feat(radio): replace robot TTS with Piper Neural voices & add instant voice preview

- Integrate local Piper ONNX neural TTS engine (Dmitri, Irina, Denis, Ruslan) - Add GET /api/radio/preview-voice endpoint with cached pre-rendered samples - Add '▶ Прослушать голос' button in Web UI for 1-click preview before generating - Add automated filter to strip LLM refusals and fall back to clean radio phrasing
2026-10-12 16:58:59 +03:00
1 parent e5afeb06cc
commit 4ed8e0c71a
@@ -30,7 +30,7 @@ class Settings(BaseSettings):

    # AI Radio DJ Settings
    RADIO_ENABLED: bool = True
-    RADIO_VOICE: str = "ru-RU-DmitryNeural"
+    RADIO_VOICE: str = "ru_RU-dmitri-medium"
    WEATHER_CITY: str = ""
    RADIO_INTRO_INTERVAL: int = 2

@@ -22,7 +22,7 @@ from app.core.auth import (
 )
 from app.wave.rotator import wave_rotator
 from app.wave.playlist import ensure_smart_playlist
-from app.radio import radio_manager
+from app.radio import radio_manager, dj_host

 app = FastAPI(title="Navidrome Music Sync", version="1.0.0")

@@ -334,3 +334,16 @@ async def generate_radio_broadcast(req: Optional[RadioGenerateRequest] = None, u
    res = await radio_manager.generate_radio_broadcast(voice=voice)
    return res

+@app.get("/api/radio/preview-voice")
+def preview_voice(voice: str = Query(...)):
+    """Stream voice sample preview."""
+    preview_path = dj_host.get_voice_preview_path(voice)
+    if not preview_path or not os.path.exists(preview_path):
+        raise HTTPException(status_code=404, detail="Превью голоса не найдено")
+    return FileResponse(
+        preview_path,
+        media_type="audio/mpeg",
+        headers={"Content-Disposition": f"inline; filename=preview_{voice}.mp3"}
+    )
+
+
@@ -2,7 +2,7 @@ import os
 import time
 import httpx
 import asyncio
-import edge_tts
+import tempfile
 from datetime import datetime
 from typing import List, Dict, Any, Optional
 from mutagen.id3 import ID3, TIT2, TPE1, TALB, ID3NoHeaderError
@@ -10,11 +10,71 @@ from mutagen.id3 import ID3, TIT2, TPE1, TALB, ID3NoHeaderError
 from app.config import settings
 from app.radio.weather import get_current_weather

+VOICE_CATALOG: Dict[str, Dict[str, Any]] = {
+    "ru_RU-dmitri-medium": {
+        "id": "ru_RU-dmitri-medium",
+        "name": "Дмитрий (Энергичный DJ)",
+        "gender": "Мужской",
+        "desc": "Бодрый и харизматичный голос ведущего радиоэфира",
+        "engine": "piper",
+        "model_file": "ru_RU-dmitri-medium.onnx"
+    },
+    "ru_RU-irina-medium": {
+        "id": "ru_RU-irina-medium",
+        "name": "Ирина (Естественный женский)",
+        "gender": "Женский",
+        "desc": "Мягкий, естественный и дружелюбный женский тембр",
+        "engine": "piper",
+        "model_file": "ru_RU-irina-medium.onnx"
+    },
+    "ru_RU-denis-medium": {
+        "id": "ru_RU-denis-medium",
+        "name": "Денис (Глубокий баритон)",
+        "gender": "Мужской",
+        "desc": "Плотный, уверенный радио-голос с глубоким низом",
+        "engine": "piper",
+        "model_file": "ru_RU-denis-medium.onnx"
+    },
+    "ru_RU-ruslan-medium": {
+        "id": "ru_RU-ruslan-medium",
+        "name": "Руслан (Спокойный ведущий)",
+        "gender": "Мужской",
+        "desc": "Размеренный, интеллигентный и чилловый голос",
+        "engine": "piper",
+        "model_file": "ru_RU-ruslan-medium.onnx"
+    },
+    "gtts_ru": {
+        "id": "gtts_ru",
+        "name": "Google (Стандартный робот)",
+        "gender": "Женский",
+        "desc": "Классический синтез Google Translate",
+        "engine": "gtts",
+        "model_file": None
+    }
+}
+
 class RadioDJ:
    def __init__(self):
        self.ollama_url = f"{settings.OLLAMA_URL}/api/generate"
        self.model = settings.OLLAMA_MODEL
-        self.default_voice = settings.RADIO_VOICE or "ru-RU-DmitryNeural"
+        self.default_voice = settings.RADIO_VOICE or "ru_RU-dmitri-medium"
+        self.voices_dir = os.path.join(settings.DATA_DIR, "voices")
+        self.previews_dir = os.path.join(self.voices_dir, "previews")
+        os.makedirs(self.voices_dir, exist_ok=True)
+        os.makedirs(self.previews_dir, exist_ok=True)
+
+    def get_available_voices(self) -> List[Dict[str, Any]]:
+        """Return list of available voice profiles."""
+        return list(VOICE_CATALOG.values())
+
+    def get_voice_preview_path(self, voice_id: str) -> Optional[str]:
+        """Return absolute path to pre-generated preview mp3 for the given voice."""
+        if voice_id not in VOICE_CATALOG:
+            voice_id = self.default_voice
+        path = os.path.join(self.previews_dir, f"{voice_id}.mp3")
+        if os.path.exists(path) and os.path.getsize(path) > 1000:
+            return path
+        return None

    def get_time_of_day_info(self) -> Dict[str, str]:
        """Determine time of day greeting."""
@@ -76,7 +136,10 @@ class RadioDJ:
                    lines = [l.strip() for l in text.split("\n") if l.strip()]
                    if lines:
                        cleaned = " ".join(lines)
-                        if len(cleaned) > 10:
+                        refusal_words = ["извините", "не могу", "авторских прав", "я языковая модель", "как ии", "служб безопасности", "отказаться", "i cannot", "i am unable"]
+                        if any(w in cleaned.lower() for w in refusal_words):
+                            print(f"[RadioDJ] Filtered out LLM refusal: {cleaned}")
+                        elif len(cleaned) > 10:
                            return cleaned
        except Exception as e:
            print(f"[RadioDJ] Ollama generation fallback due to: {e}")
@@ -105,6 +168,55 @@ class RadioDJ:
        except Exception as tag_err:
            print(f"[RadioDJ] Warning tagging audio: {tag_err}")

+    async def _synthesize_piper(self, text: str, model_filename: str, output_file: str) -> bool:
+        """Synthesize audio using local neural Piper TTS model and convert to MP3 via ffmpeg."""
+        model_path = os.path.join(self.voices_dir, model_filename)
+        if not os.path.exists(model_path):
+            print(f"[RadioDJ] Piper model {model_filename} not found at {model_path}")
+            return False
+
+        with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp_wav:
+            wav_path = tmp_wav.name
+
+        try:
+            cmd = [
+                "piper",
+                "--model", model_path,
+                "--output_file", wav_path
+            ]
+            proc = await asyncio.create_subprocess_exec(
+                *cmd,
+                stdin=asyncio.subprocess.PIPE,
+                stdout=asyncio.subprocess.PIPE,
+                stderr=asyncio.subprocess.PIPE
+            )
+            stdout, stderr = await proc.communicate(input=text.encode("utf-8"))
+            if proc.returncode != 0:
+                print(f"[RadioDJ] Piper error: {stderr.decode('utf-8', errors='ignore')}")
+                return False
+
+            # Convert WAV to MP3 using ffmpeg
+            ffmpeg_cmd = [
+                "ffmpeg", "-y",
+                "-i", wav_path,
+                "-codec:a", "libmp3lame",
+                "-b:a", "192k",
+                output_file
+            ]
+            ff_proc = await asyncio.create_subprocess_exec(
+                *ffmpeg_cmd,
+                stdout=asyncio.subprocess.PIPE,
+                stderr=asyncio.subprocess.PIPE
+            )
+            await ff_proc.communicate()
+            return os.path.exists(output_file) and os.path.getsize(output_file) > 500
+        except Exception as e:
+            print(f"[RadioDJ] Piper synthesis exception: {e}")
+            return False
+        finally:
+            if os.path.exists(wav_path):
+                os.remove(wav_path)
+
    async def synthesize_speech(
        self,
        text: str,
@@ -112,18 +224,17 @@ class RadioDJ:
        title: str = "Ведущий в эфире",
        voice: Optional[str] = None
    ) -> bool:
-        """Synthesize speech using edge-tts or rock-solid gTTS fallback and write ID3 tags for Navidrome."""
-        selected_voice = voice or self.default_voice
+        """Synthesize speech using high-fidelity Piper TTS or fallback to gTTS, and write ID3 tags for Navidrome."""
+        selected_voice_id = voice or self.default_voice
+        voice_info = VOICE_CATALOG.get(selected_voice_id, VOICE_CATALOG.get(self.default_voice))
        
-        # 1. Attempt edge-tts
-        try:
-            tts = edge_tts.Communicate(text, selected_voice)
-            await tts.save(output_file)
-            if os.path.exists(output_file) and os.path.getsize(output_file) > 500:
+        # 1. Attempt Piper TTS if configured
+        if voice_info and voice_info["engine"] == "piper" and voice_info["model_file"]:
+            success = await self._synthesize_piper(text, voice_info["model_file"], output_file)
+            if success:
                self._apply_id3_tags(output_file, title)
                return True
-        except Exception as e:
-            print(f"[RadioDJ] edge-tts error ({e}), trying gTTS fallback...")
+            print(f"[RadioDJ] Piper synthesis failed for {selected_voice_id}, falling back to gTTS...")

        # 2. Resilient fallback to gTTS
        try:
@@ -146,11 +146,8 @@ class RadioManager:
            "playlist_ready": playlist_ready,
            "speech_clips_count": len(speech_files),
            "speech_clips": speech_files,
-            "default_voice": settings.RADIO_VOICE,
-            "available_voices": [
-                {"id": "ru-RU-DmitryNeural", "name": "Дмитрий (Харизматичный DJ)", "gender": "Мужской"},
-                {"id": "ru-RU-SvetlanaNeural", "name": "Светлана (Мягкий вечерний голос)", "gender": "Женский"}
-            ]
+            "default_voice": settings.RADIO_VOICE or "ru_RU-dmitri-medium",
+            "available_voices": dj_host.get_available_voices()
        }

 radio_manager = RadioManager()
@@ -501,6 +501,7 @@ document.addEventListener("DOMContentLoaded", () => {
  // --- AI Radio Host Controller (Phase 3) ---
  const radioGenerateBtn = document.getElementById("radioGenerateBtn");
  const radioVoiceSelect = document.getElementById("radioVoiceSelect");
+  const radioVoicePreviewBtn = document.getElementById("radioVoicePreviewBtn");
  const radioStatusLabel = document.getElementById("radioStatusLabel");
  const radioClipsCountLabel = document.getElementById("radioClipsCountLabel");
  const radioSegmentsContainer = document.getElementById("radioSegmentsContainer");
@@ -508,10 +509,29 @@ document.addEventListener("DOMContentLoaded", () => {

  fetchRadioStatus();

+  // Voice sample preview playback
+  if (radioVoicePreviewBtn) {
+    radioVoicePreviewBtn.addEventListener("click", () => {
+      const selectedVoice = radioVoiceSelect ? radioVoiceSelect.value : "ru_RU-dmitri-medium";
+      const audioUrl = `api/radio/preview-voice?voice=${encodeURIComponent(selectedVoice)}`;
+      const selectedOption = (radioVoiceSelect && radioVoiceSelect.options[radioVoiceSelect.selectedIndex]) 
+        ? radioVoiceSelect.options[radioVoiceSelect.selectedIndex].text 
+        : "Голос ведущего";
+      playPreviewTrack(
+        audioUrl,
+        `Сэмпл: ${selectedOption}`,
+        "AI Радио (Предпрослушивание)",
+        "",
+        "",
+        audioUrl
+      );
+    });
+  }
+
  if (radioGenerateBtn) {
    radioGenerateBtn.addEventListener("click", async () => {
      radioGenerateBtn.disabled = true;
-      const voice = radioVoiceSelect ? radioVoiceSelect.value : "ru-RU-DmitryNeural";
+      const voice = radioVoiceSelect ? radioVoiceSelect.value : "ru_RU-dmitri-medium";
      if (radioStatusLabel) {
        radioStatusLabel.textContent = "🎙️ Ведущий пишет реплики и сводит эфир...";
        radioStatusLabel.className = "badge badge-llm";
@@ -560,6 +580,17 @@ document.addEventListener("DOMContentLoaded", () => {
      if (radioClipsCountLabel) {
        radioClipsCountLabel.textContent = `${data.speech_clips_count || 0} записей`;
      }
+
+      // Populate voice selector dynamically
+      if (radioVoiceSelect && data.available_voices && data.available_voices.length > 0) {
+        const currentVal = radioVoiceSelect.value || data.default_voice;
+        radioVoiceSelect.innerHTML = data.available_voices.map(v => `
+          <option value="${escapeHtml(v.id)}" ${v.id === currentVal ? "selected" : ""}>
+            🎙️ ${escapeHtml(v.name)}
+          </option>
+        `).join("");
+      }
+
      if (radioStatusLabel) {
        if (data.playlist_ready && data.speech_clips_count > 0) {
          radioStatusLabel.textContent = "🟢 Эфир в ротации";
@@ -147,7 +147,7 @@
      <section class="card radio-card">
        <div class="radio-header">
          <div class="radio-header-title">
-            <h2>🎙️ AI Радио <span class="badge badge-radio">AI DJ Эфир</span> <span class="badge badge-llm">🤖 Qwen 2.5 + Edge-TTS</span></h2>
+            <h2>🎙️ AI Радио <span class="badge badge-radio">AI DJ Эфир</span> <span class="badge badge-llm">🤖 Qwen 2.5 + Piper Neural</span></h2>
            <p class="radio-subtitle">
              Персональный радиоэфир: AI-ведущий объявляет погоду, время суток и делает динамичные подводки перед песнями в отдельном плейлисте «AI Радио».
            </p>
@@ -159,12 +159,18 @@

        <!-- Radio Host Voice & Context Row -->
        <div class="radio-settings-row">
-          <div class="radio-voice-group">
+          <div class="radio-voice-group" style="flex-wrap: wrap;">
            <label for="radioVoiceSelect" class="radio-voice-label">Голос ведущего:</label>
            <select id="radioVoiceSelect" class="radio-voice-select">
-              <option value="ru-RU-DmitryNeural" selected>🎙️ Дмитрий (Харизматичный)</option>
-              <option value="ru-RU-SvetlanaNeural">🎙️ Светлана (Мягкий)</option>
+              <option value="ru_RU-dmitri-medium" selected>🎙️ Дмитрий (Энергичный DJ)</option>
+              <option value="ru_RU-irina-medium">🎙️ Ирина (Естественный женский)</option>
+              <option value="ru_RU-denis-medium">🎙️ Денис (Глубокий баритон)</option>
+              <option value="ru_RU-ruslan-medium">🎙️ Руслан (Спокойный ведущий)</option>
+              <option value="gtts_ru">🎙️ Google (Стандартный робот)</option>
            </select>
+            <button id="radioVoicePreviewBtn" type="button" class="btn-sm btn-play-preview" title="Послушать сэмпл выбранного голоса">
+              ▶ Прослушать голос
+            </button>
          </div>
          <div class="radio-status-badge-wrap">
            <span id="radioStatusLabel" class="badge badge-context">Эфир готов к запуску</span>