diff --git a/tests/dfx/perf/tests/test_minicpmo_4_5.json b/tests/dfx/perf/tests/test_minicpmo_4_5.json
index a984c2f32..edc509a81 100644
--- a/tests/dfx/perf/tests/test_minicpmo_4_5.json
+++ b/tests/dfx/perf/tests/test_minicpmo_4_5.json
@@ -41,7 +41,7 @@
         "no_oversample": true,
         "disable_shuffle": true,
         "trust_remote_code": true,
-        "seed_tts_locale": "en",
+        "seed_tts_locale": "zh",
         "extra_body": {
           "modalities": [
             "text",
@@ -126,7 +126,7 @@
         "no_oversample": true,
         "disable_shuffle": true,
         "trust_remote_code": true,
-        "seed_tts_locale": "en",
+        "seed_tts_locale": "zh",
         "extra_body": {
           "modalities": [
             "text",
diff --git a/tests/dfx/perf/tests/test_minicpmo_4_5_duplex_seed_tts.json b/tests/dfx/perf/tests/test_minicpmo_4_5_duplex_seed_tts.json
index 986d0764d..8476628bf 100644
--- a/tests/dfx/perf/tests/test_minicpmo_4_5_duplex_seed_tts.json
+++ b/tests/dfx/perf/tests/test_minicpmo_4_5_duplex_seed_tts.json
@@ -37,7 +37,7 @@
         "no_oversample": true,
         "disable_shuffle": true,
         "trust_remote_code": true,
-        "seed_tts_locale": "en",
+        "seed_tts_locale": "zh",
         "seed_tts_turns_per_session": 4,
         "expected_duplex_audio_turns_per_session": 4,
         "extra_body": {
diff --git a/tests/dfx/perf/tests/test_minicpmo_4_5_ready.json b/tests/dfx/perf/tests/test_minicpmo_4_5_ready.json
index 15d52e6ad..cd919e02f 100644
--- a/tests/dfx/perf/tests/test_minicpmo_4_5_ready.json
+++ b/tests/dfx/perf/tests/test_minicpmo_4_5_ready.json
@@ -39,7 +39,7 @@
         "no_oversample": true,
         "disable_shuffle": true,
         "trust_remote_code": true,
-        "seed_tts_locale": "en",
+        "seed_tts_locale": "zh",
         "extra_body": {
           "modalities": [
             "text",
diff --git a/tests/e2e/accuracy/minicpmo_4_5/test_minicpmo_4_5.py b/tests/e2e/accuracy/minicpmo_4_5/test_minicpmo_4_5.py
index ce14f19fe..1940f77ac 100644
--- a/tests/e2e/accuracy/minicpmo_4_5/test_minicpmo_4_5.py
+++ b/tests/e2e/accuracy/minicpmo_4_5/test_minicpmo_4_5.py
@@ -1,14 +1,17 @@
 # SPDX-License-Identifier: Apache-2.0
 """MiniCPM-o 4.5 Daily-Omni + Seed-TTS + Video-MME accuracy regression coverage.
 
-Daily-Omni settings follow the MiniCPM interleaved AV recipe that reaches
-~78% overall accuracy on Daily-Omni (``minicpm-interleave``, ``temperature=0``,
-text modalities with ``use_tts_template``, and server ``--interleave-mm-strings``
-+ 1fps / 128-frame media-io kwargs, also pinned in ``minicpmo_4_5.yaml``).
+Daily-Omni settings follow the MiniCPM interleaved AV recipe (``minicpm-interleave``,
+``temperature=0``, text modalities with ``use_tts_template``, and server
+``--interleave-mm-strings`` + 1fps / 128-frame media-io kwargs, also pinned in
+``minicpmo_4_5.yaml``). Baseline **79.5**; gate **77.5** (drop ≤ 2pp).
 
 Video-MME settings follow OpenBMB OmniEvalKit MiniCPM ``videomme`` (w/o subs):
 ``minicpm-frames``, ``max_frames=96``, ``temperature=0``, ``output_len=128``.
-Official MiniCPM-o 4.5 reports **70.4**; the gate is **0.68** (~2pp margin).
+Baseline **69.0**; gate **67.0** (drop ≤ 2pp).
+
+Seed-TTS (zh, seed-tts-eval protocol): WER baseline **1.414%** / gate **≤ 1.56%**
+(+10%).
 """
 
 from __future__ import annotations
@@ -26,7 +29,6 @@ from tests.e2e.accuracy.qwen3_omni.qwen3_omni_acc_bench_core import (
     find_vllm_cli,
 )
 from tests.helpers.mark import hardware_test
-from tests.helpers.minicpmo_4_5_duplex import SERVER_PARAMS as DUPLEX_TEST_PARAMS
 from tests.helpers.runtime import OmniServerParams
 from tests.helpers.stage_config import get_deploy_config_path
 
@@ -39,10 +41,11 @@ _RESULT_DIR = Path(
     )
 )
 
-_MIN_DAILY_OMNI_ACCURACY = 0.78
-# MiniCPM-o 4.5 reports 70.4 on Video-MME (w/o subs); ~2pp margin like Daily-Omni.
-_MIN_VIDEOMME_ACCURACY = float(os.environ.get("ACC_BENCH_MIN_VIDEOMME_ACCURACY", "0.68"))
-_MAX_SEED_TTS_MEAN_WER = 0.02
+# Accuracy gates: baseline minus allowed drop (≤2pp / ≤10% WER).
+_MIN_DAILY_OMNI_ACCURACY = 0.775  # 79.5 - 2pp
+_MIN_VIDEOMME_ACCURACY = float(os.environ.get("ACC_BENCH_MIN_VIDEOMME_ACCURACY", "0.67"))
+# jiwer reports a 0-1 ratio; table 1.414 / 1.56 is percent → 0.0156.
+_MAX_SEED_TTS_MEAN_WER = 0.0156
 # Full Seed-TTS zh split (seed-tts-eval/zh/meta.lst) is 2020 rows.
 _SEED_TTS_LOCALE = "zh"
 _SEED_TTS_NUM_PROMPTS = 2020
@@ -291,41 +294,3 @@ def test_minicpmo_4_5_seed_tts_wer_bench(omni_server) -> None:
     )
 
     assert _acc_bench.run_acc_benchmark(_acc_bench.parse_acc_benchmark_args(argv)) == 0
-
-
-@hardware_test(res={"cuda": "H100"}, num_cards=2)
-@pytest.mark.parametrize("omni_server", DUPLEX_TEST_PARAMS, indirect=True)
-def test_minicpmo_4_5_duplex_seed_tts_wer_bench(omni_server) -> None:
-    """Gate Seed-TTS WER through the explicit Realtime TTS contract."""
-    _require_vllm_cli()
-    pytest.importorskip("huggingface_hub")
-
-    argv = build_acc_benchmark_cli_argv(
-        omni_server,
-        skip_seed=False,
-        skip_daily=True,
-        skip_videomme=True,
-        num_prompts=50,
-        max_concurrency=1,
-    )
-    argv.extend(
-        [
-            "--result-dir",
-            str(_RESULT_DIR),
-            "--max-seed-tts-mean-wer",
-            str(_MAX_SEED_TTS_MEAN_WER),
-            "--seed-tts-locale",
-            _SEED_TTS_LOCALE,
-            "--seed-extra-body-json",
-            json.dumps({"save_duplex_request_metrics": True}, separators=(",", ":")),
-            "--seed-backend",
-            "openai-realtime-tts",
-            "--seed-endpoint",
-            "/v1/realtime",
-            "--seed-tts-turns-per-session",
-            "4",
-            "--trust-remote-code",
-        ]
-    )
-
-    assert _acc_bench.run_acc_benchmark(_acc_bench.parse_acc_benchmark_args(argv)) == 0