Lucy-TTS/F5: Skripte + Batches versionieren, schwere Assets ignoriert

- pocket_server.py (Produktions-TTS mit Stimmen-Waechter), text_norm, Bench-/Diag-Skripte - lucy-f5: f5_server/f5_test/bench_dml (DirectML-Experiment, Phase C/D offen) - .gitignore: venvs/Modelle/Audio/Logs der beiden Ordner + box_recon/gemma_swap-Scratch Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-02 10:29:33 +02:00
parent aff0105700
commit 09a1c98514
49 changed files with 8231 additions and 0 deletions
@@ -0,0 +1,59 @@
+# -*- coding: utf-8 -*-
+"""Test der ROBUSTEN Stream-Kopf-Logik: fester 1.8s-Kopf + Crop am ersten SUBSTANZIELLEN Wort-Segment
+(ignoriert Mini-Blips) + FIXER Gain. Ziel: Start immer echtes Wort, out_rms stabil ~0.09, kein Clip."""
+import os, numpy as np, librosa, soundfile as sf
+from pocket_tts import TTSModel
+from faster_whisper import WhisperModel
+
+BASE = r"F:\Coding Stuff\mission-control-2\client\lucy-tts"
+src,_ = librosa.load(os.path.join(BASE,"ref.mp3"), sr=24000, mono=True)
+srt,_ = librosa.effects.trim(src, top_db=30); ref=os.path.join(BASE,"ref.wav")
+sf.write(ref, srt[:int(18*24000)], 24000)
+TARGET_RMS=0.09; LEAD="Tja. "
+BASE_GAIN = TARGET_RMS / 0.18         # ~0.5 (raw full-rms ~0.18 konsistent)
+GMIN, GMAX = 0.7*BASE_GAIN, 1.4*BASE_GAIN  # Gain-Clamp -> kann NIE explodieren
+HEAD_FIXED = int(2.0*24000)
+m = TTSModel.load_model(language="german_24l", lsd_decode_steps=6, temp=0.9)
+vs = m.get_state_for_audio_prompt(ref); sr=m.sample_rate
+w = WhisperModel("small", device="cpu", compute_type="int8")
+TEXTS = ["Hallo Commander, ich höre dich.",
+         "Natürlich, Commander. Das Backup ist sauber durchgelaufen."]
+
+def crop_and_gain(a):
+    """Crop am ersten SUBSTANZIELLEN Wort-Segment nach dem Lead (>=0.2s, ignoriert Mini-Blips);
+    Gain aus voiced-rms, GECLAMPT (kann nie explodieren)."""
+    iv = librosa.effects.split(a, top_db=35)
+    voiced = np.concatenate([a[s:e] for s,e in iv]) if len(iv) else a
+    rms = float(np.sqrt(np.mean(voiced**2))) or 1e-9
+    gain = float(np.clip(TARGET_RMS/rms, GMIN, GMAX))
+    cut = 0
+    if len(iv) >= 2:
+        for k in range(1, len(iv)):
+            if (iv[k][1]-iv[k][0]) >= int(0.20*sr):
+                cut = max(iv[k-1][1], iv[k][0]-int(0.06*sr)); break
+    return a[cut:], round(cut/sr,2), round(gain,2)
+
+def run_once(text):
+    head=[]; hl=0; head_done=False; head_arr=None; rest=[]
+    for c in m.generate_audio_stream(vs, LEAD+text, frames_after_eos=4):
+        c=c.numpy() if hasattr(c,"numpy") else np.asarray(c); c=np.asarray(c,dtype=np.float32).reshape(-1)
+        if not head_done:
+            head.append(c); hl+=c.size
+            if hl>=HEAD_FIXED: head_arr=np.concatenate(head); head_done=True
+        else: rest.append(c)
+    if head_arr is None: head_arr=np.concatenate(head)
+    cropped_head, cutpt, gain = crop_and_gain(head_arr)
+    full = np.concatenate([cropped_head]+rest)
+    out=np.clip(full*gain,-0.95,0.95)
+    sf.write(os.path.join(BASE,"out_diag_s3.wav"), out, sr)
+    seg,_=w.transcribe(os.path.join(BASE,"out_diag_s3.wav"), language="de", beam_size=5)
+    start=(" ".join(x.text for x in seg)).strip()[:32]
+    return dict(cut=cutpt, gain=gain, out_rms=round(float(np.sqrt(np.mean(out**2))),3),
+                out_peak=round(float(np.abs(out).max()),3), start=start)
+
+print(f"gain-clamp=[{round(GMIN,2)},{round(GMAX,2)}] head={HEAD_FIXED/24000}s")
+for text in TEXTS:
+    print(f"=== {text[:28]!r} ===")
+    for i in range(5):
+        print(f"  run{i}:", run_once(text))
+print("DIAG3_DONE")