Russian narration for the Enshrouded notes: mod 0.4.0, installer, generation pipeline

ensvoice/assemble.py

@@ -0,0 +1,105 @@
+"""Step 7: join every note from its picked takes and encode it for the game (work/qoa, work/wav48).
+
+- each take trimmed to its speech (top_db 40), 60 ms kept at both ends, 8 ms fades;
+- pauses as the EN narrator makes them: PAUSE["s"] inside a paragraph, PAUSE["l"] after a line break,
+  PAUSE["p"] between paragraphs (EN medians: phrases ~0.6 s, paragraphs ~0.9 s);
+- 24 -> 48 kHz with soxr VHQ; loudness: active-speech RMS matched to the EN original, soft limiter
+  at -1 dBFS; 100 ms of silence at both ends; frameless QOA.
+The QOA encoder is pure Python: this step is CPU-bound (assemble.workers processes).
+"""
+import json
+import os
+import sys
+from multiprocessing import Pool
+
+import librosa
+import numpy as np
+import soundfile as sf
+
+from .config import load
+from .qoa import encode
+
+PAUSE = {"s": 0.45, "l": 0.70, "p": 0.90}
+CFG = None
+
+
+def active_rms(x):
+    v = x[::5]
+    v = v[np.abs(v) > 200 / 32768]
+    return float(np.sqrt(np.mean(v * v))) if len(v) else 1e-3
+
+
+def trim(y, sr):
+    iv = librosa.effects.split(y, top_db=40, frame_length=1024, hop_length=240)
+    if not len(iv):
+        return y
+    y = y[max(0, iv[0][0] - int(0.06 * sr)): min(len(y), iv[-1][1] + int(0.06 * sr))].copy()
+    f = int(0.008 * sr)
+    y[:f] *= np.linspace(0, 1, f)
+    y[-f:] *= np.linspace(1, 0, f)
+    return y
+
+
+def job(args):
+    n, segs, pick = args
+    cfg = CFG or load()
+    meta = cfg.w("qoa", f"{n['k']}.json")
+    if meta.exists():
+        return
+    parts = []
+    for s in segs:
+        y, sr = sf.read(cfg.w("takes", f"{s['take']}_s{pick[s['id']]}.wav"), dtype="float32")
+        if parts:
+            parts.append(np.zeros(int(PAUSE.get(s["join"], 0.45) * sr), np.float32))
+        parts.append(trim(y, sr))
+    y = librosa.resample(np.concatenate(parts), orig_sr=24000, target_sr=48000, res_type="soxr_vhq")
+    en, _ = sf.read(cfg.w("en24", f"{n['k']}.wav"), dtype="float32")
+    gain = active_rms(en) / active_rms(y)
+    y = y * gain
+    lim = 10 ** (-1 / 20)
+    knee = 0.7 * lim
+    over = np.abs(y) > knee
+    y[over] = np.sign(y[over]) * (knee + (lim - knee) * np.tanh((np.abs(y[over]) - knee) / (lim - knee)))
+    pad = np.zeros(4800, np.float32)
+    y = np.concatenate([pad, y, pad])
+    sf.write(cfg.w("wav48", f"{n['k']}.wav"), y, 48000, subtype="PCM_16")
+    pcm = np.clip(np.round(y * 32767), -32768, 32767).astype(np.int32).tolist()
+    cfg.w("qoa", f"{n['sound']}.qoa").write_bytes(encode(pcm))
+    json.dump(dict(k=n["k"], name=n["name"], vo=n["vo"], sound=n["sound"], frames=len(pcm),
+                   seconds=len(pcm) / 48000, gain=round(gain, 3), segments=len(segs)), open(meta, "w"))
+
+
+def main():
+    cfg = load()
+    notes = json.load(open(cfg.w("notes.json"), encoding="utf-8"))
+    segs = json.load(open(cfg.w("segs.json"), encoding="utf-8"))
+    pick = json.load(open(cfg.w("pick.json")))
+    by = {}
+    for s in segs:
+        by.setdefault(s["k"], []).append(s)
+    notes = [n for n in notes if n["k"] in by]  # segments may cover a subset (ENSVOICE_NOTES)
+    if "--force" in sys.argv:  # re-encode every note (after a new pick)
+        for f in cfg.w("qoa", "x").parent.glob("n*.json"):
+            f.unlink()
+    else:  # re-encode the notes whose picked takes changed
+        for n in notes:
+            m = cfg.w("qoa", f"{n['k']}.json")
+            if m.exists() and json.load(open(m)).get("takes") != [f"{s['take']}_s{pick.get(s['id'])}" for s in by[n["k"]]]:
+                m.unlink()
+    jobs = [(n, by[n["k"]], pick) for n in notes if not cfg.w("qoa", f"{n['k']}.json").exists()]
+    print("notes to encode", len(jobs), flush=True)
+    with Pool(int(cfg.get("assemble", "workers", max(1, (os.cpu_count() or 2) - 1)))) as p:
+        for i, _ in enumerate(p.imap_unordered(job, jobs, chunksize=1)):
+            if i % 25 == 0:
+                print(f"{i + 1}/{len(jobs)}", flush=True)
+    for n in notes:  # remember which takes each note was built from
+        m = cfg.w("qoa", f"{n['k']}.json")
+        d = json.load(open(m))
+        d["takes"] = [f"{s['take']}_s{pick[s['id']]}" for s in by[n["k"]]]
+        json.dump(d, open(m, "w"))
+    total = sum(json.load(open(cfg.w("qoa", f"{n['k']}.json")))["seconds"] for n in notes)
+    print("ASSEMBLED", len(notes), "notes,", round(total / 60, 1), "minutes")
+
+
+if __name__ == "__main__":
+    main()