Speak announcements with Kokoro, leading silence against clipped first word

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

1 files changed, 28 insertions, 1 deletions

diff --git a/cal-notif b/cal-notif
index 3726200..d577d13 100755
--- a/cal-notif
+++ b/cal-notif
@@ -296,13 +296,40 @@ def announce(occ, now, vcfg):
 
 # ---- voice ----
 
+def say(text, vcfg):
+    import numpy as np
+    from kokoro_onnx import Kokoro
+    k = Kokoro(vcfg["model"], vcfg["voices"])
+    style = sum(w * k.get_voice_style(n) for n, w in vcfg["blend"].items())
+    audio, sr = k.create(text, voice=style, lang=vcfg["lang"])
+    # A new playback stream wakes the suspended sink and loses its first
+    # ~0.2 s: lead with silence so it eats that, not the first word.
+    lead = bytes(int(sr * vcfg["lead_silence"]) * 2)
+    pcm = (np.clip(audio, -1, 1) * 32767).astype("<i2").tobytes()
+    subprocess.run(["aplay", "-q", "-r", str(sr), "-f", "S16_LE", "-c", "1"],
+                   input=lead + pcm, check=True)
+
+
+def voice_ready(vcfg):
+    if not vcfg["enabled"]:
+        return False
+    missing = [p for p in (vcfg["model"], vcfg["voices"]) if not Path(p).exists()]
+    if missing or not shutil.which("aplay"):
+        print(f"cal-notif: voice off, missing {missing or 'aplay'}", file=sys.stderr)
+        return False
+    return True
+
 # ---- daemon ----
 
 # ---- picker ----
 
 
 def main():
-    sys.exit(__doc__)
+    cmd = sys.argv[1:2]
+    if cmd == ["say"] and len(sys.argv) > 2:
+        say(" ".join(sys.argv[2:]), load_config()["voice"])
+    else:
+        sys.exit(__doc__)
 
 
 if __name__ == "__main__":