diff --git a/litellm-rust/crates/core/src/providers/bedrock/audio_transcription.rs b/litellm-rust/crates/core/src/providers/bedrock/audio_transcription.rs
index 9bf1f73a74d..a33670c5113 100644
--- a/litellm-rust/crates/core/src/providers/bedrock/audio_transcription.rs
+++ b/litellm-rust/crates/core/src/providers/bedrock/audio_transcription.rs
@@ -70,6 +70,12 @@ impl AudioTranscriptionProviderConfig for BedrockAudioTranscriptionConfig {
         if let Some(temperature) = optional_params.get("temperature") {
             inference_config.insert("temperature".to_string(), temperature.clone());
         }
+        // Bedrock's Converse API rejects a `system` block on any request that
+        // also carries an audio content block (HTTP 400: "Found system
+        // messages ... and audio chunks ... This is not allowed prior to the
+        // tokenizer version 13."). Voxtral takes the transcription
+        // instruction from the user turn's `text` block above, so no system
+        // prompt is needed here.
         Ok(AudioTranscriptionRequestData {
             body: json!({
                 "messages": [{
@@ -79,7 +85,6 @@ impl AudioTranscriptionProviderConfig for BedrockAudioTranscriptionConfig {
                         {"text": instruction}
                     ]
                 }],
-                "system": [{"text": "You are a transcription assistant."}],
                 "inferenceConfig": inference_config,
             }),
         })
@@ -179,12 +184,33 @@ mod tests {
                         {"text": "Transcribe the audio. Respond with only the transcript. The audio language is en. Additional context: Speaker names"}
                     ]
                 }],
-                "system": [{"text": "You are a transcription assistant."}],
                 "inferenceConfig": {"maxTokens": 4096, "temperature": 0}
             })
         );
     }
 
+    #[test]
+    fn request_omits_system_message_alongside_audio_content() {
+        // Regression test: Bedrock's Converse API 400s on a `system` block
+        // combined with an audio content block ("Found system messages ...
+        // and audio chunks ... This is not allowed prior to the tokenizer
+        // version 13."). Confirmed against live Bedrock that dropping
+        // `system` while keeping everything else identical makes the
+        // request succeed.
+        let result = BEDROCK_AUDIO_TRANSCRIPTION_CONFIG
+            .transform_transcription_request(
+                "mistral.voxtral-mini-3b-2507",
+                json!({"data": "AQI=", "format": "wav", "filename": "sample.wav"}),
+                Map::new(),
+            )
+            .expect("request");
+        assert!(
+            result.body.get("system").is_none(),
+            "request body must not carry a `system` block alongside an audio content block: {}",
+            result.body
+        );
+    }
+
     #[test]
     fn response_concatenates_content_blocks() {
         let result = BEDROCK_AUDIO_TRANSCRIPTION_CONFIG
diff --git a/tests/test_litellm/test_audio_transcription_rust_bridge.py b/tests/test_litellm/test_audio_transcription_rust_bridge.py
index bbeb6c38f78..40ea20c17ef 100644
--- a/tests/test_litellm/test_audio_transcription_rust_bridge.py
+++ b/tests/test_litellm/test_audio_transcription_rust_bridge.py
@@ -149,3 +149,37 @@ async def test_bedrock_atranscription_uses_rust_only_path() -> None:
         rust_bridge.configure_rust_transcription(transcription=None, atranscription=None)
 
     assert response.text == "rust"
+
+
+def test_bedrock_transcription_dispatch_adds_no_system_message() -> None:
+    """Guard the Python/Rust dispatch boundary for the Bedrock Voxtral fix.
+
+    Bedrock's Converse API rejects a request that carries both a `system`
+    block and an audio content block. The fix removes that `system` block
+    from the request the Rust transform builds (see
+    litellm-rust/crates/core/src/providers/bedrock/audio_transcription.rs),
+    since this dispatch layer mocks the whole bridge callable and never sees
+    that internal JSON. This test instead pins the boundary Python does
+    control: it must not hand the bridge a `system` message (directly or via
+    optional_params) that could reintroduce the same conflict from this side.
+    """
+    captured: dict[str, object] = {}
+
+    def capture(**kwargs: object) -> dict[str, object]:
+        captured.update(kwargs)
+        return {"text": "ok"}
+
+    rust_bridge.configure_rust_transcription(transcription=capture, atranscription=None)
+    try:
+        litellm.transcription(
+            model="bedrock/mistral.voxtral-mini-3b-2507",
+            file=("audio.wav", b"audio", "audio/wav"),
+            language="en",
+        )
+    finally:
+        rust_bridge.configure_rust_transcription(transcription=None, atranscription=None)
+
+    assert "system" not in captured
+    optional_params = captured.get("optional_params")
+    assert isinstance(optional_params, dict)
+    assert "system" not in optional_params