Smooth Hermes TTS sentence transitions

platform/hermes/025-cron-profile-root-fix

@@ -16,21 +16,55 @@ mkdir -p "$cli_dir"
 ln -sfn /opt/hermes/.venv/bin/hermes "$cli_dir/hermes"
 ln -sfn /opt/hermes/.venv/bin/hermes /usr/local/bin/hermes
 
-# Hermes streams each detected sentence to TTS separately. Its generic
-# sentence splitter treats periods in German abbreviations as sentence ends,
-# causing the following words to become unrelated utterances. Expand these
-# abbreviations before splitting.
+# Hermes' generic sentence splitter treats periods in German abbreviations as
+# sentence ends. It also submits every sentence as a separate generative TTS
+# request, which creates long gaps with local Qwen3-TTS. Keep the first sentence
+# immediate, then group following complete sentences into modest ~240-character
+# chunks so playback remains responsive but substantially more continuous.
 tts_target=/opt/hermes/tools/tts_streaming.py
-tts_old='        self.buf = _THINK_BLOCK_RE.sub("", self.buf + delta)'
-tts_new='        self.buf = _THINK_BLOCK_RE.sub("", self.buf + delta)
-        self.buf = re.sub(r"\bca\.(?=\s)", "circa", self.buf, flags=re.IGNORECASE)
-        self.buf = re.sub(r"\bv\.\s*a\.(?=\s)", "vor allem", self.buf, flags=re.IGNORECASE)
-        self.buf = re.sub(r"\bmax\.(?=\s)", "maximal", self.buf, flags=re.IGNORECASE)'
+if [ -f "$tts_target" ] && grep -Fq 'self.buf = _THINK_BLOCK_RE.sub("", self.buf + delta)' "$tts_target"; then
+    TTS_TARGET="$tts_target" python3 <<'PY'
+import os
+from pathlib import Path
 
-if [ -f "$tts_target" ] && grep -Fq "$tts_old" "$tts_target"; then
-    TTS_TARGET="$tts_target" TTS_OLD="$tts_old" TTS_NEW="$tts_new" python3 -c \
-        'import os; from functools import reduce; from pathlib import Path; p = Path(os.environ["TTS_TARGET"]); s = p.read_text(); lines = ["        self.buf = re.sub(r\"\\bca\\.(?=\\s)\", \"circa\", self.buf, flags=re.IGNORECASE)", "        self.buf = re.sub(r\"\\bv\\.\\s*a\\.(?=\\s)\", \"vor allem\", self.buf, flags=re.IGNORECASE)", "        self.buf = re.sub(r\"\\bmax\\.(?=\\s)\", \"maximal\", self.buf, flags=re.IGNORECASE)"]; s = reduce(lambda value, line: value.replace("\n" + line, ""), lines, s); p.write_text(s.replace(os.environ["TTS_OLD"], os.environ["TTS_NEW"], 1))'
-    echo "[tts-streaming-fix] expanded German abbreviations before sentence splitting"
+p = Path(os.environ["TTS_TARGET"])
+s = p.read_text()
+
+base_feed = '        self.buf = _THINK_BLOCK_RE.sub("", self.buf + delta)'
+abbr_lines = [
+    '        self.buf = re.sub(r"\\bca\\.(?=\\s)", "circa", self.buf, flags=re.IGNORECASE)',
+    '        self.buf = re.sub(r"\\bv\\.\\s*a\\.(?=\\s)", "vor allem", self.buf, flags=re.IGNORECASE)',
+    '        self.buf = re.sub(r"\\bmax\\.(?=\\s)", "maximal", self.buf, flags=re.IGNORECASE)',
+]
+for line in abbr_lines:
+    s = s.replace("\n" + line, "")
+
+init_old = '        self.buf = ""'
+init_extra = (
+    '\n        self.followup_target_len = 240'
+    '\n        self._emitted_first = False'
+)
+s = s.replace(init_old + init_extra, init_old)
+
+threshold_old = '            if len(head.strip()) < self.min_len:'
+threshold_new = (
+    '            threshold = (self.min_len if not self._emitted_first '
+    'else self.followup_target_len)\n'
+    '            if len(head.strip()) < threshold:'
+)
+s = s.replace(threshold_new, threshold_old)
+
+append_old = '            out.append(head)'
+append_new = append_old + '\n            self._emitted_first = True'
+s = s.replace(append_new, append_old)
+
+s = s.replace(base_feed, base_feed + "\n" + "\n".join(abbr_lines), 1)
+s = s.replace(init_old, init_old + init_extra, 1)
+s = s.replace(threshold_old, threshold_new, 1)
+s = s.replace(append_old, append_new, 1)
+p.write_text(s)
+PY
+    echo "[tts-streaming-fix] enabled German normalization and adaptive sentence grouping"
 else
     echo "[tts-streaming-fix] upstream code already fixed or layout changed; no action"
 fi