diff --git a/e2e_test/chat_completions/test_transcription_tokenspeed.py b/e2e_test/chat_completions/test_transcription_tokenspeed.py
new file mode 100644
index 000000000..bfede7cb5
--- /dev/null
+++ b/e2e_test/chat_completions/test_transcription_tokenspeed.py
@@ -0,0 +1,119 @@
+"""Audio transcription E2E tests — TokenSpeed Qwen3-ASR.
+
+Exercises the ``POST /v1/audio/transcriptions`` endpoint end-to-end against a
+locally-hosted **TokenSpeed** gRPC worker serving ``Qwen/Qwen3-ASR-1.7B``.
+
+This is the only serving path for audio transcription in SMG: the gRPC router's
+transcription adapter (``model_gateway/src/routers/grpc/router.rs``) is Qwen3-ASR
+only, and audio multimodal inputs are accepted only on TokenSpeed workers
+(SGLang / vLLM / TRT-LLM reject audio batches in
+``model_gateway/src/routers/grpc/multimodal/assemble.rs``). The realtime
+WebSocket ASR path (``e2e_test/realtime/test_realtime_local.py``) is a separate
+vLLM-only serving path and does not cover this REST endpoint.
+
+The model is wired through the standard ``setup_backend`` gRPC fixture, so the
+test runs in the ``e2e-1gpu-chat (tokenspeed)`` CI lane — the only lane with
+TokenSpeed installed — and needs no bespoke gateway plumbing. Requests are sent
+with raw ``httpx`` multipart (rather than the OpenAI SDK's transcription
+helper), matching the gateway's ``multipart/form-data`` contract exactly and
+keeping assertions independent of the SDK's streaming/response-format overloads.
+
+Prerequisites:
+- 1 GPU able to serve Qwen3-ASR under TokenSpeed.
+- ``E2E_RUNTIME=tokenspeed`` (set by the CI lane); the ``engine`` marker keeps
+  this test out of the sglang/vllm/trtllm lanes.
+"""
+
+from __future__ import annotations
+
+import logging
+from pathlib import Path
+
+import httpx
+import pytest
+
+logger = logging.getLogger(__name__)
+
+MODEL = "Qwen/Qwen3-ASR-1.7B"
+
+# Reuse the committed 8s / 16 kHz mono PCM16 clip of "Mary had a little lamb"
+# already used by the realtime ASR e2e (no new binary fixture needed).
+AUDIO_WAV = Path(__file__).resolve().parents[1] / "realtime" / "fixtures" / "mary_had_lamb_16k.wav"
+
+# Generous: TokenSpeed model warmup + whole-file ASR decode.
+REQUEST_TIMEOUT = 120.0
+
+
+def _audio_part() -> tuple[str, bytes, str]:
+    """multipart ``file`` part: (filename, bytes, content-type)."""
+    return (AUDIO_WAV.name, AUDIO_WAV.read_bytes(), "audio/wav")
+
+
+def _post_transcription(base_url: str, data: dict[str, str]) -> httpx.Response:
+    """POST the fixture clip to /v1/audio/transcriptions with form fields ``data``."""
+    return httpx.post(
+        f"{base_url}/v1/audio/transcriptions",
+        files={"file": _audio_part()},
+        data={"model": MODEL, **data},
+        timeout=REQUEST_TIMEOUT,
+    )
+
+
+@pytest.fixture()
+def gateway(setup_backend):
+    """The Gateway launched by ``setup_backend`` (its 4th tuple element)."""
+    _, _, _, gw = setup_backend
+    return gw
+
+
+@pytest.mark.e2e
+@pytest.mark.slow
+@pytest.mark.gpu(1)
+@pytest.mark.engine("tokenspeed")
+@pytest.mark.model(MODEL)
+@pytest.mark.gateway(extra_args=["--history-backend", "memory"])
+@pytest.mark.parametrize("setup_backend", ["grpc"], indirect=True)
+class TestTokenSpeedTranscription:
+    """``/v1/audio/transcriptions`` against a TokenSpeed Qwen3-ASR worker."""
+
+    def test_transcription_round_trip(self, gateway):
+        """A whole-file transcription returns HTTP 200 with non-empty ``text``.
+
+        The exact wording is left to the ASR model (it varies by model
+        revision / build), so the transcription is logged and only asserted to
+        be non-empty — matching the realtime e2e's robustness stance. When the
+        model behaves as expected the clip's words show up, so a best-effort
+        substring check is logged without failing the test.
+        """
+        resp = _post_transcription(
+            gateway.base_url,
+            {"language": "en", "response_format": "json", "temperature": "0"},
+        )
+        assert resp.status_code == 200, resp.text
+
+        text = resp.json()["text"]
+        logger.info("transcription (json): %s", text)
+        assert isinstance(text, str)
+        assert text.strip(), "expected a non-empty transcription"
+        if not any(word in text.lower() for word in ("mary", "lamb", "little")):
+            logger.warning("transcription did not contain expected words: %r", text)
+
+    def test_transcription_response_format_text(self, gateway):
+        """``response_format="text"`` yields a plain-text (non-empty) 200 body."""
+        resp = _post_transcription(gateway.base_url, {"response_format": "text"})
+        assert resp.status_code == 200, resp.text
+        assert resp.headers["content-type"].startswith("text/plain")
+
+        text = resp.text
+        logger.info("transcription (text): %s", text)
+        assert text.strip(), "expected a non-empty transcription"
+
+    def test_streaming_transcription_rejected(self, gateway):
+        """TokenSpeed Qwen3-ASR is whole-file only — streaming must be rejected."""
+        resp = _post_transcription(gateway.base_url, {"stream": "true"})
+        assert resp.status_code == 400, resp.text
+
+    def test_unsupported_language_rejected(self, gateway):
+        """An out-of-allow-list language hint is rejected with 400."""
+        resp = _post_transcription(gateway.base_url, {"language": "zz"})
+        assert resp.status_code == 400, resp.text