diff --git a/tests/audio/tts-utils.test.ts b/tests/audio/tts-utils.test.ts
new file mode 100644
index 000000000..64c5ef231
--- /dev/null
+++ b/tests/audio/tts-utils.test.ts
@@ -0,0 +1,83 @@
+import { describe, expect, it } from 'vitest';
+
+import {
+  splitLongSpeechText,
+  splitLongSpeechActions,
+  TTS_MAX_TEXT_LENGTH,
+} from '@/lib/audio/tts-utils';
+import type { Action, SpeechAction } from '@/lib/types/action';
+
+describe('splitLongSpeechText', () => {
+  it('returns the trimmed text unchanged when within the limit', () => {
+    expect(splitLongSpeechText('  hello world  ', 100)).toEqual(['hello world']);
+  });
+
+  it('splits at sentence boundaries, keeping the punctuation', () => {
+    expect(splitLongSpeechText('句子一。句子二！句子三？', 6)).toEqual([
+      '句子一。',
+      '句子二！',
+      '句子三？',
+    ]);
+  });
+
+  it('packs consecutive short sentences up to the limit', () => {
+    const out = splitLongSpeechText('aa。bb。cc。dd。', 6);
+    expect(out).toEqual(['aa。bb。', 'cc。dd。']);
+    expect(out.join('')).toBe('aa。bb。cc。dd。');
+  });
+
+  it('falls back to clause punctuation for an over-long sentence', () => {
+    const out = splitLongSpeechText('一，二，三，四，五', 4);
+    expect(out.every((c) => c.length <= 4)).toBe(true);
+    expect(out.join('').replace(/，/g, '')).toContain('一');
+  });
+
+  it('hard-splits a long run with no punctuation', () => {
+    expect(splitLongSpeechText('x'.repeat(25), 10)).toEqual(['xxxxxxxxxx', 'xxxxxxxxxx', 'xxxxx']);
+  });
+
+  it('never emits a chunk longer than maxLength, nor an empty chunk (invariant)', () => {
+    const text = '这是一个很长的句子，包含很多子句、标点。'.repeat(50) + 'z'.repeat(120);
+    for (const max of [8, 16, 64, 200]) {
+      const out = splitLongSpeechText(text, max);
+      expect(
+        out.every((c) => c.length <= max),
+        `max=${max}`,
+      ).toBe(true);
+      expect(
+        out.some((c) => c.length === 0),
+        `max=${max}`,
+      ).toBe(false);
+    }
+  });
+});
+
+describe('splitLongSpeechActions', () => {
+  const speech = (id: string, text: string): SpeechAction =>
+    ({ id, type: 'speech', text }) as SpeechAction;
+
+  it('returns actions untouched for a provider with no length limit', () => {
+    const actions: Action[] = [speech('a', 'x'.repeat(5000))];
+    expect(splitLongSpeechActions(actions, 'openai-tts')).toBe(actions);
+  });
+
+  it('leaves short speech and non-speech actions unchanged', () => {
+    const actions: Action[] = [speech('s', 'short'), { id: 'sp', type: 'spotlight' } as Action];
+    expect(splitLongSpeechActions(actions, 'glm-tts')).toBe(actions);
+  });
+
+  it('splits an over-limit speech action into chunked sub-actions', () => {
+    const max = TTS_MAX_TEXT_LENGTH['glm-tts']!; // 1024
+    const long = '句子。'.repeat(400); // 1200 chars > 1024
+    const out = splitLongSpeechActions([speech('a', long)], 'glm-tts') as SpeechAction[];
+
+    expect(out.length).toBeGreaterThan(1);
+    expect(out.every((a) => a.type === 'speech' && a.text.length <= max)).toBe(true);
+    // deterministic sub-action ids…
+    expect(out.map((a) => a.id)).toEqual(out.map((_, i) => `a_tts_${i + 1}`));
+    // …each gets its own audio (the parent audioId is dropped)…
+    expect(out.every((a) => a.audioId === undefined)).toBe(true);
+    // …and the text is preserved across the split.
+    expect(out.map((a) => a.text).join('')).toBe(long);
+  });
+});