diff --git a/tools/server/README.md b/tools/server/README.md
index 56d1ea5fe885..deda7c5b4ff9 100644
--- a/tools/server/README.md
+++ b/tools/server/README.md
@@ -1588,7 +1588,7 @@ Audio must be PCM16, mono, at the model sample rate. The sample rate is returned
 
 *Client events:*
 
-- `session.update`: accepted for compatibility, the server replies with `session.updated`
+- `session.update`: accepted for compatibility, the server replies with `session.updated`. If `session.audio.input.format.rate` is set to a rate other than the model sample rate, the server replies with an `error` instead
 - `input_audio_buffer.append`: `audio` holds base64 PCM16 samples
 - `input_audio_buffer.commit`: flushes the audio received so far, so the words still held back by the model's delay are produced. The server replies with `input_audio_buffer.committed` and starts a new item. With `"final": true`, the input ends instead and the last item is completed once the stream drains.
 - `input_audio_buffer.clear`: accepted for compatibility, audio already sent cannot be discarded
@@ -1603,7 +1603,7 @@ Closing the websocket also ends the input.
 - `input_audio_buffer.committed`: `item_id` of the committed item, `previous_item_id` of the item committed before it (`null` for the first one)
 - `error`: `error.message` describes the problem
 
-Note that words held back by the model's delay at the time of a commit are attributed to the next item.
+The `completed` event of a commit includes the words held back by the model's delay at the time of the commit.
 
 ### POST `/v1/responses`: OpenAI-compatible Responses API
 
diff --git a/tools/server/server-context.cpp b/tools/server/server-context.cpp
index a5cba1d71d46..916b8d035374 100644
--- a/tools/server/server-context.cpp
+++ b/tools/server/server-context.cpp
@@ -6106,6 +6106,14 @@ void server_routes::init_routes() {
                 }
                 const std::string type = json_value(ev, "type", std::string());
                 if (type == "session.update" || type == "transcription_session.update") {
+                    // audio sent at another rate would still be read at the model rate
+                    const int32_t rate = mtmd_get_audio_sample_rate(ctx_server.mctx);
+                    const json format = ev.value("session", json()).value("audio", json()).value("input", json()).value("format", json());
+                    // json_value() would fall back to the default for a string and truncate a fraction
+                    if (format.contains("rate") && !(format.at("rate").is_number_integer() && format.at("rate").get<int64_t>() == rate)) {
+                        send_error_event(string_format("audio.input.format.rate must be %d", rate), "invalid_request_error");
+                        continue;
+                    }
                     send_event(json {{"type", "session.updated"}, {"session", session_config}});
                 } else if (type == "input_audio_buffer.append") {
                     std::string pcm;
diff --git a/tools/server/tests/unit/test_realtime.py b/tools/server/tests/unit/test_realtime.py
index 608904b070ad..cc3669bb0c62 100644
--- a/tools/server/tests/unit/test_realtime.py
+++ b/tools/server/tests/unit/test_realtime.py
@@ -81,6 +81,9 @@ def test_realtime_control_events():
     start_voxtral_rt()
     events = run_realtime([
         {"type": "session.update", "session": {}},
+        {"type": "session.update", "session": {"audio": {"input": {"format": {"type": "audio/pcm", "rate": 24000}}}}},
+        {"type": "session.update", "session": {"audio": {"input": {"format": {"type": "audio/pcm", "rate": "24000"}}}}},
+        {"type": "session.update", "session": {"audio": {"input": {"format": {"type": "audio/pcm", "rate": 16000}}}}},
         {"type": "input_audio_buffer.clear"},
         {"type": "not.an.event"},
         {"type": "input_audio_buffer.commit", "final": True},
@@ -88,11 +91,16 @@ def test_realtime_control_events():
     assert [ev["type"] for ev in events if ev["type"] != EV_DELTA] == [
         "session.created",
         "session.updated",
+        "error",
+        "error",
+        "session.updated",
         "input_audio_buffer.cleared",
         "error",
         EV_COMPLETED,
     ]
     assert events[1]["session"] == events[0]["session"]
+    assert "must be 16000" in events[2]["error"]["message"]
+    assert "must be 16000" in events[3]["error"]["message"]
 
 
 # ref: https://github.com/ggml-org/llama.cpp/issues/30968