step3

.env.example

@@ -39,6 +39,19 @@ TTS_TOP_K=
 # are left untouched. Empty disables the prefix.
 TTS_STYLE_TAG="[conversational, casual]"
 
+# --- Custom voice post-processing (/customvoice page) ---
+# A freshly recorded reference is cleaned and level-matched before it is
+# installed. Disable the whole chain to store the plain transcode.
+CUSTOMVOICE_POSTPROCESS=true
+# Noise reduction (FFT denoise) + de-essing. Turn off for a clean studio take.
+CUSTOMVOICE_DENOISE=true
+# Integrated loudness target (LUFS) and true-peak ceiling (dBTP). Every take
+# lands on the same level, so recordings never jump in volume. Kept below the
+# shipped reference (solis.wav measures ≈-23.4 LUFS): a user take is far more
+# dynamic, so at equal LUFS its hotter peaks make the cloned voice sound louder.
+CUSTOMVOICE_TARGET_LUFS=-26
+CUSTOMVOICE_TRUE_PEAK=-5
+
 # --- Kokoro (TTS_ENGINE=kokoro) ---
 KOKORO_URL=http://localhost:8880
 KOKORO_VOICE=af_heart
@@ -64,11 +77,22 @@ VAD_NEG_THRESHOLD=0.35
 # Speech required (ms) to confirm an utterance start.
 VAD_MIN_SPEECH_MS=250
 # Silence required (ms) to close an utterance (end-of-utterance window).
-VAD_MIN_SILENCE_MS=300
+VAD_MIN_SILENCE_MS=700
 # Audio kept (ms) before the detected onset, so the first phoneme is not clipped.
 VAD_SPEECH_PAD_MS=30
 # Speech required (ms) to fire the fast barge-in onset (must be <= VAD_MIN_SPEECH_MS).
 VAD_BARGE_IN_MIN_SPEECH_MS=150
+# Utterance gate applied before transcription. A VAD segment is dropped (no STT,
+# no LLM turn) when it is shorter than VAD_MIN_UTTERANCE_MS ms (0 disables) or
+# its average speech probability is below VAD_MIN_AVG_PROB (0.0 disables). The
+# average is taken over the voiced windows only (prob >= VAD_NEG_THRESHOLD), so
+# the min_silence hangover does not dilute short words ("uno", "due"). This
+# filters the VAD false-triggers (breaths, clicks, acoustic echo, noise) that the
+# STT turns into hallucinated text ("Bye.", "She", <transcript 0 tokens>) and
+# that would each otherwise cost a full LLM turn. Raise the probability to be
+# stricter; lower it if genuine soft speech is being dropped.
+VAD_MIN_UTTERANCE_MS=300
+VAD_MIN_AVG_PROB=0.25
 # Feed the STT engine with VAD-segmented utterances (true) instead of letting
 # the worker run its own RMS segmentation (false).
 VAD_SEGMENT_STT=true
@@ -95,14 +119,15 @@ FILLER_GAIN_DUCK=0.15
 # the reply has not reached the mixer within START_FILLER_GO_DELAY ms, one of the
 # phrases below (picked at random) is played so the agent acknowledges the turn
 # while the LLM/TTS are still working. If audio arrives in time the cue is
-# dropped. It shares the filler gain, so it ducks out of the way once the reply
-# starts.
+# dropped. It shares the filler gain and is faded out (then cut) as soon as the
+# reply starts playing, so it does not overlap the answer.
 START_FILLER_ENABLED=true
 # How long (ms) to wait for the first audio chunk of the reply before playing the
 # start filler (default 1200). Measured from chunk ARRIVAL (not from playback
 # start, which would also include the TTS_PREBUFFER_MS jitter buffer). Same idea
-# as FILLER_DELAY_MS, but shorter.
-START_FILLER_GO_DELAY=500
+# as FILLER_DELAY_MS, but shorter. Raised above the default so the cue only fires
+# when the first audio is genuinely late.
+START_FILLER_GO_DELAY=1800
 # Phrases spoken by the start filler, ONE PER LINE (a single line with "|"
 # separators also works). One clip is synthesized per phrase and a RANDOM one is
 # played when the reply is late, so the cue varies turn to turn. The bank follows