diff --git a/.gitignore b/.gitignore
index db41c35..f393843 100644
--- a/.gitignore
+++ b/.gitignore
@@ -31,3 +31,4 @@ Thumbs.db
 # Compiled binaries and output
 find_all_keys_macos
 decoded_voices/
+voice_transcriptions.json
diff --git a/mcp_server.py b/mcp_server.py
index c464a06..15ec657 100644
--- a/mcp_server.py
+++ b/mcp_server.py
@@ -6,7 +6,7 @@ Runs on Windows Python (needs access to D:\ WeChat databases).
 """
 
 import io
-import os, sys, json, time, sqlite3, tempfile, struct, hashlib, atexit, re
+import os, sys, json, time, sqlite3, tempfile, struct, hashlib, atexit, re, threading
 import wave
 import hmac as hmac_mod
 from contextlib import closing
@@ -1890,9 +1890,92 @@ def decode_voice(chat_name: str, local_id: int) -> str:
     )
 
 
+# ============ 语音转录缓存 ============
+#
+# Whisper 转录耗时（CPU 下每条数秒到数十秒），且结果是确定性的
+# （同一段 voice_data → 同一段 text），非常适合缓存。
+#
+# 缓存 key 用 json.dumps([username, local_id])：local_id 在单个 username 下
+# 稳定唯一，套一层 JSON 序列化保证 username 里若含分隔符也不会与其它条目碰撞。
+#
+# 写入走 temp + os.replace 原子替换，避免进程中途被杀导致整份缓存损坏
+# （Whisper 的单次代价远高于 DBCache，破档不可接受）。
+#
+# 条目里记录 model_size：Whisper 升级默认模型后，旧条目自动视为失效并重跑。
+
+VOICE_TRANSCRIPTION_CACHE_FILE = os.path.join(SCRIPT_DIR, "voice_transcriptions.json")
+
+_voice_transcription_cache = None  # 懒加载 dict；None 表示尚未加载
+_voice_transcription_cache_lock = threading.Lock()
+_voice_transcription_save_warned = False  # 写失败仅首次写 stderr，避免刷屏
+
+
+def _voice_transcription_cache_key(username, local_id):
+    """构造缓存 key。用 json.dumps 兜底 username 里可能出现的分隔符。"""
+    return json.dumps([username, int(local_id)], ensure_ascii=False)
+
+
+def _load_voice_transcription_cache():
+    """加载缓存到模块级 dict，返回该 dict。
+
+    文件不存在 → 空 dict。JSON 损坏或 payload 非 dict → 空 dict
+    （与上游 DBCache 的容错风格一致：缓存坏了不要拖垮工具调用）。
+    """
+    global _voice_transcription_cache
+    with _voice_transcription_cache_lock:
+        if _voice_transcription_cache is not None:
+            return _voice_transcription_cache
+        if not os.path.exists(VOICE_TRANSCRIPTION_CACHE_FILE):
+            _voice_transcription_cache = {}
+            return _voice_transcription_cache
+        try:
+            with open(VOICE_TRANSCRIPTION_CACHE_FILE, encoding="utf-8") as f:
+                loaded = json.load(f)
+            _voice_transcription_cache = loaded if isinstance(loaded, dict) else {}
+        except (json.JSONDecodeError, OSError):
+            _voice_transcription_cache = {}
+        return _voice_transcription_cache
+
+
+def _save_voice_transcription_cache():
+    """持久化缓存到磁盘。
+
+    - 原子写：先写 .tmp 再 os.replace，避免 crash 中途留下半截文件。
+    - 未加载过也允许保存：此时把 module 状态初始化为空 dict，避免上层
+      代码因调用顺序错误而静默丢数据。
+    - OSError 不抛：避免转录成功但落盘失败时让工具调用也失败；但首次
+      失败会在 stderr 打一行警告，用户知道磁盘满 / 权限问题需要处理。
+    """
+    global _voice_transcription_cache, _voice_transcription_save_warned
+    with _voice_transcription_cache_lock:
+        if _voice_transcription_cache is None:
+            _voice_transcription_cache = {}
+        tmp_path = VOICE_TRANSCRIPTION_CACHE_FILE + ".tmp"
+        try:
+            with open(tmp_path, "w", encoding="utf-8") as f:
+                json.dump(_voice_transcription_cache, f, ensure_ascii=False)
+            os.replace(tmp_path, VOICE_TRANSCRIPTION_CACHE_FILE)
+        except OSError as exc:
+            if not _voice_transcription_save_warned:
+                print(
+                    f"[voice_cache] 写入失败（后续不再提示）: {exc}",
+                    file=sys.stderr,
+                    flush=True,
+                )
+                _voice_transcription_save_warned = True
+            # 清理可能残留的 .tmp
+            try:
+                if os.path.exists(tmp_path):
+                    os.unlink(tmp_path)
+            except OSError:
+                pass
+
+
+DEFAULT_WHISPER_MODEL = "base"
+
 _whisper_model = None
 
-def _get_whisper_model(model_size="base"):
+def _get_whisper_model(model_size=DEFAULT_WHISPER_MODEL):
     global _whisper_model
     if _whisper_model is None:
         import whisper
@@ -1904,8 +1987,10 @@ def _get_whisper_model(model_size="base"):
 def transcribe_voice(chat_name: str, local_id: int) -> str:
     """将微信语音消息转录为文字（自动检测语言，保留原语言）。
 
-    会先解码 SILK 语音为 WAV，再用 Whisper 转录。
-    首次运行会下载 Whisper 模型（约 145MB）。
+    首次转录会先解码 SILK 语音为 WAV，再用 Whisper 转录；结果缓存到
+    voice_transcriptions.json，重复调用直接返回缓存（跳过 SILK 解码
+    和 Whisper 推理）。若 Whisper 默认模型升级（如 base → small），
+    旧条目自动视为失效并重新转录。首次运行会下载 Whisper 模型（约 145MB）。
 
     依赖: pip install silk-python openai-whisper (silk-python 的 import 名为 pysilk)
 
@@ -1913,6 +1998,29 @@ def transcribe_voice(chat_name: str, local_id: int) -> str:
         chat_name: 聊天对象的名字、备注名或wxid
         local_id: 语音消息的 local_id（从 get_voice_messages 获取）
     """
+    username = resolve_username(chat_name)
+    if not username:
+        return f"找不到聊天对象: {chat_name}"
+
+    cache_key = _voice_transcription_cache_key(username, local_id)
+    cache = _load_voice_transcription_cache()
+    entry = cache.get(cache_key)
+    if (
+        isinstance(entry, dict)
+        and "text" in entry
+        and entry.get("model_size") == DEFAULT_WHISPER_MODEL
+    ):
+        # 命中缓存：跳过 DB 查询、SILK 解码、Whisper 推理。
+        # 条目里存了 create_time，即使源 DB 中消息已被清理仍能返回历史转录。
+        lang = entry.get("language", "unknown")
+        cached_ts = entry.get("create_time")
+        if isinstance(cached_ts, int):
+            time_label = datetime.fromtimestamp(cached_ts).strftime('%Y-%m-%d %H:%M')
+        else:
+            time_label = "-"
+        return f"[{time_label}] ({lang})\n{entry['text']}"
+
+    # 未命中：只有这条路径才需要 whisper / pysilk。
     try:
         import whisper  # noqa: F401
     except ImportError:
@@ -1922,10 +2030,6 @@ def transcribe_voice(chat_name: str, local_id: int) -> str:
     except ImportError:
         return "缺少依赖: pip install silk-python"
 
-    username = resolve_username(chat_name)
-    if not username:
-        return f"找不到聊天对象: {chat_name}"
-
     row = _fetch_voice_row(username, local_id)
     if row is None:
         return f"找不到 local_id={local_id} 的语音消息"
@@ -1938,6 +2042,16 @@ def transcribe_voice(chat_name: str, local_id: int) -> str:
     lang = result.get("language", "unknown")
     text = result.get("text", "").strip()
 
+    # 写缓存：即使 text 为空也缓存（Whisper 偶尔对静音/极短片段返回空），
+    # 配合 model_size 字段，升级模型后会自动重转，避免永久钉死空结果。
+    cache[cache_key] = {
+        "text": text,
+        "language": lang,
+        "create_time": int(create_time),
+        "model_size": DEFAULT_WHISPER_MODEL,
+    }
+    _save_voice_transcription_cache()
+
     time_label = datetime.fromtimestamp(create_time).strftime('%Y-%m-%d %H:%M')
     return f"[{time_label}] ({lang})\n{text}"
 
diff --git a/tests/test_voice_transcription_cache.py b/tests/test_voice_transcription_cache.py
new file mode 100644
index 0000000..957596b
--- /dev/null
+++ b/tests/test_voice_transcription_cache.py
@@ -0,0 +1,274 @@
+import json
+import os
+import tempfile
+import threading
+import unittest
+from unittest.mock import patch
+
+import mcp_server
+
+
+class _CacheIsolationMixin:
+    """所有测试共享：隔离 module-level 缓存状态 + 指向 tempdir 的 cache 文件。"""
+
+    def setUp(self):
+        self._saved_cache = mcp_server._voice_transcription_cache
+        self._saved_path = mcp_server.VOICE_TRANSCRIPTION_CACHE_FILE
+        self._saved_warned = mcp_server._voice_transcription_save_warned
+
+        mcp_server._voice_transcription_cache = None
+        mcp_server._voice_transcription_save_warned = False
+
+        self._tmp = tempfile.TemporaryDirectory()
+        self.addCleanup(self._tmp.cleanup)
+        mcp_server.VOICE_TRANSCRIPTION_CACHE_FILE = os.path.join(
+            self._tmp.name, "voice_transcriptions.json"
+        )
+
+    def tearDown(self):
+        mcp_server._voice_transcription_cache = self._saved_cache
+        mcp_server.VOICE_TRANSCRIPTION_CACHE_FILE = self._saved_path
+        mcp_server._voice_transcription_save_warned = self._saved_warned
+
+
+class VoiceTranscriptionCachePersistenceTests(_CacheIsolationMixin, unittest.TestCase):
+    """_load_voice_transcription_cache / _save_voice_transcription_cache 的持久化行为。"""
+
+    def test_load_missing_file_returns_empty_dict(self):
+        self.assertEqual(mcp_server._load_voice_transcription_cache(), {})
+
+    def test_save_and_reload_roundtrip(self):
+        cache = mcp_server._load_voice_transcription_cache()
+        cache["wxid_foo:42"] = {
+            "text": "你好",
+            "language": "zh",
+            "create_time": 1700000000,
+            "model_size": "base",
+        }
+        mcp_server._save_voice_transcription_cache()
+
+        # 强制下一次 load 从磁盘读
+        mcp_server._voice_transcription_cache = None
+        reloaded = mcp_server._load_voice_transcription_cache()
+        self.assertEqual(reloaded["wxid_foo:42"]["text"], "你好")
+        self.assertEqual(reloaded["wxid_foo:42"]["language"], "zh")
+
+    def test_corrupt_file_returns_empty_dict(self):
+        with open(mcp_server.VOICE_TRANSCRIPTION_CACHE_FILE, "w", encoding="utf-8") as f:
+            f.write("{{ not valid json")
+        self.assertEqual(mcp_server._load_voice_transcription_cache(), {})
+
+    def test_non_dict_payload_returns_empty_dict(self):
+        with open(mcp_server.VOICE_TRANSCRIPTION_CACHE_FILE, "w", encoding="utf-8") as f:
+            json.dump(["not", "a", "dict"], f)
+        self.assertEqual(mcp_server._load_voice_transcription_cache(), {})
+
+    def test_utf8_preserved_on_disk(self):
+        # ensure_ascii=False 必须生效，否则中文会被转义成 \uXXXX
+        cache = mcp_server._load_voice_transcription_cache()
+        cache["wxid_bar:1"] = {"text": "中文测试", "language": "zh"}
+        mcp_server._save_voice_transcription_cache()
+
+        with open(mcp_server.VOICE_TRANSCRIPTION_CACHE_FILE, "rb") as f:
+            raw = f.read()
+        self.assertIn("中文测试".encode("utf-8"), raw)
+
+    def test_save_without_prior_load_persists_empty_dict(self):
+        # 从未 load 过就直接 save：应落盘一个空 dict，而不是静默丢弃。
+        mcp_server._voice_transcription_cache = None
+        mcp_server._save_voice_transcription_cache()
+        self.assertTrue(os.path.exists(mcp_server.VOICE_TRANSCRIPTION_CACHE_FILE))
+        with open(mcp_server.VOICE_TRANSCRIPTION_CACHE_FILE, encoding="utf-8") as f:
+            self.assertEqual(json.load(f), {})
+
+
+class VoiceTranscriptionCacheAtomicityTests(_CacheIsolationMixin, unittest.TestCase):
+    """原子写 + crash-during-save 行为。"""
+
+    def test_write_is_atomic_via_rename(self):
+        # 先写入一份已有缓存
+        cache = mcp_server._load_voice_transcription_cache()
+        cache["wxid_x:1"] = {"text": "initial", "language": "zh", "model_size": "base"}
+        mcp_server._save_voice_transcription_cache()
+
+        # 模拟：写 .tmp 正常但 os.replace 阶段失败
+        original_replace = os.replace
+
+        def flaky_replace(src, dst):
+            raise OSError("disk full during rename")
+
+        cache["wxid_x:1"] = {"text": "MUTATED", "language": "zh", "model_size": "base"}
+        with patch.object(os, "replace", side_effect=flaky_replace):
+            mcp_server._save_voice_transcription_cache()  # 不应抛
+
+        # 磁盘上应仍然是 initial，不是 MUTATED，也不是损坏的半截文件
+        with open(mcp_server.VOICE_TRANSCRIPTION_CACHE_FILE, encoding="utf-8") as f:
+            disk = json.load(f)
+        self.assertEqual(disk["wxid_x:1"]["text"], "initial")
+
+        # .tmp 应该被清理，避免污染目录
+        tmp_path = mcp_server.VOICE_TRANSCRIPTION_CACHE_FILE + ".tmp"
+        # 注：patch 生效期间 os.replace 失败，finally 里会尝试 unlink
+        _ = original_replace  # 防 lint 警告
+        self.assertFalse(os.path.exists(tmp_path))
+
+    def test_early_save_error_preserves_existing_file(self):
+        # json.dump 在 .tmp 上抛异常时（模拟磁盘满 / 权限问题），主文件应保持原样；
+        # 注意此测试不是"写到一半中断"而是"写前就失败"的场景。
+        cache = mcp_server._load_voice_transcription_cache()
+        cache["wxid_y:1"] = {"text": "survives", "language": "zh", "model_size": "base"}
+        mcp_server._save_voice_transcription_cache()
+
+        cache["wxid_y:1"] = {"text": "DO NOT SEE", "language": "zh", "model_size": "base"}
+
+        def boom(*args, **kwargs):
+            raise OSError("disk full")
+
+        with patch.object(mcp_server.json, "dump", side_effect=boom):
+            mcp_server._save_voice_transcription_cache()  # 静默降级，不抛
+
+        # 主文件没被破坏：仍然可 json.load 出原先内容
+        mcp_server._voice_transcription_cache = None
+        reloaded = mcp_server._load_voice_transcription_cache()
+        self.assertEqual(reloaded["wxid_y:1"]["text"], "survives")
+
+
+class VoiceTranscriptionCacheConcurrencyTests(_CacheIsolationMixin, unittest.TestCase):
+    """多线程下的 load/save 行为。"""
+
+    def test_concurrent_load_returns_same_dict_instance(self):
+        # 16 个线程同时触发首次 load，应当只实际化一份 dict（lock 生效）
+        barrier = threading.Barrier(16)
+        results = []
+        results_lock = threading.Lock()
+
+        def worker():
+            barrier.wait()
+            d = mcp_server._load_voice_transcription_cache()
+            with results_lock:
+                results.append(id(d))
+
+        threads = [threading.Thread(target=worker) for _ in range(16)]
+        for t in threads:
+            t.start()
+        for t in threads:
+            t.join()
+
+        self.assertEqual(len(set(results)), 1, "并发 load 应返回同一个 dict 对象")
+
+    def test_concurrent_save_does_not_corrupt(self):
+        # 多个线程同时 save，磁盘上最终文件必须是合法 JSON（原子写 + lock 保障）
+        cache = mcp_server._load_voice_transcription_cache()
+        for i in range(100):
+            cache[f"wxid_z:{i}"] = {
+                "text": f"msg-{i}",
+                "language": "zh",
+                "model_size": "base",
+            }
+
+        def worker():
+            mcp_server._save_voice_transcription_cache()
+
+        threads = [threading.Thread(target=worker) for _ in range(8)]
+        for t in threads:
+            t.start()
+        for t in threads:
+            t.join()
+
+        with open(mcp_server.VOICE_TRANSCRIPTION_CACHE_FILE, encoding="utf-8") as f:
+            disk = json.load(f)  # 必须能解析
+        self.assertEqual(len(disk), 100)
+
+
+class TranscribeVoiceCacheHitTests(_CacheIsolationMixin, unittest.TestCase):
+    """transcribe_voice 的缓存命中 / 失效路径。"""
+
+    def _seed(self, key, entry):
+        cache = mcp_server._load_voice_transcription_cache()
+        cache[key] = entry
+        mcp_server._save_voice_transcription_cache()
+
+    def test_cache_hit_skips_fetch_and_transcribe(self):
+        key = mcp_server._voice_transcription_cache_key("wxid_test", 7)
+        self._seed(key, {
+            "text": "缓存命中文本",
+            "language": "zh",
+            "create_time": 1700000000,
+            "model_size": mcp_server.DEFAULT_WHISPER_MODEL,
+        })
+
+        with patch.object(mcp_server, "resolve_username", return_value="wxid_test") as mock_resolve, \
+             patch.object(mcp_server, "_fetch_voice_row") as mock_fetch, \
+             patch.object(mcp_server, "_silk_to_wav") as mock_silk, \
+             patch.object(mcp_server, "_get_whisper_model") as mock_model:
+            result = mcp_server.transcribe_voice("test_contact", 7)
+
+        mock_resolve.assert_called_once_with("test_contact")
+        mock_fetch.assert_not_called()
+        mock_silk.assert_not_called()
+        mock_model.assert_not_called()
+        self.assertIn("缓存命中文本", result)
+        self.assertIn("(zh)", result)
+
+    def test_cache_hit_uses_placeholder_when_create_time_missing(self):
+        # 旧条目若没有 create_time 字段，不应崩溃
+        key = mcp_server._voice_transcription_cache_key("wxid_test", 8)
+        self._seed(key, {
+            "text": "历史条目",
+            "language": "zh",
+            "model_size": mcp_server.DEFAULT_WHISPER_MODEL,
+        })
+
+        with patch.object(mcp_server, "resolve_username", return_value="wxid_test"), \
+             patch.object(mcp_server, "_fetch_voice_row") as mock_fetch:
+            result = mcp_server.transcribe_voice("test_contact", 8)
+
+        mock_fetch.assert_not_called()
+        self.assertIn("历史条目", result)
+
+    def test_cache_hit_returns_empty_text_without_retranscribing(self):
+        # Whisper 返回空也要缓存；再次调用应直接返回空，不进入 miss 路径
+        key = mcp_server._voice_transcription_cache_key("wxid_test", 9)
+        self._seed(key, {
+            "text": "",
+            "language": "zh",
+            "create_time": 1700000000,
+            "model_size": mcp_server.DEFAULT_WHISPER_MODEL,
+        })
+
+        with patch.object(mcp_server, "resolve_username", return_value="wxid_test"), \
+             patch.object(mcp_server, "_fetch_voice_row") as mock_fetch:
+            result = mcp_server.transcribe_voice("test_contact", 9)
+
+        mock_fetch.assert_not_called()
+        self.assertIn("(zh)", result)
+
+    def test_model_mismatch_is_treated_as_miss(self):
+        # 缓存条目的 model_size 和当前 DEFAULT_WHISPER_MODEL 不一致时，
+        # 不应命中；进入 miss 路径（这里无 whisper 依赖，应落到"缺少依赖"分支）。
+        key = mcp_server._voice_transcription_cache_key("wxid_test", 10)
+        self._seed(key, {
+            "text": "旧模型结果",
+            "language": "zh",
+            "create_time": 1700000000,
+            "model_size": "OUTDATED_MODEL",
+        })
+
+        with patch.object(mcp_server, "resolve_username", return_value="wxid_test"), \
+             patch.dict("sys.modules", {"whisper": None}):
+            # whisper=None 时 `import whisper` 触发 ImportError
+            result = mcp_server.transcribe_voice("test_contact", 10)
+
+        # 走了 miss 路径 → 返回缺依赖提示，而不是返回旧缓存文本
+        self.assertNotIn("旧模型结果", result)
+        self.assertIn("缺少依赖", result)
+
+    def test_cache_key_handles_colon_in_username(self):
+        # 若上游未来的 resolve_username 放出带 ':' 的 username，也不会和其他条目冲突
+        key_a = mcp_server._voice_transcription_cache_key("wxid:foo", 1)
+        key_b = mcp_server._voice_transcription_cache_key("wxid", 1)
+        self.assertNotEqual(key_a, key_b)
+
+
+if __name__ == "__main__":
+    unittest.main()