storyshot 0.4.2 → 0.6.0

package/src/zh_g2p.py

@@ -0,0 +1,118 @@
+# Mandarin text -> Kokoro v1.1-zh phonemes, for storyshot's Chinese voice (src/kokoro-zh.mjs).
+# Runs in storyshot's own venv (~/.cache/storyshot/zh-venv), once per render for every line.
+#   stdin:  {"lines": ["...", ...], "target": 300, "max": 510}
+#   stdout: {"pieces": [["<phonemes>", ...], ...]}   one list per line; long lines come back in pieces
+# G2P: misaki ZHG2P(version='1.1'); English words and single letters go through misaki's espeak-ng
+# fallback (letter names: "m x" -> ˈɛm ˈɛks), instead of being dropped as the stock frontend does.
+import json
+import logging
+import os
+import re
+import shutil
+import sys
+
+
+def espeak_data_path():
+    # espeak-ng truncates its data path at 160 characters; use a copy at a shorter path if needed
+    import espeakng_loader
+    data = espeakng_loader.get_data_path()
+    if len(data) <= 120:
+        return data
+    import tempfile
+    from importlib.metadata import version
+    for base in (os.path.join(os.path.expanduser('~'), '.cache', 'storyshot'), tempfile.gettempdir()):
+        short = os.path.join(base, f"espeak-ng-data-{version('espeakng-loader')}")
+        if len(short) > 120:
+            continue
+        if not os.path.exists(os.path.join(short, 'phontab')):
+            shutil.rmtree(short, ignore_errors=True)
+            shutil.copytree(data, short)
+        return short
+    return data
+
+
+def make_g2p():
+    import jieba
+    jieba.setLogLevel(logging.ERROR)
+    from phonemizer.backend.espeak.wrapper import EspeakWrapper
+    import misaki.espeak  # noqa: F401  (points phonemizer at the bundled espeak-ng library)
+    EspeakWrapper.set_data_path(espeak_data_path())
+    from misaki.espeak import EspeakFallback
+    from misaki import zh
+
+    esp = EspeakFallback(british=False)
+
+    class Token:
+        def __init__(self, text):
+            self.text = text
+
+    def en_callable(run):  # the v1.1 frontend hands over Latin runs like "m x" or "GeoGebra"
+        return ' '.join(esp(Token(w))[0] or '' for w in re.findall(r"[A-Za-z']+", run))
+
+    g2p = zh.ZHG2P(version='1.1', en_callable=en_callable)
+    return lambda text: (g2p(text)[0] or '').strip()
+
+
+SENTENCE = re.compile(r'(?<=[。！？；!?;])')
+CLAUSE = re.compile(r'(?<=[，、：,:])')
+
+
+def units(g2p, text, target, hard):
+    """Split text into (text, phoneme length) units no longer than `target`: sentences first, then clauses,
+    then (rarely) plain character slices."""
+    out = []
+    for sentence in [s for s in SENTENCE.split(text) if s.strip()]:
+        n = len(g2p(sentence))
+        if n <= target:
+            out.append((sentence, n))
+            continue
+        for clause in [c for c in CLAUSE.split(sentence) if c.strip()]:
+            m = len(g2p(clause))
+            if m <= hard:
+                out.append((clause, m))
+                continue
+            k = -(-m // target)
+            step = -(-len(clause) // k)
+            for i in range(0, len(clause), step):
+                part = clause[i:i + step]
+                out.append((part, len(g2p(part))))
+    return out
+
+
+def pieces(g2p, text, target, hard):
+    text = text.strip()
+    if not text:
+        return []
+    ps = g2p(text)
+    if len(ps) <= target:
+        return [ps] if ps else []
+    us = units(g2p, text, target, hard)
+    total = sum(n for _, n in us)
+    k = -(-total // target)
+    goal = total / k
+    groups, cur, n = [], [], 0
+    for t, m in us:
+        # close the group when the next unit would overflow, or would land further from an even split
+        if cur and (n + m > target or abs(n + m - goal) > abs(n - goal)):
+            groups.append(''.join(cur))
+            cur, n = [], 0
+        cur.append(t)
+        n += m
+    if cur:
+        groups.append(''.join(cur))
+    return [p[:hard] for p in (g2p(g) for g in groups) if p]
+
+
+def main():
+    req = json.loads(sys.stdin.buffer.read().decode('utf-8'))
+    out = sys.stdout
+    sys.stdout = sys.stderr  # keep any library chatter off the JSON channel
+    g2p = make_g2p()
+    target, hard = int(req.get('target', 300)), int(req.get('max', 510))
+    result = [pieces(g2p, line, target, hard) for line in req['lines']]
+    out.buffer.write(json.dumps({'pieces': result}, ensure_ascii=False).encode('utf-8'))
+    out.flush()
+
+
+if __name__ == '__main__':
+    main()