Hacker-Software
/
ARIA-AGENT
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases 71 Wiki Activity

fix(f5tts): cfg_strength hochgezogen damit Deutsch nicht ins Spanische rutscht 

F5TTS_v1_Base ist hauptsaechlich auf Englisch+Chinesisch trainiert; bei
Deutsch (oder anderen Romance/Germanic-Sprachen) schwimmt der Generator
ohne starkes Conditioning gerne in eine andere Sprache.

- cfg_strength 2.0 → 2.5 (per ENV F5TTS_CFG_STRENGTH ueberschreibbar)
- nfe_step bleibt 32 (per ENV ueberschreibbar)
- F5TTS_CKPT_FILE / F5TTS_VOCAB_FILE als ENV — damit man eine Community-
  German-Checkpoint einhaengen kann ohne Code-Aenderung

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
duffyduck 2026-04-24 15:30:08 +02:00
parent 22fa4b3ccf
commit c1a5518fb7
2 changed files with 27 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
xtts/docker-compose.yml
View File

@ -39,7 +39,11 @@ services:
- RVS_TLS_FALLBACK=${RVS_TLS_FALLBACK:-true}
- RVS_TOKEN=${RVS_TOKEN}
- F5TTS_MODEL=${F5TTS_MODEL:-F5TTS_v1_Base}
- F5TTS_CKPT_FILE=${F5TTS_CKPT_FILE:-}
- F5TTS_VOCAB_FILE=${F5TTS_VOCAB_FILE:-}
- F5TTS_DEVICE=${F5TTS_DEVICE:-cuda}
- F5TTS_CFG_STRENGTH=${F5TTS_CFG_STRENGTH:-2.5}
- F5TTS_NFE_STEP=${F5TTS_NFE_STEP:-32}
- VOICES_DIR=/voices
restart: unless-stopped

26
xtts/f5tts/bridge.py
View File

@ -53,7 +53,14 @@ RVS_TLS_FALLBACK = os.getenv("RVS_TLS_FALLBACK", "true").lower() == "true"
RVS_TOKEN = os.getenv("RVS_TOKEN", "").strip()
F5TTS_MODEL = os.getenv("F5TTS_MODEL", "F5TTS_v1_Base")
F5TTS_CKPT_FILE = os.getenv("F5TTS_CKPT_FILE", "") # optional: HF-Repo oder lokales .pt
F5TTS_VOCAB_FILE = os.getenv("F5TTS_VOCAB_FILE", "") # optional: zugehoerige vocab.txt
F5TTS_DEVICE = os.getenv("F5TTS_DEVICE", "cuda")
# cfg_strength: wie stark der Generator am Referenz-Voice klebt.
# Default F5-TTS = 2.0. Bei nicht-EN/CN Sprachen (Deutsch!) hilft >2.5
# damit das Modell nicht in eine andere Sprache abrutscht.
F5TTS_CFG_STRENGTH = float(os.getenv("F5TTS_CFG_STRENGTH", "2.5"))
F5TTS_NFE_STEP = int(os.getenv("F5TTS_NFE_STEP", "32"))
VOICES_DIR = Path(os.getenv("VOICES_DIR", "/voices"))
PCM_CHUNK_BYTES = 8192 # ~170ms @ 24kHz mono s16
@ -82,10 +89,17 @@ class F5Runner:
def _load_blocking(self) -> None:
cls = _get_f5tts_cls()
logger.info("Lade F5-TTS '%s' (device=%s)...", F5TTS_MODEL, F5TTS_DEVICE)
logger.info("Lade F5-TTS '%s' (device=%s, ckpt=%s)...",
F5TTS_MODEL, F5TTS_DEVICE, F5TTS_CKPT_FILE or "default")
t0 = time.time()
self.model = cls(model=F5TTS_MODEL, device=F5TTS_DEVICE)
logger.info("F5-TTS geladen in %.1fs", time.time() - t0)
kwargs = {"model": F5TTS_MODEL, "device": F5TTS_DEVICE}
if F5TTS_CKPT_FILE:
kwargs["ckpt_file"] = F5TTS_CKPT_FILE
if F5TTS_VOCAB_FILE:
kwargs["vocab_file"] = F5TTS_VOCAB_FILE
self.model = cls(**kwargs)
logger.info("F5-TTS geladen in %.1fs (cfg_strength=%.1f, nfe=%d)",
time.time() - t0, F5TTS_CFG_STRENGTH, F5TTS_NFE_STEP)
async def ensure_loaded(self) -> None:
async with self._lock:
@ -95,12 +109,18 @@ class F5Runner:
await loop.run_in_executor(None, self._load_blocking)
def _infer_blocking(self, gen_text: str, ref_wav: str, ref_text: str) -> tuple[np.ndarray, int]:
# cfg_strength + nfe_step erhoeht damit das Modell nicht in andere
# Sprachen abrutscht (Bug bei Deutsch: rutscht ohne Verstaerkung
# gerne ins Spanische ab, weil F5TTS_v1_Base hauptsaechlich auf EN+CN
# trainiert ist).
wav, sr, _ = self.model.infer(
ref_file=ref_wav,
ref_text=ref_text,
gen_text=gen_text,
remove_silence=True,
seed=-1,
cfg_strength=F5TTS_CFG_STRENGTH,
nfe_step=F5TTS_NFE_STEP,
)
# F5-TTS gibt float32 1D-Array — auf 24kHz sample-rate standard
if not isinstance(wav, np.ndarray):