fix(cli): split Kokoro text on token overflow

packages/cli/src/tts/synthesize.test.ts

@@ -1,3 +1,5 @@
+import { readFileSync } from "node:fs";
+import { fileURLToPath } from "node:url";
 import { beforeEach, describe, expect, it, vi } from "vitest";
 import type { SupportedLang } from "./manager.js";
 
@@ -121,3 +123,14 @@ describe("synthesize — espeak-ng language code translation", () => {
     expect(getCapturedArgv()?.[LANG_ARGV_INDEX]).toBe("es");
   });
 });
+
+describe("synthesize — Kokoro voice token overflow", () => {
+  it("adaptively splits only the 510-entry voice overflow and concatenates the audio", () => {
+    const source = readFileSync(fileURLToPath(new URL("./synthesize.ts", import.meta.url)), "utf8");
+
+    expect(source).toContain("def is_voice_token_overflow(error):");
+    expect(source).toContain("def synthesize_text(value):");
+    expect(source).toContain("samples.extend(right_samples)");
+    expect(source).toContain('const SCRIPT_PATH = join(SCRIPT_DIR, "synth-v3.py")');
+  });
+});

packages/cli/src/tts/synthesize.ts

@@ -20,7 +20,7 @@ import { findPython, hasPythonPackage } from "./python.js";
 // We pass it conditionally so older installs that only accept `voice=`/`speed=`
 // continue to work (falling back to Kokoro's default phonemization).
 const SYNTH_SCRIPT = `
-import sys, json, inspect
+import sys, json, inspect, re
 
 model_path = sys.argv[1]
 voices_path = sys.argv[2]
@@ -40,7 +40,39 @@ supports_lang = "lang" in inspect.signature(model.create).parameters
 if lang and supports_lang:
     kwargs["lang"] = lang
 
-samples, sample_rate = model.create(text, **kwargs)
+def is_voice_token_overflow(error):
+    return bool(re.search(
+        r"index \\d+ is out of bounds for axis 0 with size 510",
+        str(error),
+        re.IGNORECASE,
+    ))
+
+def split_for_retry(value):
+    midpoint = len(value) // 2
+    boundaries = [
+        index + 1
+        for index, char in enumerate(value)
+        if (char in "。！？!?\\n" or char.isspace()) and 0 < index < len(value) - 1
+    ]
+    split_at = min(boundaries, key=lambda index: abs(index - midpoint)) if boundaries else midpoint
+    return value[:split_at], value[split_at:]
+
+def synthesize_text(value):
+    try:
+        return model.create(value, **kwargs)
+    except IndexError as error:
+        if not is_voice_token_overflow(error) or len(value) < 2:
+            raise
+        left, right = split_for_retry(value)
+        left_samples, left_rate = synthesize_text(left)
+        right_samples, right_rate = synthesize_text(right)
+        if left_rate != right_rate:
+            raise RuntimeError("Kokoro returned inconsistent sample rates across text chunks")
+        samples = list(left_samples)
+        samples.extend(right_samples)
+        return samples, left_rate
+
+samples, sample_rate = synthesize_text(text)
 sf.write(output_path, samples, sample_rate)
 
 duration = len(samples) / sample_rate
@@ -63,7 +95,7 @@ const ESPEAK_LANG_OVERRIDES: Partial<Record<SupportedLang, string>> = {
 // The filename carries a version suffix so older installs automatically
 // upgrade when the script body changes (e.g., adding the `lang` kwarg).
 const SCRIPT_DIR = join(homedir(), ".cache", "hyperframes", "tts");
-const SCRIPT_PATH = join(SCRIPT_DIR, "synth-v2.py");
+const SCRIPT_PATH = join(SCRIPT_DIR, "synth-v3.py");
 
 function ensureSynthScript(): string {
   if (!existsSync(SCRIPT_PATH)) {