Lock MVP voice (Michael, native settings); native /tts route

2026-10-17 16:44:28 -05:00
parent 444cee7f76
commit ca5592b964
4 changed files with 136 additions and 6 deletions
@@ -0,0 +1,103 @@
+"""Voice-tuning clips: one paragraph, many voice variants, so a human can pick by ear.
+Standard library only. Usage:
+  python3 tune_clips.py --text-file <chapter-01.txt> --paragraph 3 --out <dir> [--chatterbox URL] [--kokoro URL]
+Writes <out>/NN-name.wav (and .mp3 if ffmpeg exists) plus <out>/clips.md describing each clip."""
+import argparse
+import json
+import os
+import shutil
+import subprocess
+import time
+import urllib.request
+
+SEED = 7  # same seed everywhere so differences come from the settings, not luck
+
+# Chatterbox: stock male voices, then setting changes on one voice. exaggeration = expressiveness (lower = calmer),
+# cfg_weight = pacing (lower = slower, more deliberate), temperature = variation (lower = steadier).
+CHATTERBOX = [
+    ("henry-default", "Henry", dict(exaggeration=0.5, cfg_weight=0.5, temperature=0.8), "Henry, server defaults."),
+    ("henry-calm", "Henry", dict(exaggeration=0.3, cfg_weight=0.3, temperature=0.6), "Henry, calmer and slower: low expressiveness, low pace weight, steadier."),
+    ("gabriel-calm", "Gabriel", dict(exaggeration=0.3, cfg_weight=0.3, temperature=0.6), "Gabriel, same calm settings as Henry calm."),
+    ("adrian-calm", "Adrian", dict(exaggeration=0.3, cfg_weight=0.3, temperature=0.6), "Adrian, same calm settings."),
+    ("everett-calm", "Everett", dict(exaggeration=0.3, cfg_weight=0.3, temperature=0.6), "Everett, same calm settings."),
+    ("henry-warmer", "Henry", dict(exaggeration=0.4, cfg_weight=0.4, temperature=0.7), "Henry, between default and calm: a bit more life."),
+]
+# Kokoro: male voices. speed < 1 is slower.
+KOKORO = [
+    ("kokoro-adam", "am_adam", 0.92, "Kokoro am_adam at 0.92 speed."),
+    ("kokoro-michael", "am_michael", 0.92, "Kokoro am_michael at 0.92 speed."),
+    ("kokoro-george", "bm_george", 0.92, "Kokoro bm_george (British) at 0.92 speed."),
+]
+
+
+def post(url, body, timeout=600):
+    req = urllib.request.Request(url, json.dumps(body).encode(), {"Content-Type": "application/json"})
+    with urllib.request.urlopen(req, timeout=timeout) as r:
+        return r.read()
+
+
+def get_json(url):
+    with urllib.request.urlopen(url, timeout=30) as r:
+        return json.loads(r.read())
+
+
+def chatterbox_clip(base, text, voice, p):
+    body = dict(text=text, voice_mode="predefined", predefined_voice_id=voice + ".wav" if not voice.endswith(".wav") else voice,
+                output_format="wav", split_text=True, chunk_size=240, seed=SEED, **p)
+    return post(base.rstrip("/") + "/tts", body)
+
+
+def kokoro_clip(base, text, voice, speed):
+    return post(base.rstrip("/") + "/v1/audio/speech",
+                dict(model="kokoro", input=text, voice=voice, speed=speed, response_format="wav"))
+
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--text-file", required=True)
+    ap.add_argument("--paragraph", type=int, default=3, help="1-based paragraph number")
+    ap.add_argument("--out", required=True)
+    ap.add_argument("--chatterbox", default="http://127.0.0.1:8090")
+    ap.add_argument("--kokoro", default="http://127.0.0.1:8091")
+    a = ap.parse_args()
+    paras = [p.strip() for p in open(a.text_file, encoding="utf-8").read().split("\n\n") if p.strip()]
+    text = paras[a.paragraph - 1]
+    os.makedirs(a.out, exist_ok=True)
+    rows = ["# Voice-tuning clips", "", "Text: Clark, Chapter I, paragraph %d (%d words). Same text in every clip." % (a.paragraph, len(text.split())), "", "> " + text, ""]
+    ff = shutil.which("ffmpeg")
+    n = 0
+    jobs = [(nm, "Chatterbox", lambda v=v, p=p: chatterbox_clip(a.chatterbox, text, v, p), d) for nm, v, p, d in CHATTERBOX]
+    try:
+        have = {x if isinstance(x, str) else x.get("id", x.get("name")) for x in get_json(a.kokoro.rstrip("/") + "/v1/audio/voices").get("voices", [])}
+    except Exception as e:
+        print("kokoro voice list failed:", e)
+        have = None
+    for nm, v, sp, d in KOKORO:
+        if have is not None and v not in have:
+            print("skip", v, "(not on server)")
+            continue
+        jobs.append((nm, "Kokoro", lambda v=v, sp=sp: kokoro_clip(a.kokoro, text, v, sp), d))
+    for nm, eng, fn, d in jobs:
+        n += 1
+        base = os.path.join(a.out, "%02d-%s" % (n, nm))
+        t0 = time.time()
+        try:
+            data = fn()
+        except Exception as e:
+            print("FAILED", nm, e)
+            rows.append("- %02d-%s: FAILED (%s)" % (n, nm, e))
+            continue
+        with open(base + ".wav", "wb") as f:
+            f.write(data)
+        final = base + ".wav"
+        if ff:
+            subprocess.run([ff, "-y", "-loglevel", "error", "-i", base + ".wav", "-ac", "1", "-ar", "44100", "-b:a", "192k", base + ".mp3"], check=True)
+            final = base + ".mp3"
+        print("ok %s %.1fs render, %s" % (nm, time.time() - t0, final))
+        rows.append("- **%02d-%s** (%s): %s" % (n, nm, eng, d))
+    with open(os.path.join(a.out, "clips.md"), "w", encoding="utf-8") as f:
+        f.write("\n".join(rows) + "\n")
+
+
+if __name__ == "__main__":
+    main()