v1.18.1 — TTS: strip JA full-stops before Kokoro misaki phonemizer

asr-worker/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "asr-worker"
-version = "1.18.0"
+version = "1.18.1"
 description = "Local ASR worker for english-trainer (Parakeet EN/JA routing)"
 readme = "README.md"
 requires-python = ">=3.10"

mcp-server/dist/index.js

@@ -3,7 +3,7 @@ import { randomUUID } from 'node:crypto';
 import { createFramedMessageParser, writeMessage } from './mcp_stdio.js';
 
 const SERVER_NAME = 'minimum-headroom';
-const SERVER_VERSION = '1.18.0';
+const SERVER_VERSION = '1.18.1';
 const PROTOCOL_VERSION = '2024-11-05';
 const FACE_WS_URL = process.env.FACE_WS_URL ?? 'ws://127.0.0.1:8765/ws';
 const FACE_AUTH_TOKEN = (() => {

package.json

@@ -1,6 +1,6 @@
 {
   "name": "minimum-headroom",
-  "version": "1.18.0",
+  "version": "1.18.1",
   "private": true,
   "type": "module",
   "scripts": {

tts-worker/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "minimum-headroom-tts-worker"
-version = "1.18.0"
+version = "1.18.1"
 description = "Minimum Headroom TTS worker (Kokoro ONNX default, optional Qwen3-TTS)"
 readme = "README.md"
 requires-python = ">=3.12"

tts-worker/src/tts_worker/kokoro_engine.py

@@ -11,6 +11,7 @@
 
 from .chunking import TextChunk, split_text_chunks
 from .engine import EngineMetadata
+from .kokoro_text import strip_japanese_silent_punctuation
 
 
 @dataclass(frozen=True)
@@ -105,6 +106,8 @@ def synthesize_chunks(self, chunks: Iterable[TextChunk], *, voice_override: str
       source_text = chunk.text
       if chunk.is_phonemes:
         source_text = self._to_ja_phonemes(chunk.text)
+        if not source_text:
+          continue
 
       audio, chunk_rate = self._kokoro_create(
         source_text,
@@ -130,10 +133,14 @@ def synthesize_chunks(self, chunks: Iterable[TextChunk], *, voice_override: str
     return combined.astype(np.float32, copy=False), sample_rate
 
   def _to_ja_phonemes(self, text: str) -> str:
+    cleaned = strip_japanese_silent_punctuation(text).strip()
+    if not cleaned:
+      return ''
+
     capture = io.StringIO()
     # Some pyopenjtalk-backed helpers print progress text to stdout; keep protocol stdout JSON-only.
     with contextlib.redirect_stdout(capture):
-      raw = self._ja_g2p(text)
+      raw = self._ja_g2p(cleaned)
 
     echoed = capture.getvalue().strip()
     if echoed:

tts-worker/src/tts_worker/kokoro_text.py

@@ -0,0 +1,17 @@
+from __future__ import annotations
+
+import re
+
+# Misaki's pyopenjtalk-backed Japanese G2P maps the JA full-stop 「。」
+# (and its fullwidth ASCII twin 「．」) to an audible phoneme rather than
+# silence, so Kokoro renders chunk endings as a short "ye"-like sound.
+# Chunk boundaries already separate sentences, so dropping these
+# characters before phonemization removes the artifact without losing
+# meaningful prosody. Other JA punctuation (「、」「！」「？」「・」) is
+# left in place because it either drives prosodic pausing or has not
+# been observed to produce an artifact.
+_SILENT_PUNCT_RE = re.compile(r'[。．]+')
+
+
+def strip_japanese_silent_punctuation(text: str) -> str:
+  return _SILENT_PUNCT_RE.sub('', text)

tts-worker/tests/test_kokoro_text.py

@@ -0,0 +1,54 @@
+from __future__ import annotations
+
+import sys
+import unittest
+from pathlib import Path
+
+
+ROOT_DIR = Path(__file__).resolve().parents[2]
+SRC_DIR = ROOT_DIR / 'tts-worker' / 'src'
+if str(SRC_DIR) not in sys.path:
+  sys.path.insert(0, str(SRC_DIR))
+
+from tts_worker.kokoro_text import strip_japanese_silent_punctuation
+
+
+class StripJapaneseSilentPunctuationTests(unittest.TestCase):
+  def test_drops_trailing_full_stop(self) -> None:
+    self.assertEqual(strip_japanese_silent_punctuation('あ。'), 'あ')
+
+  def test_drops_runs_of_full_stops(self) -> None:
+    self.assertEqual(strip_japanese_silent_punctuation('あ。。。。。'), 'あ')
+
+  def test_drops_internal_and_trailing_full_stops(self) -> None:
+    self.assertEqual(
+      strip_japanese_silent_punctuation('あ。い。う。'), 'あいう'
+    )
+
+  def test_drops_fullwidth_period(self) -> None:
+    self.assertEqual(strip_japanese_silent_punctuation('あ．'), 'あ')
+
+  def test_preserves_other_japanese_punctuation(self) -> None:
+    # Comma, exclamation, question mark, middle dot, and ellipsis are
+    # kept because they either drive prosody or have not been observed
+    # to produce an artifact.
+    self.assertEqual(
+      strip_japanese_silent_punctuation('あ、い！う？え・お…'),
+      'あ、い！う？え・お…',
+    )
+
+  def test_returns_empty_for_punctuation_only_input(self) -> None:
+    self.assertEqual(strip_japanese_silent_punctuation('。。。'), '')
+
+  def test_passes_through_empty_string(self) -> None:
+    self.assertEqual(strip_japanese_silent_punctuation(''), '')
+
+  def test_passes_through_ascii_text(self) -> None:
+    self.assertEqual(
+      strip_japanese_silent_punctuation('Hello, world.'),
+      'Hello, world.',
+    )
+
+
+if __name__ == '__main__':
+  unittest.main()