# -*- coding: utf-8 -*-
"""Naturalness-Sweep gegen 'Roboter/abgehackt': temperature, lsd hoch, + 2 Referenz-Fenster.
Wörter sind laut Whisper alle da -> Problem ist Prosodie/Timbre, nicht Content."""
import os, time, numpy as np, soundfile as sf, librosa
from pocket_tts import TTSModel

BASE = r"F:\Coding Stuff\mission-control-2\client\lucy-tts"
OUT = r"C:\Users\TobisPC\Desktop\lucy_natural"; os.makedirs(OUT, exist_ok=True)

# Zwei Referenz-Fenster aus der EL-Saber-MP3
src, _ = librosa.load(os.path.join(BASE, "ref.mp3"), sr=24000, mono=True)
srt, _ = librosa.effects.trim(src, top_db=30)
ref18 = os.path.join(BASE, "ref.wav")          # aktuell: erste 18s
sf.write(ref18, srt[:int(18*24000)], 24000)
refC = os.path.join(BASE, "ref_C.wav")          # XTTS-Liebling: Sek 18..27 (9s)
segC = src[int(18*24000):int(27*24000)]
segCt, _ = librosa.effects.trim(segC, top_db=30)
sf.write(refC, segCt, 24000)

SENT = {
 "kurz":  "Hallo Commander, ich höre dich.",
 "mittel":"Guten Morgen, Commander. Das Backup ist sauber durchgelaufen und es gab keine Fehler.",
}

def cleanup_v2(a, sr):
    a = np.asarray(a, dtype=np.float32).reshape(-1)
    if a.size == 0: return a
    peak = float(np.max(np.abs(a)))
    if peak > 0: a = a * (0.95 / peak)
    yt, _ = librosa.effects.trim(a, top_db=45); a = yt if yt.size else a
    fi = min(int(0.01*sr), a.size//2)
    if fi > 0:
        a[:fi] *= np.linspace(0.,1.,fi,dtype=np.float32)
        a[-fi:] *= np.linspace(1.,0.,fi,dtype=np.float32)
    pad = np.zeros(int(0.09*sr), dtype=np.float32)
    return np.concatenate([pad, a, pad])

# (tag, lsd, temp, ref)
COMBOS = [
 ("A_lsd6_t070_ref18", 6, 0.70, ref18),
 ("B_lsd6_t090_ref18", 6, 0.90, ref18),
 ("C_lsd6_t105_ref18", 6, 1.05, ref18),
 ("D_lsd8_t090_ref18", 8, 0.90, ref18),
 ("E_lsd12_t090_ref18",12, 0.90, ref18),
 ("F_lsd6_t090_refC",  6, 0.90, refC),
]

for tag, lsd, temp, ref in COMBOS:
    t0 = time.time()
    m = TTSModel.load_model(language="german_24l", lsd_decode_steps=lsd, temp=temp)
    vs = m.get_state_for_audio_prompt(ref)
    sr = m.sample_rate
    print(f"== {tag} (load {time.time()-t0:.1f}s) ==", flush=True)
    for name, text in SENT.items():
        t0 = time.time()
        audio = m.generate_audio(vs, text, frames_after_eos=4); dt = time.time()-t0
        a = audio.numpy() if hasattr(audio,"numpy") else np.asarray(audio)
        a = cleanup_v2(a, sr); secs = a.size/sr
        sf.write(os.path.join(OUT, f"{tag}_{name}.wav"), a, sr)
        print(f"   [{name:6}] gen={dt:5.2f}s audio={secs:5.2f}s RTF={dt/max(secs,0.01):.2f}", flush=True)
print("NATURAL_DONE", flush=True)