diff --git a/content/browser/speech/speech_recognition_dispatcher_host.cc b/content/browser/speech/speech_recognition_dispatcher_host.cc
index 3cc9651bc822e..5321a3a331d99 100644
--- a/content/browser/speech/speech_recognition_dispatcher_host.cc
+++ b/content/browser/speech/speech_recognition_dispatcher_host.cc
@@ -32,6 +32,7 @@
 #include "mojo/public/cpp/bindings/message.h"
 #include "mojo/public/cpp/bindings/self_owned_receiver.h"
 #include "services/network/public/cpp/shared_url_loader_factory.h"
+#include "services/network/public/mojom/permissions_policy/permissions_policy_feature.mojom.h"
 
 #if !BUILDFLAG(IS_ANDROID)
 #include "components/soda/soda_util.h"
@@ -175,6 +176,13 @@ void SpeechRecognitionDispatcherHost::StartRequestOnUI(
       storage_partition == browser_context->GetDefaultStoragePartition()
           ? true
           : !rfh->GetLastCommittedURL().SchemeIsHTTPOrHTTPS();
+  // Browser-side enforcement of the on-device-speech-recognition
+  // Permissions-Policy. The renderer is expected to gate `on_device` itself,
+  // but a compromised renderer can lie, so re-check authoritatively here.
+  if (!rfh->IsFeatureEnabled(network::mojom::PermissionsPolicyFeature::
+                                 kOnDeviceSpeechRecognition)) {
+    can_render_frame_use_on_device = false;
+  }
   const std::string& language =
       SpeechRecognitionDispatcherHost::GetAcceptedLanguages(
           params->language,
diff --git a/third_party/blink/renderer/modules/speech/speech_recognition.cc b/third_party/blink/renderer/modules/speech/speech_recognition.cc
index 386be511d2335..befcccb0170aa 100644
--- a/third_party/blink/renderer/modules/speech/speech_recognition.cc
+++ b/third_party/blink/renderer/modules/speech/speech_recognition.cc
@@ -605,10 +605,18 @@ void SpeechRecognition::StartController(
       GetExecutionContext()->GetTaskRunner(TaskType::kMiscPlatformAPI));
   receiver_.set_disconnect_handler(BindOnce(
       &SpeechRecognition::OnConnectionError, WrapWeakPersistent(this)));
+  // Only request on-device recognition when the frame is permitted to use it
+  // (same gates as available()/install()): the on-device-speech-recognition
+  // Permissions-Policy must be granted and the frame must not be a cross-site
+  // subframe.
+  const bool on_device_allowed =
+      DomWindow() && !DomWindow()->IsCrossSiteSubframeIncludingScheme() &&
+      DomWindow()->IsFeatureEnabled(network::mojom::PermissionsPolicyFeature::
+                                        kOnDeviceSpeechRecognition);
   auto params = controller_->BuildStartSpeechRecognitionRequestParams(
       std::move(session_receiver), std::move(session_client), *grammars_,
       phrases_.Get(), lang_, continuous_, interim_results_, max_alternatives_,
-      /*on_device=*/true,  // On-device speech recognition is always preferred.
+      /*on_device=*/on_device_allowed,
       /*allow_cloud_fallback=*/!process_locally_,
       SpeechRecognitionQualityToMojom(quality_),
       std::move(audio_forwarder_receiver), std::move(audio_parameters));