Vorlesen: Sprachauswahl DE/EN mit englischer Piper-Stimme (SW v43)

Reader-Header bekommt ein #reader-tts-lang-Dropdown (Auswahl, keine
Sprach-Erkennung — persistiert in localStorage); /tts/synthesize/ nimmt jetzt
{"lang": "de"|"en"} entgegen. tts/piper_engine.py lädt pro Sprache eine eigene
PiperVoice lazy (settings.TTS_VOICES), Docker-Image lädt zusätzlich
en_US-lessac-medium beim Build.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Xb6yX2S9aTGepYA9JFba9x

Dockerfile

@@ -10,16 +10,20 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
 COPY requirements.txt .
 RUN pip install --no-cache-dir -r requirements.txt
 
-# Piper voice for the reader's read-aloud feature (see tts/piper_engine.py).
+# Piper voices for the reader's read-aloud feature (see tts/piper_engine.py).
 # Fetched at build time rather than kept in git or bind-mounted — there's no
 # existing volume-mount pattern for extra binary assets in this repo, and
-# baking it into the image keeps the watchtower "just pull the new image"
+# baking them into the image keeps the watchtower "just pull the new image"
 # deploy flow working unchanged.
 RUN mkdir -p /app/tts_models && \
     curl -fsSL -o /app/tts_models/de_DE-thorsten-medium.onnx \
       https://huggingface.co/rhasspy/piper-voices/resolve/main/de/de_DE/thorsten/medium/de_DE-thorsten-medium.onnx && \
     curl -fsSL -o /app/tts_models/de_DE-thorsten-medium.onnx.json \
-      https://huggingface.co/rhasspy/piper-voices/resolve/main/de/de_DE/thorsten/medium/de_DE-thorsten-medium.onnx.json
+      https://huggingface.co/rhasspy/piper-voices/resolve/main/de/de_DE/thorsten/medium/de_DE-thorsten-medium.onnx.json && \
+    curl -fsSL -o /app/tts_models/en_US-lessac-medium.onnx \
+      https://huggingface.co/rhasspy/piper-voices/resolve/main/en/en_US/lessac/medium/en_US-lessac-medium.onnx && \
+    curl -fsSL -o /app/tts_models/en_US-lessac-medium.onnx.json \
+      https://huggingface.co/rhasspy/piper-voices/resolve/main/en/en_US/lessac/medium/en_US-lessac-medium.onnx.json
 
 ARG BUILD_TIME
 ENV BUILD_TIME=${BUILD_TIME}

diora/settings.py

@@ -152,6 +152,10 @@ WEBDAV_ALLOW_PRIVATE_HOSTS = os.environ.get('WEBDAV_ALLOW_PRIVATE_HOSTS', 'False
 
 BUILD_TIME = os.environ.get('BUILD_TIME', '')
 
-# Piper voice model for the reader's read-aloud feature (see tts/piper_engine.py).
-TTS_MODEL_PATH = os.environ.get(
-    'TTS_MODEL_PATH', str(BASE_DIR / 'tts_models' / 'de_DE-thorsten-medium.onnx'))
+# Piper voice models for the reader's read-aloud feature (see tts/piper_engine.py).
+TTS_VOICES = {
+    'de': os.environ.get(
+        'TTS_MODEL_PATH_DE', str(BASE_DIR / 'tts_models' / 'de_DE-thorsten-medium.onnx')),
+    'en': os.environ.get(
+        'TTS_MODEL_PATH_EN', str(BASE_DIR / 'tts_models' / 'en_US-lessac-medium.onnx')),
+}

static/css/app.css

@@ -1636,6 +1636,12 @@ body.dnd-mode .timer-display {
   padding: 4px 6px; font-size: 0.82rem; cursor: pointer;
 }
 
+.tts-lang-select {
+  background: var(--surface, #1e1e2e); color: var(--fg, #fff);
+  border: 1px solid var(--border, #444); border-radius: 4px;
+  padding: 2px 4px; font-size: 0.8rem; cursor: pointer;
+}
+
 .reader-marker-btn-mobile { display: none; }
 
 @media (max-width: 600px) {

static/js/app.js

@@ -4150,6 +4150,12 @@ let ttsRunToken = 0;
 let ttsAudio = null;
 let ttsCurrentMark = null;
 let ttsBarEl = null;
+let ttsLang = localStorage.getItem('diora_tts_lang') || 'de';
+
+function setTtsLang(lang) {
+  ttsLang = lang;
+  localStorage.setItem('diora_tts_lang', lang);
+}
 
 function _ttsSplitSentences(text) {
   const raw = text.split(/(?<=[.!?])\s+/).map(s => s.trim()).filter(Boolean);
@@ -4172,7 +4178,7 @@ async function _ttsFetchAudio(sentence) {
   const resp = await fetch('/tts/synthesize/', {
     method: 'POST',
     headers: {'Content-Type': 'application/json'},
-    body: JSON.stringify({text: sentence}),
+    body: JSON.stringify({text: sentence, lang: ttsLang}),
   });
   if (!resp.ok) throw new Error('tts request failed');
   return await resp.blob();
@@ -6769,6 +6775,10 @@ function openRadioSidebar() {
     setVolume(vol);
   }
 
+  // Restore persisted read-aloud language
+  const ttsLangSelect = $('reader-tts-lang');
+  if (ttsLangSelect) ttsLangSelect.value = ttsLang;
+
   // Load recommendations on page load
   loadRecommendations();
 

static/js/sw.js

@@ -2,7 +2,7 @@
  * diora service worker — caches the app shell for offline use.
  */
 
-const CACHE = 'diora-v42';
+const CACHE = 'diora-v43';
 const PODCAST_CACHE = 'diora-podcast-v1';
 const SHELL = [
   '/static/css/app.css',

templates/radio/player.html

@@ -354,6 +354,10 @@
       </span>
       <button class="btn-icon" id="reader-search-btn"   onclick="toggleReaderSearch()"        title="Search">⌕</button>
       <button class="btn-icon" id="reader-settings-btn" onclick="toggleSettingsPanel()"        title="Font &amp; layout">⚙</button>
+      <select id="reader-tts-lang" class="tts-lang-select" title="Vorlese-Sprache" onchange="setTtsLang(this.value)">
+        <option value="de">DE</option>
+        <option value="en">EN</option>
+      </select>
       <button class="btn-icon" id="reader-tts-btn"      onclick="toggleReadAloud()"           title="Vorlesen">▶</button>
       <button class="btn-icon" id="reader-bookmark-btn" onclick="addBookmark()"               title="Bookmark">★</button>
       <button class="btn-icon" id="reader-bm-list-btn"  onclick="openBookmarksSidebar()"      title="Bookmarks">▤</button>

tts/piper_engine.py

@@ -4,31 +4,36 @@ import wave
 
 from django.conf import settings
 
-# Lazy per-process singleton: each gunicorn worker loads its own copy on first
-# use rather than all workers loading the ONNX model at startup (the host runs
-# several other containers with limited spare RAM).
-_voice = None
+DEFAULT_LANGUAGE = 'de'
+SUPPORTED_LANGUAGES = tuple(settings.TTS_VOICES.keys())
+
+# Lazy per-process singletons, one per language: each gunicorn worker loads a
+# voice only once it's actually requested, rather than all workers loading
+# every model at startup (the host runs several other containers with limited
+# spare RAM). One lock guards both the lazy-load and the inference call below
+# — a single self-hosted user never needs concurrent synthesis across
+# languages, so there's no reason for a lock per voice.
+_voices = {}
 _lock = threading.Lock()
 
 
-def _get_voice():
-    global _voice
-    if _voice is None:
+def _get_voice(lang):
+    if lang not in _voices:
         with _lock:
-            if _voice is None:
+            if lang not in _voices:
                 from piper import PiperVoice
-                _voice = PiperVoice.load(str(settings.TTS_MODEL_PATH))
-    return _voice
+                _voices[lang] = PiperVoice.load(str(settings.TTS_VOICES[lang]))
+    return _voices[lang]
 
 
-def synthesize_wav(text):
-    """Synthesize `text` to WAV bytes.
+def synthesize_wav(text, lang=DEFAULT_LANGUAGE):
+    """Synthesize `text` (in `lang`) to WAV bytes.
 
     Never persists or logs `text` — callers must not log it either. The lock
     also serializes inference, since one onnxruntime session isn't meant to
     run concurrent calls within a process.
     """
-    voice = _get_voice()
+    voice = _get_voice(lang)
     buf = io.BytesIO()
     with _lock:
         with wave.open(buf, 'wb') as wav_file:

tts/tests.py

@@ -31,11 +31,25 @@ class TtsSynthesizeTests(TestCase):
         self.assertEqual(resp.status_code, 400)
 
     @patch.object(piper_engine, 'synthesize_wav', return_value=b'RIFF....WAVEfmt fake')
-    def test_synthesizes_audio(self, mock_synth):
+    def test_synthesizes_audio_default_lang(self, mock_synth):
         self.client.force_login(self.user)
         resp = self.client.post(
             '/tts/synthesize/', {'text': 'Hallo Welt.'}, content_type='application/json')
         self.assertEqual(resp.status_code, 200)
         self.assertEqual(resp['Content-Type'], 'audio/wav')
         self.assertEqual(b''.join(resp.streaming_content), b'RIFF....WAVEfmt fake')
-        mock_synth.assert_called_once_with('Hallo Welt.')
+        mock_synth.assert_called_once_with('Hallo Welt.', 'de')
+
+    @patch.object(piper_engine, 'synthesize_wav', return_value=b'RIFF....WAVEfmt fake')
+    def test_synthesizes_audio_explicit_lang(self, mock_synth):
+        self.client.force_login(self.user)
+        resp = self.client.post(
+            '/tts/synthesize/', {'text': 'Hello world.', 'lang': 'en'}, content_type='application/json')
+        self.assertEqual(resp.status_code, 200)
+        mock_synth.assert_called_once_with('Hello world.', 'en')
+
+    def test_rejects_unsupported_lang(self):
+        self.client.force_login(self.user)
+        resp = self.client.post(
+            '/tts/synthesize/', {'text': 'Hallo', 'lang': 'fr'}, content_type='application/json')
+        self.assertEqual(resp.status_code, 400)

tts/views.py

@@ -42,8 +42,12 @@ def synthesize(request):
     if len(text) > MAX_TEXT_LENGTH:
         return JsonResponse({'error': f'text exceeds {MAX_TEXT_LENGTH} characters'}, status=400)
 
+    lang = body.get('lang', piper_engine.DEFAULT_LANGUAGE)
+    if lang not in piper_engine.SUPPORTED_LANGUAGES:
+        return JsonResponse({'error': 'unsupported lang'}, status=400)
+
     try:
-        audio = _synth_pool.apply(piper_engine.synthesize_wav, (text,))
+        audio = _synth_pool.apply(piper_engine.synthesize_wav, (text, lang))
     except Exception:
         return JsonResponse({'error': 'synthesis failed'}, status=500)