feat(f5tts): HF-URL Support fuer Custom Checkpoints (aihpi/F5-TTS-German)

_resolve_hf_path wandelt hf://user/repo/path → lokaler Download via
huggingface_hub.hf_hub_download. So kann man in Diagnostic einfach die
HF-Pfade fuer custom Modelle reinschreiben, ohne erst manuell zu
downloaden + zu mounten.

Format: hf://aihpi/F5-TTS-German/F5TTS_Base/model_365000.safetensors
        hf://aihpi/F5-TTS-German/vocab.txt

Diagnostic UI: Placeholders + Labels angepasst mit Beispiel-HF-Pfaden
und Hinweis dass fuer Fine-Tunes "F5TTS_Base" statt "F5TTS_v1_Base"
als Architektur-Name gesetzt werden muss.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

diagnostic/index.html

@@ -469,23 +469,25 @@
               Hardcoded Defaults: F5TTS_v1_Base, cfg_strength=2.5, nfe_step=32.
             </div>
 
-            <label style="color:#8888AA;font-size:12px;">Modell-ID:</label>
+            <label style="color:#8888AA;font-size:12px;">
+              Modell-Architektur (F5TTS_v1_Base = Default multilingual, F5TTS_Base = fuer die meisten Fine-Tunes):
+            </label>
             <input type="text" id="diag-f5tts-model"
               placeholder="F5TTS_v1_Base"
               style="background:#1E1E2E;color:#fff;border:1px solid #2A2A3E;border-radius:6px;padding:6px 10px;font-size:13px;">
 
             <label style="color:#8888AA;font-size:12px;">
-              Custom Checkpoint (HF-Repo "user/repo" oder Container-Pfad, leer = Default):
+              Custom Checkpoint — HF-Pfad (hf://user/repo/file) oder lokaler Container-Pfad. Leer = Default.
             </label>
             <input type="text" id="diag-f5tts-ckpt"
-              placeholder="z.B. aoxo/F5-TTS-German"
+              placeholder="z.B. hf://aihpi/F5-TTS-German/F5TTS_Base/model_365000.safetensors"
               style="background:#1E1E2E;color:#fff;border:1px solid #2A2A3E;border-radius:6px;padding:6px 10px;font-size:13px;">
 
             <label style="color:#8888AA;font-size:12px;">
-              Custom Vocab (passend zum Checkpoint, optional):
+              Custom Vocab — muss zum Checkpoint passen. Leer = Default.
             </label>
             <input type="text" id="diag-f5tts-vocab"
-              placeholder="leer = Default"
+              placeholder="z.B. hf://aihpi/F5-TTS-German/vocab.txt"
               style="background:#1E1E2E;color:#fff;border:1px solid #2A2A3E;border-radius:6px;padding:6px 10px;font-size:13px;">
 
             <div style="display:flex;gap:12px;">

xtts/f5tts/bridge.py

@@ -99,6 +99,33 @@ def _get_f5tts_cls():
     return _F5TTS_cls
 
 
+def _resolve_hf_path(p: str) -> str:
+    """Wenn p mit 'hf://' anfaengt → aus HuggingFace Hub runterladen,
+    lokalen Pfad zurueckgeben. Sonst unveraendert.
+
+    Format: hf://user/repo/path/to/file.ext
+    Beispiel: hf://aihpi/F5-TTS-German/F5TTS_Base/model_365000.safetensors
+    """
+    if not p or not p.startswith("hf://"):
+        return p
+    try:
+        from huggingface_hub import hf_hub_download
+        rest = p[5:]
+        parts = rest.split("/", 2)
+        if len(parts) < 3:
+            logger.warning("Ungueltiges hf:// Format: %s (erwarte hf://user/repo/path)", p)
+            return p
+        repo_id = f"{parts[0]}/{parts[1]}"
+        filename = parts[2]
+        logger.info("HF-Download: %s aus %s", filename, repo_id)
+        local = hf_hub_download(repo_id=repo_id, filename=filename)
+        logger.info("HF-Download fertig: %s", local)
+        return local
+    except Exception as e:
+        logger.exception("HF-Download fehlgeschlagen fuer %s: %s", p, e)
+        return p
+
+
 class F5Runner:
     """Haelt das F5-TTS-Modell. Synthese laeuft im Executor (blocking).
 
@@ -122,14 +149,16 @@ class F5Runner:
 
     def _load_blocking(self) -> None:
         cls = _get_f5tts_cls()
+        ckpt_resolved = _resolve_hf_path(self.ckpt_file) if self.ckpt_file else ""
+        vocab_resolved = _resolve_hf_path(self.vocab_file) if self.vocab_file else ""
         logger.info("Lade F5-TTS '%s' (device=%s, ckpt=%s)...",
-                    self.model_id, F5TTS_DEVICE, self.ckpt_file or "default")
+                    self.model_id, F5TTS_DEVICE, ckpt_resolved or "default")
         self._load_started_at = time.time()
         kwargs = {"model": self.model_id, "device": F5TTS_DEVICE}
-        if self.ckpt_file:
+        if ckpt_resolved:
-            kwargs["ckpt_file"] = self.ckpt_file
+            kwargs["ckpt_file"] = ckpt_resolved
-        if self.vocab_file:
+        if vocab_resolved:
-            kwargs["vocab_file"] = self.vocab_file
+            kwargs["vocab_file"] = vocab_resolved
         self.model = cls(**kwargs)
         elapsed = time.time() - self._load_started_at
         logger.info("F5-TTS geladen in %.1fs (cfg_strength=%.1f, nfe=%d)",
@@ -256,39 +285,48 @@ def voice_paths(name: str) -> tuple[Path, Path]:
 
 def normalize_ref_wav(src_wav: Path, max_seconds: float = REF_MAX_SECONDS) -> tuple[Path, bool]:
     """Bringt die Referenz-WAV in F5-TTS-freundliche Form:
-    24kHz mono + max max_seconds Dauer. Original wird ueberschrieben wenn
+
-    Aenderungen noetig waren.
+      * 24kHz mono
+      * max max_seconds Dauer
+      * Stille am Anfang + Ende abgeschnitten (silenceremove-Filter)
+      * Lautheit auf -16 LUFS normalisiert (loudnorm-Filter) damit
+        das Modell konsistente Amplituden sieht
+
+    F5-TTS reagiert empfindlich auf leise / verrauschte / zerhackte
+    Referenzen. Konsistente, saubere Input-Lautheit hilft der Quali.
 
     Returns:
         (path, was_modified) — was_modified=True wenn die Datei wirklich
         geaendert wurde (Caller sollte dann den passenden .txt invalidieren).
     """
-    try:
-        info = sf.info(str(src_wav))
-        # Schon gut? Sample-Rate, Kanaele und Dauer passen?
-        if (info.samplerate == TARGET_SR and info.channels == 1
-                and info.duration <= max_seconds + 0.1):
-            return src_wav, False
-    except Exception:
-        info = None
-
     tmp_out = src_wav.with_suffix(".conv.wav")
+    # silenceremove am Anfang: bis -50dB gesprochen wird
+    # silenceremove am Ende: ueber -50dB rein, dann 0.5s stille als Cutoff
+    # loudnorm: EBU R128, Ziel -16 LUFS
+    af = ("silenceremove=start_periods=1:start_duration=0.05:start_threshold=-50dB,"
+          "silenceremove=stop_periods=1:stop_duration=0.5:stop_threshold=-50dB,"
+          "loudnorm=I=-16:TP=-1.5:LRA=11")
     cmd = ["ffmpeg", "-y", "-i", str(src_wav),
+           "-af", af,
            "-ar", str(TARGET_SR), "-ac", "1",
            "-t", str(max_seconds),
            "-f", "wav", str(tmp_out)]
     r = subprocess.run(cmd, capture_output=True, timeout=30)
     if r.returncode != 0:
         logger.warning("ffmpeg-Normalisierung von %s fehlgeschlagen: %s",
-                       src_wav, r.stderr.decode(errors="replace")[:200])
+                       src_wav, r.stderr.decode(errors="replace")[:300])
         try:
             tmp_out.unlink()
         except OSError:
             pass
         return src_wav, False
     os.replace(tmp_out, src_wav)
-    logger.info("Referenz-WAV normalisiert: %s (24kHz mono, max %.1fs)",
+    try:
-                src_wav.name, max_seconds)
+        info = sf.info(str(src_wav))
+        logger.info("Referenz-WAV normalisiert: %s (%.1fs, %dHz mono, -16 LUFS, silence getrimmt)",
+                    src_wav.name, info.duration, info.samplerate)
+    except Exception:
+        logger.info("Referenz-WAV normalisiert: %s", src_wav.name)
     return src_wav, True