Hitonabi
09a1c98514
- 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>
68 lines
2.9 KiB
Python
68 lines
2.9 KiB
Python
# -*- coding: utf-8 -*-
|
|
"""bench_b2.py — verifiziert B2 (Worker-Pool) end-to-end mit den ECHTEN pocket_server-Funktionen.
|
|
Misst die Wall-Clock einer langen, mehrsätzigen Antwort SERIELL (1 Modell) vs. 3 WORKER parallel,
|
|
prüft die Reihenfolge (geordnete Ausgabe) und legt beide Ergebnis-WAVs auf den Desktop.
|
|
"""
|
|
import os, time
|
|
import numpy as np, soundfile as sf, librosa
|
|
from concurrent.futures import ProcessPoolExecutor
|
|
from pocket_tts import TTSModel
|
|
import pocket_server as ps
|
|
|
|
DESK = os.path.join(os.path.expanduser("~"), "Desktop", "lucy_samples")
|
|
os.makedirs(DESK, exist_ok=True)
|
|
|
|
LONG = ("Guten Morgen, Commander. Das nächtliche Backup ist sauber durchgelaufen. "
|
|
"Es gab keine Fehler in den Protokollen. Der Dienst läuft stabil weiter. "
|
|
"Ich habe die Modelle vorgewärmt und die Engine antwortet zügig. "
|
|
"Wenn du möchtest, starte ich jetzt den Tagesbericht und fasse die offenen Punkte zusammen.")
|
|
|
|
|
|
def _serial():
|
|
m = TTSModel.load_model(language=ps.LANG, lsd_decode_steps=ps.LSD, temp=ps.TEMP,
|
|
noise_clamp=ps.NOISE_CLAMP, quantize=ps.QUANTIZE)
|
|
ref = ps._prep_ref()
|
|
try:
|
|
vs = m.get_state_for_audio_prompt(ps.VOICE_ST)
|
|
except Exception:
|
|
vs = m.get_state_for_audio_prompt(ref)
|
|
ref_audio, _ = librosa.load(ref, sr=m.sample_rate, mono=True)
|
|
ps.STATE.clear()
|
|
ps.STATE.update(m=m, vs=vs, sr=m.sample_rate, ref_fp=ps._fingerprint(ref_audio, m.sample_rate))
|
|
sents = ps._split_sentences(LONG)
|
|
t0 = time.time()
|
|
parts = list(ps._gen_sentences_ordered(sents)) # pool=None -> serieller Zweig (LOCK)
|
|
dt = time.time() - t0
|
|
a = np.concatenate(parts)
|
|
sf.write(os.path.join(DESK, "b2_serial.wav"), a, m.sample_rate)
|
|
ps.STATE.clear(); del m
|
|
return dt, a.size / 24000, len(sents)
|
|
|
|
|
|
def _parallel(nworkers):
|
|
ps.STATE.clear()
|
|
pool = ProcessPoolExecutor(max_workers=nworkers, initializer=ps._worker_init)
|
|
list(pool.map(ps._warmup, range(nworkers))) # alle Worker vorab hochfahren
|
|
ref = ps._prep_ref()
|
|
ref_audio, _ = librosa.load(ref, sr=24000, mono=True)
|
|
ps.STATE.update(sr=24000, ref_fp=ps._fingerprint(ref_audio, 24000), pool=pool)
|
|
sents = ps._split_sentences(LONG)
|
|
t0 = time.time()
|
|
parts = list(ps._gen_sentences_ordered(sents)) # pool -> parallel, geordnet
|
|
dt = time.time() - t0
|
|
a = np.concatenate(parts)
|
|
sf.write(os.path.join(DESK, f"b2_parallel_{nworkers}w.wav"), a, 24000)
|
|
pool.shutdown(wait=True); ps.STATE.clear()
|
|
return dt, a.size / 24000, len(sents)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
print(f"Text: {len(LONG)} Zeichen")
|
|
ds, audio_s, n = _serial()
|
|
print(f"SERIELL : {ds:5.2f}s wall ({n} Saetze, {audio_s:4.1f}s Audio) -> b2_serial.wav")
|
|
dp, _, _ = _parallel(3)
|
|
print(f"3 WORKER: {dp:5.2f}s wall -> b2_parallel_3w.wav")
|
|
if dp > 0:
|
|
print(f"Speedup : {ds/dp:.2f}x (Audio gleich lang -> nur Generierzeit zaehlt)")
|
|
print("B2_DONE")
|