diff --git a/custom_components/speaker_recognition/__init__.py b/custom_components/speaker_recognition/__init__.py
index d8ef158..218ddfa 100644
--- a/custom_components/speaker_recognition/__init__.py
+++ b/custom_components/speaker_recognition/__init__.py
@@ -2,9 +2,10 @@
 
 from __future__ import annotations
 
+from homeassistant.components import persistent_notification
 from homeassistant.config_entries import ConfigEntry
 from homeassistant.const import Platform
-from homeassistant.core import HomeAssistant
+from homeassistant.core import HomeAssistant, callback
 
 from .const import (
     CONF_BACKEND_URL,
@@ -18,6 +19,31 @@
 
 type SpeakerRecognitionConfigEntry = ConfigEntry[SpeakerRecognition]
 
+TRAINING_NOTIFICATION_ID = "speaker_recognition_training"
+
+
+@callback
+def _async_update_training_notification(
+    hass: HomeAssistant, recognition: SpeakerRecognition
+) -> None:
+    """Notify the user if the last training attempt failed; clear it otherwise.
+
+    Training happens in the background, so without this a failure (backend
+    unreachable, audio files not found, ...) would only appear in the logs.
+    """
+    if recognition.last_train_error:
+        persistent_notification.async_create(
+            hass,
+            (
+                "Speaker Recognition could not train the configured voices:\n\n"
+                f"{recognition.last_train_error}"
+            ),
+            title="Speaker Recognition",
+            notification_id=TRAINING_NOTIFICATION_ID,
+        )
+    else:
+        persistent_notification.async_dismiss(hass, TRAINING_NOTIFICATION_ID)
+
 
 def _get_main_entry(hass: HomeAssistant) -> ConfigEntry | None:
     """Get the main config entry."""
@@ -50,6 +76,7 @@ async def async_setup_main_entry(
 
     if voice_samples:
         await recognition.async_train()
+        _async_update_training_notification(hass, recognition)
 
     entry.runtime_data = recognition
     entry.async_on_unload(entry.add_update_listener(async_update_main_listener))
diff --git a/custom_components/speaker_recognition/recognition.py b/custom_components/speaker_recognition/recognition.py
index f474660..0f5ffbd 100644
--- a/custom_components/speaker_recognition/recognition.py
+++ b/custom_components/speaker_recognition/recognition.py
@@ -43,10 +43,15 @@ def __init__(
         self.hass = hass
         self.voice_samples = voice_samples
         self._trained = False
+        # Human-readable reason the last training attempt failed, or None if the
+        # last attempt succeeded (or there was nothing to train). Surfaced to the
+        # user as a notification by __init__.py.
+        self.last_train_error: str | None = None
         self._client = SpeakerRecognitionClient(base_url=base_url, timeout=300.0)
 
     async def async_train(self) -> None:
         """Train the speaker recognition model with configured voice samples."""
+        self.last_train_error = None
         _LOGGER.debug(
             "Training speaker recognition with %d voice samples",
             len(self.voice_samples),
@@ -91,6 +96,11 @@ async def async_train(self) -> None:
             if not voice_sample_models:
                 _LOGGER.warning("No valid training samples prepared")
                 self._trained = False
+                self.last_train_error = (
+                    "No valid audio files were found for the configured voice "
+                    "samples. Make sure the selected files exist under Home "
+                    "Assistant's media folder."
+                )
                 return
 
             request = TrainingRequest(voice_samples=voice_sample_models)
@@ -99,6 +109,7 @@ async def async_train(self) -> None:
         except (OSError, ValueError, TypeError) as error:
             _LOGGER.error("Error during training: %s", error)
             self._trained = False
+            self.last_train_error = str(error)
         else:
             self._trained = True
             _LOGGER.info(