stt-server.py

@@ -16,13 +16,15 @@ Every WebSocket connection receives the full event stream from the moment it
 connects — no subscription handshake required.
 
 All log/status messages go to stderr. Stdout is machine-readable events only.
+Pass --verbose to enable info logging (startup, VAD events, transcripts).
+Errors always go to stderr regardless of verbosity.
 
 Environment:
   STT_PORT   WebSocket port (default: 11501)
 
 Usage:
   ./stt-server.py
-  ./stt-server.py --model large-v3 --device cuda --compute-type int8_float16
+  ./stt-server.py --model large-v3 --device cuda --compute-type int8_float16 --verbose
 """
 
 import sys
@@ -45,9 +47,10 @@ HISTORY_SAMPLES  = 960000   # 60s ring buffer for pre-roll
 PORT             = int(__import__('os').environ.get('STT_PORT', 11501))
 
 
-def log(msg):
-	sys.stderr.write(f'[stt] {msg}\n')
-	sys.stderr.flush()
+def log(msg, error=False):
+	if error or verbose:
+		sys.stderr.write(f'[stt] {msg}\n')
+		sys.stderr.flush()
 
 
 # --- WebSocket broadcast ---
@@ -118,7 +121,9 @@ parser = argparse.ArgumentParser()
 parser.add_argument('--model',        default='base.en')
 parser.add_argument('--device',       default='cuda')
 parser.add_argument('--compute-type', default='int8_float16')
-args = parser.parse_args()
+parser.add_argument('--verbose', '-v', action='store_true')
+args    = parser.parse_args()
+verbose = args.verbose
 
 log(f'loading faster-whisper {args.model} ({args.device}, {args.compute_type})...')
 from faster_whisper import WhisperModel
@@ -126,7 +131,7 @@ try:
 	model = WhisperModel(args.model, device=args.device, compute_type=args.compute_type)
 	log(f'model ready on {args.device}')
 except Exception as e:
-	log(f'{args.device} failed ({e}), falling back to cpu')
+	log(f'{args.device} failed ({e}), falling back to cpu', error=True)
 	model = WhisperModel(args.model, device='cpu', compute_type='int8')
 	log('model ready on cpu')
 
@@ -197,7 +202,7 @@ def transcription_worker():
 				emit({'event': 'transcript', 'text': text.strip(), 'words': words, 'duration': round(duration, 3)})
 		except Exception:
 			msg = traceback.format_exc()
-			log(f'transcription error:\n{msg}')
+			log(f'transcription error:\n{msg}', error=True)
 			emit({'event': 'error', 'message': msg})
 		finally:
 			transcription_queue.task_done()