flags |= _winsound.SND_ASYNC
+        try:
+            _winsound.PlaySound(filename, flags)
+        except RuntimeError as exc:
+            print(f"[sound] winsound playback failed: {exc}")
+    # else: no audio backend — silently skip cue
 
     if wait_done:
-        play_obj.wait_done()
-        # time.sleep((len(wave_obj.audio_data) / wave_obj.sample_rate) + 0.5)
-        # print("END")
-        # time.sleep(0.5)
         recorder.start()
 
 
diff --git a/tts.py b/tts.py
index 9930298..35ca6f9 100644
--- a/tts.py
+++ b/tts.py
@@ -1,3 +1,14 @@
+"""TTS — Silero ru_v3 voice synth with lazy initialisation.
+
+Why lazy: torch.hub.load() blocks for 30-120 seconds on the first run
+(downloading the model) and produces no progress output beyond a single
+"Using cache found in..." line. If this runs at MODULE IMPORT time, the
+whole jarvis-python startup looks frozen — exactly the bug the user hit.
+
+Fix: defer the load until first synth call. main.py prints its startup
+banner first, then this module loads on demand. Side effect: the first
+"speak" is slow (3-4s), but every print before it is visible.
+"""
 import time
 
 import numpy as np
@@ -14,13 +25,31 @@ speaker = 'aidar'  # aidar, baya, kseniya, xenia, random
 put_accent = True
 put_yo = True
 device = torch.device('cpu')  # cpu или gpu
-text = "Хауди Хо, друзья!!!"
 
-model, _ = torch.hub.load(repo_or_dir='snakers4/silero-models',
+# Initialised on first synthesize() call.
+_model = None
+
+
+def _ensure_model():
+    """Load Silero on first use. Subsequent calls are no-ops.
+
+    Prints a one-line progress hint so the user can tell the assistant is
+    busy downloading rather than hung — critical for the first run when
+    the model isn't cached yet.
+    """
+    global _model
+    if _model is not None:
+        return _model
+    print("[tts] Loading Silero ru_v3 (first use — ~30s if model isn't cached)...")
+    started = time.time()
+    m, _ = torch.hub.load(repo_or_dir='snakers4/silero-models',
                           model='silero_tts',
                           language=language,
                           speaker=model_id)
-model.to(device)
+    m.to(device)
+    _model = m
+    print(f"[tts] Silero ready ({time.time() - started:.1f}s).")
+    return _model
 
 
 def _post_process(audio):
@@ -45,11 +74,12 @@ def _post_process(audio):
 
 
 def synthesize(what: str):
-    audio = model.apply_tts(text=what + "..",
-                            speaker=speaker,
-                            sample_rate=sample_rate,
-                            put_accent=put_accent,
-                            put_yo=put_yo)
+    m = _ensure_model()
+    audio = m.apply_tts(text=what + "..",
+                        speaker=speaker,
+                        sample_rate=sample_rate,
+                        put_accent=put_accent,
+                        put_yo=put_yo)
     return _post_process(audio)