feat: add webrtcvad-based smart command-end detection

Replace the fixed 10s post-wake-word listen window with a VAD-driven
loop. Once at least 500ms of speech have been seen, 1.2s of silence
closes the window. Hard cap at 15s as a safety net.

Tunable in config.py via VAD_AGGRESSIVENESS, COMMAND_END_SILENCE_MS,
COMMAND_MIN_SPEECH_MS, COMMAND_MIN_LISTEN_MS, COMMAND_MAX_LISTEN_MS.

config.py

@@ -26,3 +26,11 @@ CHROME_PATH = 'C:/Program Files (x86)/Google/Chrome/Application/chrome.exe %s'
 GROQ_TOKEN = os.getenv('GROQ_TOKEN')
 GROQ_BASE_URL = "https://api.groq.com/openai/v1"
 GROQ_MODEL = "llama-3.3-70b-versatile"
+
+# VAD (webrtcvad) — определяет конец команды по тишине вместо фиксированных 10 сек.
+# 0..3, выше — агрессивнее режет шум (но и обрезает речь).
+VAD_AGGRESSIVENESS = 2
+COMMAND_END_SILENCE_MS = 1200
+COMMAND_MIN_SPEECH_MS = 500
+COMMAND_MIN_LISTEN_MS = 1000
+COMMAND_MAX_LISTEN_MS = 15000

main.py

@@ -13,6 +13,7 @@ import openai
 from openai import OpenAI
 import simpleaudio as sa
 import vosk
+import webrtcvad
 import yaml
 from comtypes import CLSCTX_ALL
 from fuzzywuzzy import fuzz
@@ -54,6 +55,10 @@ wake_rec = vosk.KaldiRecognizer(model, samplerate, wake_grammar)
 kaldi_rec = vosk.KaldiRecognizer(model, samplerate)
 q = queue.Queue()
 
+VAD_FRAME_MS = 30
+VAD_FRAME_BYTES = int(samplerate * VAD_FRAME_MS / 1000) * 2
+vad = webrtcvad.Vad(config.VAD_AGGRESSIVENESS)
+
 
 def gpt_answer():
     global message_log
@@ -301,20 +306,43 @@ while True:
 
         play("greet", True)
         print("Yes, sir.")
-        # reset the command recognizer so stale audio doesn't bleed into it
         kaldi_rec.Reset()
-        ltc = time.time()
+        listen_started = time.time()
+        speech_ms = 0
+        silence_ms = 0
+        vad_buf = b""
+        finalize = False
+
+        while True:
+            elapsed_ms = (time.time() - listen_started) * 1000
+            if elapsed_ms >= config.COMMAND_MAX_LISTEN_MS:
+                break
 
-        while time.time() - ltc <= 10:
             pcm = recorder.read()
             sp = struct.pack("h" * len(pcm), *pcm)
+            kaldi_rec.AcceptWaveform(sp)
 
-            if kaldi_rec.AcceptWaveform(sp):
-                if va_respond(json.loads(kaldi_rec.Result())["text"]):
-                    ltc = time.time()
+            vad_buf += sp
+            while len(vad_buf) >= VAD_FRAME_BYTES:
+                frame = vad_buf[:VAD_FRAME_BYTES]
+                vad_buf = vad_buf[VAD_FRAME_BYTES:]
+                if vad.is_speech(frame, samplerate):
+                    speech_ms += VAD_FRAME_MS
+                    silence_ms = 0
+                else:
+                    silence_ms += VAD_FRAME_MS
 
+            if elapsed_ms < config.COMMAND_MIN_LISTEN_MS:
+                continue
+            if speech_ms >= config.COMMAND_MIN_SPEECH_MS and silence_ms >= config.COMMAND_END_SILENCE_MS:
+                finalize = True
                 break
 
+        if finalize or speech_ms > 0:
+            text = json.loads(kaldi_rec.FinalResult()).get("text", "")
+            if text:
+                va_respond(text)
+
         wake_rec.Reset()
 
     except Exception as err:

requirements.txt

@@ -4,6 +4,9 @@ pvrecorder
 # Vosk
 vosk
 
+# Voice activity detection (smart command-end)
+webrtcvad-wheels
+
 # Groq / OpenAI-compatible LLM
 openai>=1.0