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

ensvoice/refbench.py

@@ -0,0 +1,77 @@
+"""Choose the reference clip by numbers: 16 note paragraphs x 2 seeds in every candidate voice.
+
+usage: refbench [ref_id ...]      render (tts env); no ids: the configured reference + the 8 best by sim
+       refbench --report          table after `qc --jobs work/refbench/jobs.json --out work/refbench/qc.json`
+Pick the reference with low WER and few bad takes (a leaking reference shows as many takes with WER > 0.1:
+English at the start), then the highest MOS; put its id into voice.reference. Local backend only.
+"""
+import collections
+import json
+import os
+import random
+import re
+import statistics as st
+import sys
+
+from .config import load
+
+
+def report(cfg):
+    q = json.load(open(cfg.w("refbench", "qc.json"), encoding="utf-8"))
+    by = collections.defaultdict(list)
+    for k, r in q.items():
+        if "error" not in r:
+            by[k.split("|")[0]].append(r)
+    rows = []
+    for v, rs in by.items():
+        so = sum(r["stress_ok"] for r in rs) / max(1, sum(r["stress_n"] for r in rs))
+        rows.append((st.mean(r.get("mos", 0) for r in rs), v, len(rs), st.mean(r["wer"] for r in rs),
+                     sum(r["wer"] > .1 for r in rs), so, st.mean(r["sim"] for r in rs)))
+    print(f"{'MOS':>5} {'WER':>6} {'bad':>4} {'stress':>7} {'SIM':>6}  reference")
+    for mos, v, n, wer, bad, so, sim in sorted(rows, reverse=True):
+        print(f"{mos:5.2f} {wer:6.3f} {bad:4d} {so:7.3f} {sim:6.3f}  {v}")
+
+
+def main():
+    cfg = load()
+    if "--report" in sys.argv:
+        return report(cfg)
+    from .accent import acute_marks, plus_marks
+    from .lexicon import post, pre
+    from .render import reference
+    from .tts.local import LocalTTS
+    refs = json.load(open(cfg.w("refs.json"), encoding="utf-8"))
+    ids = [a for a in sys.argv[1:] if not a.startswith("-")] or \
+        list(dict.fromkeys([cfg["voice"]["reference"]] + [c["id"] for c in refs[:8]]))
+    notes = json.load(open(cfg.w("notes.json"), encoding="utf-8"))
+    tp = cfg.w("refbench", "texts.json")
+    if tp.exists():
+        sel = json.load(open(tp, encoding="utf-8"))
+    else:
+        random.seed(7)
+        paras = [(n["k"], p.strip()) for n in notes for p in re.split(r"\n\s*\n", pre(n["text"])) if 150 <= len(p.strip()) <= 380]
+        random.shuffle(paras)
+        seen, sel = set(), []
+        for k, p in paras:
+            if k not in seen and len(sel) < 16:
+                seen.add(k); sel.append(dict(id=k, text=p, plus=post(plus_marks(p))))
+        json.dump(sel, open(tp, "w", encoding="utf-8"), ensure_ascii=False, indent=1)
+    tts, jobs = None, []
+    for rid in ids:
+        wav, text = reference(cfg, rid)
+        tts = tts or LocalTTS(cfg, wav, text)
+        tts.set_voice(wav, text)
+        for seed in (1000, 1001):
+            items = [(acute_marks(s["plus"]), str(cfg.w("refbench", "takes", f"{rid}~{s['id']}~{seed}.wav")))
+                     for s in sel]
+            todo = [it for it in items if not os.path.exists(it[1])]
+            if todo:
+                tts.speak(todo, seed)
+            jobs += [dict(id=f"{rid}|{s['id']}|{seed}", text=s["text"], plus=s["plus"], wav=it[1]) for s, it in zip(sel, items)]
+    json.dump(jobs, open(cfg.w("refbench", "jobs.json"), "w", encoding="utf-8"), ensure_ascii=False)
+    print("REFBENCH_RENDERED", len(jobs), "takes; next: python run.py qc --jobs", cfg.w("refbench", "jobs.json"),
+          "--out", cfg.w("refbench", "qc.json"))
+
+
+if __name__ == "__main__":
+    main()