diff --git a/src/ui/Features/Video/TextToSpeech/Engines/AudioCppRuntime.cs b/src/ui/Features/Video/TextToSpeech/Engines/AudioCppRuntime.cs
index db0a652831..caaa0c17d3 100644
--- a/src/ui/Features/Video/TextToSpeech/Engines/AudioCppRuntime.cs
+++ b/src/ui/Features/Video/TextToSpeech/Engines/AudioCppRuntime.cs
@@ -1,4 +1,4 @@
-using Nikse.SubtitleEdit.Logic.Config;
+using Nikse.SubtitleEdit.Logic.Config;
 using System;
 using System.IO;
 
@@ -30,6 +30,25 @@ public static string GetSetEngineFolder()
     public static string GetServerExecutable() =>
         Path.Combine(GetSetEngineFolder(), OperatingSystem.IsWindows() ? "audiocpp_server.exe" : "audiocpp_server");
 
+    /// <summary>
+    /// Identity of the server binary on disk (size + last write), or empty when it is missing.
+    /// The engines record it when they launch a server and restart when it changes: an engine
+    /// update extracts a new binary while the old process keeps running, and reusing that
+    /// process means the new build (and any family it adds) is never actually used.
+    /// </summary>
+    public static string GetServerExecutableStamp()
+    {
+        try
+        {
+            var fi = new FileInfo(GetServerExecutable());
+            return fi.Exists ? $"{fi.Length}:{fi.LastWriteTimeUtc.Ticks}" : string.Empty;
+        }
+        catch
+        {
+            return string.Empty;
+        }
+    }
+
     /// <summary>
     /// ggml backend the installed archive was built for, stored at install time. The setting
     /// keeps its historical IndexTts25-prefixed name — it predates the runtime being shared —
diff --git a/src/ui/Features/Video/TextToSpeech/Engines/FishTtsAudioCpp.cs b/src/ui/Features/Video/TextToSpeech/Engines/FishTtsAudioCpp.cs
index 78036f804f..20d7217625 100644
--- a/src/ui/Features/Video/TextToSpeech/Engines/FishTtsAudioCpp.cs
+++ b/src/ui/Features/Video/TextToSpeech/Engines/FishTtsAudioCpp.cs
@@ -1,4 +1,4 @@
-using Nikse.SubtitleEdit.Features.Video.TextToSpeech.ModelLicense;
+using Nikse.SubtitleEdit.Features.Video.TextToSpeech.ModelLicense;
 using Nikse.SubtitleEdit.Features.Video.TextToSpeech.Voices;
 using Nikse.SubtitleEdit.Logic.Config;
 using Nikse.SubtitleEdit.Logic.Download;
@@ -138,6 +138,7 @@ public static string GetModelFileName(string? modelKey) =>
     // Only the model and the backend are baked into the running server — voice is per request.
     private static string? _serverModelKey;
     private static string? _serverBackend;
+    private static string? _serverExeStamp;
     private static bool _processExitHooked;
     private static readonly StringBuilder _serverLog = new();
 
@@ -512,10 +513,12 @@ public async Task<TtsResult> Speak(
     private static async Task EnsureServerRunningAsync(string modelKey, CancellationToken ct)
     {
         var backend = AudioCppRuntime.GetBackend();
+        var exeStamp = AudioCppRuntime.GetServerExecutableStamp();
 
         if (_serverProcess is { HasExited: false } && _serverPort != 0
             && string.Equals(_serverModelKey, modelKey, StringComparison.OrdinalIgnoreCase)
-            && string.Equals(_serverBackend, backend, StringComparison.OrdinalIgnoreCase))
+            && string.Equals(_serverBackend, backend, StringComparison.OrdinalIgnoreCase)
+            && string.Equals(_serverExeStamp, exeStamp, StringComparison.Ordinal))
         {
             return;
         }
@@ -525,7 +528,8 @@ private static async Task EnsureServerRunningAsync(string modelKey, Cancellation
         {
             if (_serverProcess is { HasExited: false } && _serverPort != 0
                 && string.Equals(_serverModelKey, modelKey, StringComparison.OrdinalIgnoreCase)
-                && string.Equals(_serverBackend, backend, StringComparison.OrdinalIgnoreCase))
+                && string.Equals(_serverBackend, backend, StringComparison.OrdinalIgnoreCase)
+                && string.Equals(_serverExeStamp, exeStamp, StringComparison.Ordinal))
             {
                 return;
             }
@@ -596,6 +600,7 @@ private static async Task EnsureServerRunningAsync(string modelKey, Cancellation
             _serverPort = port;
             _serverModelKey = modelKey;
             _serverBackend = backend;
+            _serverExeStamp = AudioCppRuntime.GetServerExecutableStamp();
             HookProcessExitOnce();
 
             // The config uses lazy_load, so /health answers within a second or two — the 5.9 GB
@@ -616,6 +621,7 @@ private static async Task EnsureServerRunningAsync(string modelKey, Cancellation
                     _serverLaunchCommand = null;
                     _serverModelKey = null;
                     _serverBackend = null;
+                    _serverExeStamp = null;
                     throw new InvalidOperationException(
                         $"audiocpp_server exited during startup (code {exitCode}). "
                         + AudioCppRuntime.DescribeStartupExit(exitCode, backend)
@@ -784,6 +790,7 @@ private static void StopServerInternal()
         _serverLaunchCommand = null;
         _serverModelKey = null;
         _serverBackend = null;
+        _serverExeStamp = null;
         if (p == null)
         {
             return;
diff --git a/src/ui/Features/Video/TextToSpeech/Engines/HiggsTtsAudioCpp.cs b/src/ui/Features/Video/TextToSpeech/Engines/HiggsTtsAudioCpp.cs
index 635752da22..b9fc13df1f 100644
--- a/src/ui/Features/Video/TextToSpeech/Engines/HiggsTtsAudioCpp.cs
+++ b/src/ui/Features/Video/TextToSpeech/Engines/HiggsTtsAudioCpp.cs
@@ -1,4 +1,4 @@
-using Nikse.SubtitleEdit.Features.Video.TextToSpeech.ModelLicense;
+using Nikse.SubtitleEdit.Features.Video.TextToSpeech.ModelLicense;
 using Nikse.SubtitleEdit.Features.Video.TextToSpeech.Voices;
 using Nikse.SubtitleEdit.Logic.Config;
 using Nikse.SubtitleEdit.Logic.Download;
@@ -140,6 +140,7 @@ public static string GetModelFileName(string? modelKey) =>
     // Only the model and the backend are baked into the running server — voice is per request.
     private static string? _serverModelKey;
     private static string? _serverBackend;
+    private static string? _serverExeStamp;
     private static bool _processExitHooked;
     private static readonly StringBuilder _serverLog = new();
 
@@ -500,10 +501,12 @@ public async Task<TtsResult> Speak(
     private static async Task EnsureServerRunningAsync(string modelKey, CancellationToken ct)
     {
         var backend = AudioCppRuntime.GetBackend();
+        var exeStamp = AudioCppRuntime.GetServerExecutableStamp();
 
         if (_serverProcess is { HasExited: false } && _serverPort != 0
             && string.Equals(_serverModelKey, modelKey, StringComparison.OrdinalIgnoreCase)
-            && string.Equals(_serverBackend, backend, StringComparison.OrdinalIgnoreCase))
+            && string.Equals(_serverBackend, backend, StringComparison.OrdinalIgnoreCase)
+            && string.Equals(_serverExeStamp, exeStamp, StringComparison.Ordinal))
         {
             return;
         }
@@ -513,7 +516,8 @@ private static async Task EnsureServerRunningAsync(string modelKey, Cancellation
         {
             if (_serverProcess is { HasExited: false } && _serverPort != 0
                 && string.Equals(_serverModelKey, modelKey, StringComparison.OrdinalIgnoreCase)
-                && string.Equals(_serverBackend, backend, StringComparison.OrdinalIgnoreCase))
+                && string.Equals(_serverBackend, backend, StringComparison.OrdinalIgnoreCase)
+                && string.Equals(_serverExeStamp, exeStamp, StringComparison.Ordinal))
             {
                 return;
             }
@@ -584,6 +588,7 @@ private static async Task EnsureServerRunningAsync(string modelKey, Cancellation
             _serverPort = port;
             _serverModelKey = modelKey;
             _serverBackend = backend;
+            _serverExeStamp = AudioCppRuntime.GetServerExecutableStamp();
             HookProcessExitOnce();
 
             // The config uses lazy_load, so /health answers within a second or two — the 4.7 GB
@@ -604,6 +609,7 @@ private static async Task EnsureServerRunningAsync(string modelKey, Cancellation
                     _serverLaunchCommand = null;
                     _serverModelKey = null;
                     _serverBackend = null;
+                    _serverExeStamp = null;
                     throw new InvalidOperationException(
                         $"audiocpp_server exited during startup (code {exitCode}). "
                         + AudioCppRuntime.DescribeStartupExit(exitCode, backend)
@@ -772,6 +778,7 @@ private static void StopServerInternal()
         _serverLaunchCommand = null;
         _serverModelKey = null;
         _serverBackend = null;
+        _serverExeStamp = null;
         if (p == null)
         {
             return;
diff --git a/src/ui/Features/Video/TextToSpeech/Engines/IndexTts25AudioCpp.cs b/src/ui/Features/Video/TextToSpeech/Engines/IndexTts25AudioCpp.cs
index 0cc89c386d..e3a1ba65bd 100644
--- a/src/ui/Features/Video/TextToSpeech/Engines/IndexTts25AudioCpp.cs
+++ b/src/ui/Features/Video/TextToSpeech/Engines/IndexTts25AudioCpp.cs
@@ -1,4 +1,4 @@
-using Nikse.SubtitleEdit.Features.Video.TextToSpeech.Voices;
+using Nikse.SubtitleEdit.Features.Video.TextToSpeech.Voices;
 using Nikse.SubtitleEdit.Logic.Config;
 using Nikse.SubtitleEdit.Logic.Download;
 using Nikse.SubtitleEdit.Logic.Media;
@@ -116,6 +116,7 @@ public static string GetModelFileName(string? modelKey) =>
     // Only the model and the backend are baked into the running server — voice is per request.
     private static string? _serverModelKey;
     private static string? _serverBackend;
+    private static string? _serverExeStamp;
     private static bool _processExitHooked;
     private static readonly StringBuilder _serverLog = new();
 
@@ -522,10 +523,12 @@ public async Task<TtsResult> Speak(
     private static async Task EnsureServerRunningAsync(string modelKey, CancellationToken ct)
     {
         var backend = GetBackend();
+        var exeStamp = AudioCppRuntime.GetServerExecutableStamp();
 
         if (_serverProcess is { HasExited: false } && _serverPort != 0
             && string.Equals(_serverModelKey, modelKey, StringComparison.OrdinalIgnoreCase)
-            && string.Equals(_serverBackend, backend, StringComparison.OrdinalIgnoreCase))
+            && string.Equals(_serverBackend, backend, StringComparison.OrdinalIgnoreCase)
+            && string.Equals(_serverExeStamp, exeStamp, StringComparison.Ordinal))
         {
             return;
         }
@@ -535,7 +538,8 @@ private static async Task EnsureServerRunningAsync(string modelKey, Cancellation
         {
             if (_serverProcess is { HasExited: false } && _serverPort != 0
                 && string.Equals(_serverModelKey, modelKey, StringComparison.OrdinalIgnoreCase)
-                && string.Equals(_serverBackend, backend, StringComparison.OrdinalIgnoreCase))
+                && string.Equals(_serverBackend, backend, StringComparison.OrdinalIgnoreCase)
+                && string.Equals(_serverExeStamp, exeStamp, StringComparison.Ordinal))
             {
                 return;
             }
@@ -608,6 +612,7 @@ private static async Task EnsureServerRunningAsync(string modelKey, Cancellation
             _serverPort = port;
             _serverModelKey = modelKey;
             _serverBackend = backend;
+            _serverExeStamp = AudioCppRuntime.GetServerExecutableStamp();
             HookProcessExitOnce();
 
             // The config uses lazy_load, so /health answers within a second or two — the 3.3 GB
@@ -628,6 +633,7 @@ private static async Task EnsureServerRunningAsync(string modelKey, Cancellation
                     _serverLaunchCommand = null;
                     _serverModelKey = null;
                     _serverBackend = null;
+                    _serverExeStamp = null;
                     throw new InvalidOperationException(
                         $"audiocpp_server exited during startup (code {exitCode}). "
                         + DescribeStartupExit(exitCode, backend)
@@ -798,6 +804,7 @@ private static void StopServerInternal()
         _serverLaunchCommand = null;
         _serverModelKey = null;
         _serverBackend = null;
+        _serverExeStamp = null;
         if (p == null)
         {
             return;
diff --git a/src/ui/Features/Video/TextToSpeech/TtsVoiceInstaller.cs b/src/ui/Features/Video/TextToSpeech/TtsVoiceInstaller.cs
index ab00ee2e97..0ec258fe95 100644
--- a/src/ui/Features/Video/TextToSpeech/TtsVoiceInstaller.cs
+++ b/src/ui/Features/Video/TextToSpeech/TtsVoiceInstaller.cs
@@ -1,4 +1,4 @@
-using Avalonia.Controls;
+using Avalonia.Controls;
 using Nikse.SubtitleEdit.Core.Common;
 using Nikse.SubtitleEdit.Features.Shared;
 using Nikse.SubtitleEdit.Features.Video.SpeechToText;
@@ -312,6 +312,14 @@ await MessageBox.Show(
             return false;
         }
 
+        // The three audio.cpp engines share this binary. A server started from the old build
+        // must go before the archive is extracted: on Windows the running exe cannot be
+        // overwritten at all, and elsewhere the stale process would keep answering requests
+        // (and reject any model family the new build added) until Subtitle Edit restarts.
+        IndexTts25AudioCpp.StopServer();
+        HiggsTtsAudioCpp.StopServer();
+        FishTtsAudioCpp.StopServer();
+
         var dlResult = await windowService.ShowDialogAsync<DownloadTtsWindow, DownloadTtsViewModel>(
             window, vm => vm.StartDownloadIndexTts25AudioCppEngine(backend));