Fix Whisper transcription for audio over 30 seconds (#33604)

python/sglang/srt/entrypoints/openai/audio_chunking.py

@@ -0,0 +1,102 @@
+"""Energy-aware audio chunking for ASR models with a bounded input window.
+
+Whisper-style encoders ingest a fixed-length window (30 s of audio = 3000
+mel frames); the feature extractor silently truncates anything longer, so
+long audio must be split into independent chunks before prompting. Cutting
+blindly at the window boundary can land mid-word and corrupt the
+transcription on both sides of the seam, so the splitter searches the tail
+of each window for the quietest stretch (lowest RMS energy) and cuts there.
+
+This matches the chunking behavior of vLLM's
+``OpenAISpeechToText._split_audio`` / ``_find_split_point``: chunks are
+contiguous and non-overlapping — the "search window" is only the region in
+which the cut is allowed to land, and stitching the transcripts back is a
+plain in-order concatenation.
+"""
+
+from __future__ import annotations
+
+import io
+import math
+from typing import List, Tuple
+
+import numpy as np
+import soundfile as sf
+
+from sglang.srt.utils import load_audio
+
+# Region at the tail of each max-length window in which to search for a
+# low-energy split point, in seconds.
+SPLIT_SEARCH_WINDOW_S = 1.0
+
+# RMS energy is evaluated over strides of this many samples (100 ms at
+# 16 kHz); the quietest stride in the search region wins.
+MIN_ENERGY_WINDOW_SIZE = 1600
+
+
+def find_split_point(wav: np.ndarray, start_idx: int, end_idx: int) -> int:
+    """Return the start index of the quietest energy window in
+    ``wav[start_idx:end_idx]``.
+
+    Energy is RMS over consecutive ``MIN_ENERGY_WINDOW_SIZE``-sample
+    strides; the returned index is absolute (into ``wav``). The loop bound
+    intentionally leaves the final stride of the search region unevaluated:
+    it mirrors vLLM's ``_find_split_point`` verbatim so split points stay
+    bit-identical with the old vLLM transcription endpoint.
+    """
+    segment = wav[start_idx:end_idx]
+    min_energy = math.inf
+    quietest_idx = start_idx
+    for i in range(0, len(segment) - MIN_ENERGY_WINDOW_SIZE, MIN_ENERGY_WINDOW_SIZE):
+        window = segment[i : i + MIN_ENERGY_WINDOW_SIZE]
+        energy = (window**2).mean() ** 0.5
+        if energy < min_energy:
+            quietest_idx = i + start_idx
+            min_energy = energy
+    return quietest_idx
+
+
+def split_audio_energy_aware(
+    audio_data: bytes,
+    max_clip_s: float,
+    sample_rate: int = 16000,
+) -> Tuple[List[bytes], List[float]]:
+    """Split audio into WAV chunks no longer than ``max_clip_s`` seconds.
+
+    Decodes (and resamples to ``sample_rate``) the input, walks it in
+    ``max_clip_s`` strides, and cuts each chunk at the lowest-energy point
+    within the final ``SPLIT_SEARCH_WINDOW_S`` of the stride so cuts land
+    on pauses instead of mid-word. Chunks are contiguous and
+    non-overlapping; concatenated they reproduce the full waveform.
+
+    Returns ``(chunk_wav_bytes, chunk_start_offsets_s)`` where
+    ``chunk_start_offsets_s[i]`` is the start time of chunk ``i`` in the
+    original audio.
+    """
+    if not audio_data:
+        raise ValueError("audio_data is empty")
+    audio = load_audio(audio_data, sr=sample_rate, mono=True)
+    chunk_size = int(sample_rate * max_clip_s)
+    search_size = int(sample_rate * SPLIT_SEARCH_WINDOW_S)
+    total = audio.shape[-1]
+
+    raw_chunks: List[np.ndarray] = []
+    offsets_s: List[float] = []
+    i = 0
+    while i < total:
+        offsets_s.append(i / sample_rate)
+        if i + chunk_size >= total:
+            raw_chunks.append(audio[i:])
+            break
+        search_start = i + chunk_size - search_size
+        search_end = min(i + chunk_size, total)
+        split_point = find_split_point(audio, search_start, search_end)
+        raw_chunks.append(audio[i:split_point])
+        i = split_point
+
+    chunks: List[bytes] = []
+    for chunk in raw_chunks:
+        buf = io.BytesIO()
+        sf.write(buf, chunk, sample_rate, format="WAV")
+        chunks.append(buf.getvalue())
+    return chunks, offsets_s

python/sglang/srt/entrypoints/openai/serving_transcription.py

@@ -32,6 +32,7 @@ from typing import TYPE_CHECKING, AsyncGenerator, List, Optional, Union
 from fastapi import Request, WebSocket
 from fastapi.responses import ORJSONResponse, Response, StreamingResponse
 
+from sglang.srt.entrypoints.openai.audio_chunking import split_audio_energy_aware
 from sglang.srt.entrypoints.openai.protocol import (
     DeltaMessage,
     ErrorResponse,
@@ -113,9 +114,17 @@ class OpenAIServingTranscription(OpenAIServingBase):
 
             info = sf.info(io.BytesIO(audio_data))
             return info.duration
-        except Exception as e:
+        except Exception:
-            logger.warning(f"Could not calculate audio duration: {e}")
+            # soundfile can't parse some containers (e.g. mp3 on older
-            return 0.0
+            # libsndfile builds); fall back to a full decode.
+            try:
+                from sglang.srt.utils import load_audio
+
+                audio = load_audio(audio_data, sr=16000, mono=True)
+                return audio.shape[-1] / 16000.0
+            except Exception as e:
+                logger.warning(f"Could not calculate audio duration: {e}")
+                return 0.0
 
     async def create_transcription(
         self,
@@ -135,8 +144,63 @@ class OpenAIServingTranscription(OpenAIServingBase):
         ORJSONResponse,
     ]:
         """Main entry point for transcription requests."""
-        # Calculate audio duration for usage reporting
+        # Calculate audio duration for usage reporting. Run in a thread:
-        audio_duration_s = self._get_audio_duration(audio_data)
+        # the fallback path decodes the full file and would block the
+        # event loop on long inputs.
+        audio_duration_s = await asyncio.to_thread(self._get_audio_duration, audio_data)
+
+        # Audio longer than the model's encoder window (30 s for Whisper)
+        # would be silently truncated by the feature extractor. Split it
+        # into chunks cut at low-energy points (pauses) so the cut never
+        # lands mid-word; each chunk is transcribed as an independent
+        # request and the texts are stitched back in order.
+        audio_chunks: Optional[List[bytes]] = None
+        chunk_offsets_s: Optional[List[float]] = None
+        max_clip_s = self._adapter.max_audio_clip_s
+        if max_clip_s is not None and audio_duration_s > max_clip_s:
+            split_error = (
+                f"Failed to split audio longer than {max_clip_s:g} seconds into "
+                "supported chunks."
+            )
+            try:
+                # In a thread: decodes + re-encodes the whole file, which
+                # would otherwise block the event loop on long inputs.
+                audio_chunks, chunk_offsets_s = await asyncio.to_thread(
+                    split_audio_energy_aware, audio_data, max_clip_s
+                )
+            except Exception as e:
+                logger.warning(
+                    "Failed to split %.1fs audio into chunks of <=%ss: %s",
+                    audio_duration_s,
+                    max_clip_s,
+                    e,
+                )
+                return self.create_error_response(split_error)
+            else:
+                if (
+                    not audio_chunks
+                    or len(audio_chunks) <= 1
+                    or chunk_offsets_s is None
+                    or len(chunk_offsets_s) != len(audio_chunks)
+                ):
+                    logger.error(
+                        "Audio splitter returned an invalid result for %.1fs audio: "
+                        "%d chunks and %s offsets",
+                        audio_duration_s,
+                        len(audio_chunks or []),
+                        (
+                            "no"
+                            if chunk_offsets_s is None
+                            else str(len(chunk_offsets_s))
+                        ),
+                    )
+                    return self.create_error_response(split_error)
+                logger.info(
+                    "Split %.1fs audio into %d chunks of <=%ss for transcription",
+                    audio_duration_s,
+                    len(audio_chunks),
+                    max_clip_s,
+                )
 
         # When language is not specified and the adapter supports detection,
         # use a single fused request: SGLang's structured generation (regex)
@@ -172,6 +236,9 @@ class OpenAIServingTranscription(OpenAIServingBase):
             # selection see the same boolean — and we don't recompute it on
             # every cumulative-text snapshot in streaming.
             request._fused_ts_variant = bool(timestamp_granularities)
+        if audio_chunks is not None and len(audio_chunks) > 1:
+            request._audio_chunks = audio_chunks
+            request._chunk_offsets_s = chunk_offsets_s
 
         # Use the base class handle_request pattern
         return await self.handle_request(request, raw_request)
@@ -189,6 +256,11 @@ class OpenAIServingTranscription(OpenAIServingBase):
         Response,
     ]:
         """Handle non-streaming transcription request."""
+        if getattr(request, "_audio_chunks", None):
+            return await self._handle_chunked_non_streaming_request(
+                adapted_request, request, raw_request
+            )
+
         try:
             ret = await self.tokenizer_manager.generate_request(
                 adapted_request, raw_request
@@ -196,27 +268,7 @@ class OpenAIServingTranscription(OpenAIServingBase):
         except ValueError as e:
             return self.create_error_response(str(e))
 
-        text = self._adapter.postprocess_text(ret.get("text", ""))
+        text = self._finalize_text(request, ret.get("text", ""))
-
-        # For fused auto-detect, parse_fused_output returns the scrubbed
-        # user-visible text. On parse failure (FSM abort, truncation) it
-        # returns (None, None) and we fall back to strip_special_tokens —
-        # the language stays unset rather than reporting a bogus detection.
-        if getattr(request, "_fused_autodetect", False):
-            lang, visible = self._adapter.parse_fused_output(
-                text, ts_variant=getattr(request, "_fused_ts_variant", False)
-            )
-            if visible is None:
-                logger.warning(
-                    "Fused auto-detect parse failed on non-streaming response; "
-                    "falling back to raw-text scrub."
-                )
-                text = self._adapter.strip_special_tokens(text)
-            else:
-                text = visible
-                if lang is not None:
-                    request.language = lang
-                    logger.info("Auto-detected language: '%s'", lang)
 
         usage = TranscriptionUsage(seconds=int(math.ceil(request.audio_duration_s)))
 
@@ -233,6 +285,161 @@ class OpenAIServingTranscription(OpenAIServingBase):
         # Default JSON format
         return TranscriptionResponse(text=text, usage=usage)
 
+    def _finalize_text(
+        self, request: TranscriptionRequest, raw_text: str, strip: bool = True
+    ) -> str:
+        """Postprocess one generation's raw text into user-visible text.
+
+        For fused auto-detect requests, parse_fused_output returns the
+        scrubbed user-visible text and the detected language is recorded on
+        the request (first non-empty parsed chunk wins for chunked audio). On
+        parse failure (FSM abort, truncation) it returns (None, None) and we
+        fall back to a best-effort scrub — the language stays unset rather
+        than reporting a bogus detection.
+
+        ``strip=False`` keeps the model-emitted boundary whitespace in the
+        fused path; chunk stitching needs it as the natural separator.
+        """
+        text = self._adapter.postprocess_text(raw_text)
+        if not getattr(request, "_fused_autodetect", False):
+            return text
+        lang, visible = self._adapter.parse_fused_output(
+            text,
+            ts_variant=getattr(request, "_fused_ts_variant", False),
+            strip=strip,
+        )
+        if visible is None:
+            logger.warning(
+                "Fused auto-detect parse failed on non-streaming response; "
+                "falling back to raw-text scrub."
+            )
+            return self._adapter.strip_special_tokens(text)
+        if lang is not None and visible.strip() and request.language is None:
+            request.language = lang
+            logger.info("Auto-detected language: '%s'", lang)
+        return visible
+
+    def _build_chunk_request(
+        self,
+        adapted_request: GenerateReqInput,
+        chunk_audio: bytes,
+        stream: bool,
+    ) -> GenerateReqInput:
+        """Clone the adapted request for one audio chunk.
+
+        ``sampling_params`` must be a fresh dict per chunk: the multimodal
+        processor pops transcription-level keys (language,
+        timestamp_granularities, the fused-autodetect flag) out of it while
+        building each chunk's decoder prompt.
+        """
+        sampling_params = adapted_request.sampling_params
+        assert isinstance(sampling_params, dict)
+        chunk_request = GenerateReqInput(
+            text="",
+            audio_data=chunk_audio,
+            sampling_params=dict(sampling_params),
+            stream=stream,
+            modalities=["audio"],
+            routing_key=adapted_request.routing_key,
+        )
+        chunk_request.received_time = adapted_request.received_time
+        return chunk_request
+
+    def _abort_chunk_requests(self, chunk_requests: List[GenerateReqInput]) -> None:
+        """Abort chunk generations engine-side.
+
+        The scheduler keeps decoding until a request is aborted by rid (a
+        no-op for chunks that already finished). A request whose rid has not
+        been assigned yet cannot be aborted by the immediate pass, so a second
+        abort fires after the dispatch window — the same approach as
+        ``TokenizerManager.create_abort_task``.
+        """
+
+        def abort_assigned_rids():
+            # Read rids at call time: a chunk task that had not started
+            # executing during the first pass gets its rid assigned later,
+            # so the delayed pass must not reuse an earlier snapshot.
+            for chunk_request in chunk_requests:
+                if isinstance(chunk_request.rid, str):
+                    self.tokenizer_manager.abort_request(chunk_request.rid)
+
+        abort_assigned_rids()
+
+        async def abort_after_dispatch_window():
+            await asyncio.sleep(2)
+            abort_assigned_rids()
+
+        asyncio.create_task(abort_after_dispatch_window())
+
+    async def _handle_chunked_non_streaming_request(
+        self,
+        adapted_request: GenerateReqInput,
+        request: TranscriptionRequest,
+        raw_request: Request,
+    ) -> Union[
+        TranscriptionResponse,
+        TranscriptionVerboseResponse,
+        ErrorResponse,
+        ORJSONResponse,
+        Response,
+    ]:
+        """Transcribe pre-split long audio (duration > max_audio_clip_s).
+
+        Each chunk is an independent generation. Chunks run sequentially so
+        one user-controlled upload cannot fan out into an unbounded number of
+        tokenizer/GPU requests. Results are stitched in chunk (= audio) order.
+        """
+        chunk_requests = [
+            self._build_chunk_request(adapted_request, chunk_audio, stream=False)
+            for chunk_audio in request._audio_chunks
+        ]
+        rets = []
+        try:
+            for chunk_request in chunk_requests:
+                ret = await self.tokenizer_manager.generate_request(
+                    chunk_request, raw_request
+                ).__anext__()
+                rets.append(ret)
+        except BaseException as e:
+            # Abort the in-flight request on failures or parent cancellation.
+            # Later chunks have not been dispatched.
+            self._abort_chunk_requests(chunk_requests)
+            if isinstance(e, ValueError):
+                return self.create_error_response(str(e))
+            raise
+
+        fused = getattr(request, "_fused_autodetect", False)
+        # Plain in-order concatenation: each chunk's model-emitted boundary
+        # whitespace (a leading space for spaced scripts, nothing for
+        # spaceless scripts like zh/ja/th) is the correct seam separator,
+        # so the fused path keeps it (strip=False) instead of inventing
+        # one. Only the full fused text is trimmed at the ends, matching
+        # the single-request fused response.
+        texts = [
+            self._finalize_text(request, ret.get("text", ""), strip=False)
+            for ret in rets
+        ]
+        text = "".join(texts)
+        if fused:
+            text = text.strip()
+
+        usage = TranscriptionUsage(seconds=int(math.ceil(request.audio_duration_s)))
+
+        if request.response_format == "text":
+            return Response(content=text, media_type="text/plain")
+
+        if request.response_format == "verbose_json":
+            return self._adapter.build_verbose_response_chunked(
+                request,
+                text,
+                rets,
+                request._chunk_offsets_s,
+                self.tokenizer_manager.tokenizer,
+                usage,
+            )
+
+        return TranscriptionResponse(text=text, usage=usage)
+
     async def _handle_streaming_request(
         self,
         adapted_request: GenerateReqInput,
@@ -249,6 +456,15 @@ class OpenAIServingTranscription(OpenAIServingBase):
                 ),
                 media_type="text/event-stream",
             )
+        if getattr(request, "_audio_chunks", None):
+            # Long audio pre-split into chunks, transcribed sequentially.
+            # No background abort_task: the in-flight chunk is aborted in
+            # the generator's finally on teardown, and disconnect is checked
+            # between chunks.
+            return StreamingResponse(
+                self._generate_long_audio_stream(adapted_request, request, raw_request),
+                media_type="text/event-stream",
+            )
         return StreamingResponse(
             self._generate_transcription_stream(adapted_request, request, raw_request),
             media_type="text/event-stream",
@@ -326,7 +542,11 @@ class OpenAIServingTranscription(OpenAIServingBase):
                         yield f"data: {error}\n\n"
                         yield "data: [DONE]\n\n"
                         return
-                    if lang is not None and request.language is None:
+                    if (
+                        lang is not None
+                        and visible.strip()
+                        and request.language is None
+                    ):
                         request.language = lang
                         logger.info("Auto-detected language: '%s'", lang)
                 else:
@@ -369,6 +589,136 @@ class OpenAIServingTranscription(OpenAIServingBase):
 
         yield "data: [DONE]\n\n"
 
+    async def _generate_long_audio_stream(
+        self,
+        adapted_request: GenerateReqInput,
+        request: TranscriptionRequest,
+        raw_request: Request,
+    ) -> AsyncGenerator[str, None]:
+        """Stream transcription of long audio pre-split into chunks.
+
+        Chunks are transcribed sequentially (one streaming request at a
+        time, like ``_generate_chunked_asr_stream``), so the client sees
+        the transcript in audio order with a single finish frame after the
+        last chunk. The first abnormal chunk finish_reason (length/abort)
+        wins so a truncated non-final chunk isn't masked by later chunks
+        stopping cleanly. Fused auto-detect applies per chunk; the reported
+        language is the first non-empty chunk's detection. Disconnect is
+        checked between chunks, and the in-flight chunk is aborted on teardown.
+        ``request._chunk_offsets_s`` is unused here — only the non-streaming
+        verbose_json path needs segment timing.
+        """
+        created_time = int(time.time())
+        request_id = f"{self._request_id_prefix()}{uuid.uuid4().hex}"
+        model = request.model
+        fused_mode = getattr(request, "_fused_autodetect", False)
+        ts_variant = getattr(request, "_fused_ts_variant", False)
+        incremental = getattr(
+            self.tokenizer_manager.server_args,
+            "incremental_streaming_output",
+            False,
+        )
+
+        def _frame(delta: Optional[str], finish_reason: Optional[str] = None) -> str:
+            chunk = TranscriptionStreamResponse(
+                id=request_id,
+                created=created_time,
+                model=model,
+                choices=[
+                    TranscriptionStreamChoice(
+                        delta=DeltaMessage(content=delta) if delta else DeltaMessage(),
+                        finish_reason=finish_reason,
+                    )
+                ],
+            )
+            return f"data: {chunk.model_dump_json()}\n\n"
+
+        finish_reason_type = None
+        in_flight: Optional[GenerateReqInput] = None
+        emitted_text = False
+        try:
+            for chunk_audio in request._audio_chunks:
+                if await raw_request.is_disconnected():
+                    break
+                in_flight = self._build_chunk_request(
+                    adapted_request, chunk_audio, stream=True
+                )
+                cumulative_text = ""
+                visible_buffer = ""
+                chunk_finish_reason = None
+                strip_chunk = not emitted_text
+                async for content in self.tokenizer_manager.generate_request(
+                    in_flight, raw_request
+                ):
+                    finish_reason = content["meta_info"]["finish_reason"]
+                    chunk_finish_reason = (
+                        finish_reason["type"] if finish_reason else None
+                    )
+
+                    chunk_text = content.get("text", "")
+                    if incremental:
+                        cumulative_text += chunk_text
+                    else:
+                        cumulative_text = chunk_text
+
+                    if fused_mode:
+                        # Strip the first chunk that emits visible text (no
+                        # leading space at stream start, like the single-request
+                        # path). Later visible chunks keep their model-emitted
+                        # boundary whitespace, which is the correct seam
+                        # separator for spaced and spaceless scripts alike.
+                        lang, visible = self._adapter.parse_fused_output(
+                            cumulative_text,
+                            ts_variant=ts_variant,
+                            strip=strip_chunk,
+                        )
+                        if visible is None:
+                            if not chunk_finish_reason:
+                                continue
+                            logger.warning(
+                                "Fused auto-detect stream finished before prefix "
+                                "was parseable; returning detection-failed error."
+                            )
+                            error = self.create_streaming_error_response(
+                                "language auto-detect failed: forced-prefix "
+                                "sentinel was not produced before stream end"
+                            )
+                            yield f"data: {error}\n\n"
+                            yield "data: [DONE]\n\n"
+                            return
+                        if (
+                            lang is not None
+                            and visible.strip()
+                            and request.language is None
+                        ):
+                            request.language = lang
+                            logger.info("Auto-detected language: '%s'", lang)
+                    else:
+                        visible = cumulative_text
+
+                    delta = visible[len(visible_buffer) :]
+                    visible_buffer = visible
+                    if delta:
+                        yield _frame(delta)
+                        emitted_text = True
+
+                in_flight = None
+                if finish_reason_type in (None, "stop") and chunk_finish_reason:
+                    finish_reason_type = chunk_finish_reason
+
+            yield _frame(None, finish_reason=finish_reason_type or "stop")
+        except ValueError as e:
+            error = self.create_streaming_error_response(str(e))
+            yield f"data: {error}\n\n"
+        finally:
+            # Abort the chunk still decoding if the stream is torn down
+            # (client disconnect / generator close) so it doesn't keep
+            # running in the scheduler. Only one chunk is ever in flight.
+            if in_flight is not None and isinstance(in_flight.rid, str):
+                self.tokenizer_manager.abort_request(in_flight.rid)
+
+        yield "data: [DONE]\n\n"
+
     async def _generate_chunked_asr_stream(
         self,
         adapted_request: GenerateReqInput,

python/sglang/srt/entrypoints/openai/transcription_adapters/base.py

@@ -44,7 +44,7 @@ class TranscriptionAdapter(ABC):
 
     @staticmethod
     def parse_fused_output(
-        text: str, *, ts_variant: bool = False
+        text: str, *, ts_variant: bool = False, strip: bool = True
     ) -> tuple[Optional[str], Optional[str]]:
         """Parse the fused output into ``(language_code, user_visible_text)``.
 
@@ -76,6 +76,17 @@ class TranscriptionAdapter(ABC):
         """
         return text
 
+    @property
+    def max_audio_clip_s(self) -> Optional[float]:
+        """Maximum audio duration (seconds) the model can ingest per request.
+
+        Audio longer than this is split at low-energy points into
+        contiguous chunks (see ``audio_chunking.split_audio_energy_aware``),
+        transcribed as independent requests, and stitched back in order.
+        ``None`` disables chunking.
+        """
+        return None
+
     @property
     def supports_chunked_streaming(self) -> bool:
         """Whether this model uses chunk-based streaming instead of token-level streaming."""
@@ -125,6 +136,31 @@ class TranscriptionAdapter(ABC):
     ) -> TranscriptionVerboseResponse:
         """Build a ``verbose_json`` response with segments / timestamps."""
 
+    def build_verbose_response_chunked(
+        self,
+        request: TranscriptionRequest,
+        text: str,
+        rets: List[dict],
+        chunk_offsets_s: List[float],
+        tokenizer,
+        usage: TranscriptionUsage,
+    ) -> TranscriptionVerboseResponse:
+        """Build a ``verbose_json`` response from multiple chunk results.
+
+        Called instead of ``build_verbose_response`` when the audio was
+        longer than ``max_audio_clip_s`` and split into chunks;
+        ``chunk_offsets_s[i]`` is chunk *i*'s start time in the original
+        audio. The default implementation returns the stitched text with
+        no segment timing.
+        """
+        return TranscriptionVerboseResponse(
+            language=request.language,
+            duration=round(request.audio_duration_s, 2),
+            text=text,
+            segments=[],
+            usage=usage,
+        )
+
 
 _ADAPTER_REGISTRY: dict[str, type[TranscriptionAdapter]] = {}
 _DEFAULT_ADAPTER_KEY = "Whisper"

python/sglang/srt/entrypoints/openai/transcription_adapters/whisper.py

@@ -126,6 +126,13 @@ class WhisperAdapter(TranscriptionAdapter):
     TIMESTAMP_BASE_TOKEN_ID = 50365  # <|0.00|>
     TIMESTAMP_BASE_OFFSET = 0.02  # each token step = 0.02 s
 
+    @property
+    def max_audio_clip_s(self) -> Optional[float]:
+        # Whisper's encoder ingests a fixed 30 s window (3000 mel frames);
+        # the feature extractor truncates anything longer, so longer audio
+        # must be split into chunks before prompting.
+        return 30.0
+
     def build_sampling_params(self, request: TranscriptionRequest) -> dict:
         params: dict = {
             "temperature": request.temperature,
@@ -185,7 +192,7 @@ class WhisperAdapter(TranscriptionAdapter):
 
     @staticmethod
     def parse_fused_output(
-        text: str, *, ts_variant: bool = False
+        text: str, *, ts_variant: bool = False, strip: bool = True
     ) -> tuple[Optional[str], Optional[str]]:
         """Parse fused output into ``(language_code, user_visible_text)``.
 
@@ -209,9 +216,16 @@ class WhisperAdapter(TranscriptionAdapter):
         * ``(lang, visible)`` — prefix fully parsed. ``visible`` is the
           transcription with the forced prefix removed, any embedded
           special tokens (``<|X.XX|>``, ``<|endoftext|>``) scrubbed, and
-          surrounding whitespace trimmed. It grows monotonically across
+          (with ``strip=True``, the default) surrounding whitespace
-          streaming chunks because Whisper's special tokens detokenize
+          trimmed. It grows monotonically across streaming chunks because
-          atomically, so callers can compute deltas against it directly.
+          Whisper's special tokens detokenize atomically, so callers can
+          compute deltas against it directly.
+
+        ``strip=False`` preserves the model-emitted boundary whitespace.
+        Long-audio chunk stitching relies on it: a chunk's own leading
+        space (or its absence, for spaceless scripts like zh/ja/th) is the
+        correct separator at the chunk seam, so an artificial one must not
+        be invented after stripping.
         """
         pattern = _FUSED_PREFIX_RE_TS if ts_variant else _FUSED_PREFIX_RE_NOTS
         m = pattern.match(text)
@@ -225,7 +239,7 @@ class WhisperAdapter(TranscriptionAdapter):
         # are unwanted in the user-visible text (verbose_json gets its
         # segments from _parse_segments over output_ids instead).
         transcription = _SPECIAL_TOKEN_RE.sub("", transcription)
-        return m.group(1), transcription.strip()
+        return m.group(1), transcription.strip() if strip else transcription
 
     @staticmethod
     def strip_special_tokens(text: str) -> str:
@@ -260,9 +274,41 @@ class WhisperAdapter(TranscriptionAdapter):
             usage=usage,
         )
 
+    def build_verbose_response_chunked(
+        self,
+        request: TranscriptionRequest,
+        text: str,
+        rets: List[dict],
+        chunk_offsets_s: List[float],
+        tokenizer,
+        usage: TranscriptionUsage,
+    ) -> TranscriptionVerboseResponse:
+        segments: list[TranscriptionSegment] = []
+        for ret, offset_s in zip(rets, chunk_offsets_s):
+            _, part_segments = self._parse_segments(
+                ret.get("output_ids", []),
+                tokenizer,
+                time_offset_s=offset_s,
+                seg_id_start=len(segments),
+            )
+            segments.extend(part_segments)
+        return TranscriptionVerboseResponse(
+            language=request.language,
+            duration=round(request.audio_duration_s, 2),
+            # The serving layer already stitched model text using each
+            # chunk's emitted boundary whitespace. Reconstructing it from
+            # output_ids with an ASCII join corrupts spaceless scripts.
+            text=text,
+            segments=segments,
+            usage=usage,
+        )
+
     @staticmethod
     def _parse_segments(
-        output_ids: List[int], tokenizer
+        output_ids: List[int],
+        tokenizer,
+        time_offset_s: float = 0.0,
+        seg_id_start: int = 0,
     ) -> tuple[str, List[TranscriptionSegment]]:
         """Parse Whisper timestamp tokens from *output_ids* into segments.
 
@@ -273,6 +319,11 @@ class WhisperAdapter(TranscriptionAdapter):
 
         Each timestamp token marks the end of the current segment; its value
         also becomes the start of the next segment.
+
+        ``time_offset_s`` / ``seg_id_start`` shift segment times and ids for
+        audio that was split into chunks — timestamp tokens are relative to
+        the chunk's own 30 s window, so each chunk's segments are offset by
+        the chunk's start time within the original audio.
         """
         eos_token_id = getattr(tokenizer, "eos_token_id", 50257)
         ts_base = WhisperAdapter.TIMESTAMP_BASE_TOKEN_ID
@@ -281,12 +332,13 @@ class WhisperAdapter(TranscriptionAdapter):
         segments: list[TranscriptionSegment] = []
         full_text_parts: list[str] = []
         current_text_tokens: list[int] = []
-        current_start = 0.0  # First segment starts at 0.0 (from prompt <|0.00|>)
+        # First segment starts at the chunk start (prompt anchors <|0.00|>)
-        seg_id = 0
+        current_start = time_offset_s
+        seg_id = seg_id_start
 
         for token_id in output_ids:
             if token_id >= ts_base:
-                timestamp = (token_id - ts_base) * ts_step
+                timestamp = (token_id - ts_base) * ts_step + time_offset_s
 
                 if current_text_tokens:
                     seg_text = tokenizer.decode(

python/sglang/srt/utils/common.py

@@ -998,7 +998,7 @@ def get_compiler_backend(mode=None) -> str:
     if hasattr(torch, "npu") and torch.npu.is_available():
         try:
             import torchair
-            import torchair.ge_concrete_graph.ge_converter.experimental.patch_for_hcom_allreduce
+            import torchair.ge_concrete_graph.ge_converter.experimental.patch_for_hcom_allreduce  # noqa: F401
             from torchair.configs.compiler_config import CompilerConfig
         except ImportError:
             raise ImportError(
@@ -1687,7 +1687,7 @@ CLIENT_MEDIA_EXCEPTIONS = (
 
 
 def load_audio(
-    audio_file: str, sr: Optional[int] = None, mono: bool = True
+    audio_file: Union[str, bytes], sr: Optional[int] = None, mono: bool = True
 ) -> np.ndarray:
     if sr is None:
         sr = 16000

test/registered/openai_server/basic/test_serving_transcription.py

@@ -10,9 +10,11 @@ import json
 import unittest
 from typing import List, Optional
 
+import numpy as np
 import requests
+import soundfile as sf
 
-from sglang.srt.utils import kill_process_tree
+from sglang.srt.utils import kill_process_tree, load_audio
 from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.test_utils import (
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
@@ -21,7 +23,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=60, stage="base-b", runner_config="1-gpu-small")
+register_cuda_ci(est_time=90, stage="base-b", runner_config="1-gpu-small")
 
 WHISPER_MODEL = "openai/whisper-large-v3"
 AUDIO_URL = "https://raw.githubusercontent.com/sgl-project/sgl-test-files/refs/heads/main/audios/Trump_WEF_2018_10s.mp3"
@@ -34,6 +36,21 @@ def download_audio_bytes(url=AUDIO_URL):
     return response.content
 
 
+def long_audio_wav_bytes(prefix_silence_s: float = 30.0) -> bytes:
+    """A 40 s WAV: silence, then the 10 s speech clip.
+
+    The speech sits entirely past Whisper's 30 s encoder window, so without
+    long-audio chunking the feature extractor silently truncates it away
+    and the transcript contains none of the spoken content.
+    """
+    sr = 16000
+    speech = load_audio(download_audio_bytes(), sr=sr, mono=True).astype(np.float32)
+    wav = np.concatenate([np.zeros(int(prefix_silence_s * sr), np.float32), speech])
+    buf = io.BytesIO()
+    sf.write(buf, wav, sr, format="WAV")
+    return buf.getvalue()
+
+
 class TestServingTranscription(CustomTestCase):
     """Test Whisper transcription via /v1/audio/transcriptions endpoint."""
 
@@ -61,13 +78,15 @@ class TestServingTranscription(CustomTestCase):
         language: Optional[str] = "en",
         response_format: Optional[str] = None,
         timestamp_granularities: Optional[List[str]] = None,
+        audio_bytes: Optional[bytes] = None,
     ):
         """Send a non-streaming transcription request and return the JSON response.
 
         Passing ``language=None`` omits the field entirely, which exercises
         the fused auto-detect path.
         """
-        audio_bytes = download_audio_bytes()
+        if audio_bytes is None:
+            audio_bytes = download_audio_bytes()
         data = {"model": "whisper"}
         if language is not None:
             data["language"] = language
@@ -84,9 +103,14 @@ class TestServingTranscription(CustomTestCase):
         self.assertEqual(response.status_code, 200, response.text)
         return response.json()
 
-    def _transcribe_stream(self, language: Optional[str] = None) -> List[str]:
+    def _transcribe_stream(
+        self,
+        language: Optional[str] = None,
+        audio_bytes: Optional[bytes] = None,
+    ) -> List[str]:
         """Send a streaming transcription request and return the delta strings."""
-        audio_bytes = download_audio_bytes()
+        if audio_bytes is None:
+            audio_bytes = download_audio_bytes()
         data = {"model": "whisper", "stream": "true"}
         if language is not None:
             data["language"] = language
@@ -225,6 +249,56 @@ class TestServingTranscription(CustomTestCase):
             "Streamed auto-detect text should match the non-streaming result.",
         )
 
+    # -- long audio (> 30 s encoder window) --------------------------------
+    # 30 s of silence followed by the 10 s speech clip: every spoken word is
+    # past Whisper's encoder window, so these tests fail outright unless the
+    # server splits long audio into chunks and stitches the transcripts.
+
+    KEYWORDS = ["privilege", "leader", "science", "art"]
+
+    def _assert_keywords(self, text: str):
+        matches = [kw for kw in self.KEYWORDS if kw in text.lower()]
+        self.assertGreaterEqual(
+            len(matches),
+            2,
+            f"Expected at least 2 of {self.KEYWORDS}, found {matches}. "
+            f"Full text: {text!r}",
+        )
+
+    def test_long_audio_transcribes_past_30s(self):
+        """Speech past the 30 s window must appear in the transcript."""
+        result = self._transcribe(audio_bytes=long_audio_wav_bytes())
+        self._assert_keywords(result["text"])
+        # Usage reports the full audio duration, not one chunk's.
+        self.assertGreaterEqual(result["usage"]["seconds"], 40)
+
+    def test_long_audio_verbose_json_segment_offsets(self):
+        """Segment timestamps are offset by each chunk's start time."""
+        result = self._transcribe(
+            response_format="verbose_json",
+            timestamp_granularities=["segment"],
+            audio_bytes=long_audio_wav_bytes(),
+        )
+        self._assert_keywords(result.get("text", ""))
+        segments = result.get("segments") or []
+        self.assertGreater(len(segments), 0, "Expected at least one segment")
+        # The speech starts at t=30 s; its segments must be reported in
+        # original-audio time, which is unreachable within a single 30 s
+        # window.
+        self.assertGreater(
+            max(seg["end"] for seg in segments),
+            30.0,
+            f"Expected segment timing past the 30 s window, got {segments!r}",
+        )
+
+    def test_long_audio_streaming(self):
+        """Streaming long audio emits the post-30 s content as deltas."""
+        deltas = self._transcribe_stream(
+            language="en", audio_bytes=long_audio_wav_bytes()
+        )
+        self.assertTrue(len(deltas) > 0, "Expected at least one streamed delta")
+        self._assert_keywords("".join(deltas))
+
 
 if __name__ == "__main__":
     unittest.main()

test/registered/unit/entrypoints/openai/test_audio_chunking.py

@@ -0,0 +1,122 @@
+"""Unit tests for energy-aware audio chunking (long-audio transcription).
+
+Whisper's encoder window is 30 s; longer audio must be split before
+prompting. The splitter must cut at low-energy points (pauses) inside the
+tail search window of each stride — never blindly at the stride boundary,
+which could land mid-word — and the chunks must be contiguous,
+non-overlapping, and reproduce the full waveform when concatenated.
+"""
+
+from sglang.test.test_utils import maybe_stub_sgl_kernel
+
+maybe_stub_sgl_kernel()  # must precede any import that pulls in sgl_kernel
+
+import io
+import unittest
+
+import numpy as np
+import soundfile as sf
+
+from sglang.srt.entrypoints.openai.audio_chunking import (
+    find_split_point,
+    split_audio_energy_aware,
+)
+from sglang.test.ci.ci_register import register_cpu_ci
+from sglang.test.test_utils import CustomTestCase
+
+register_cpu_ci(est_time=4, suite="base-a-test-cpu")
+
+SR = 16000
+
+
+def _tone_with_silences(duration_s: float, silences: list, sr: int = SR) -> np.ndarray:
+    """A 440 Hz tone with zeroed-out gaps at the given (start_s, end_s) spans."""
+    t = np.arange(int(duration_s * sr)) / sr
+    wav = (0.5 * np.sin(2 * np.pi * 440.0 * t)).astype(np.float32)
+    for start_s, end_s in silences:
+        wav[int(start_s * sr) : int(end_s * sr)] = 0.0
+    return wav
+
+
+def _wav_bytes(wav: np.ndarray, sr: int = SR) -> bytes:
+    buf = io.BytesIO()
+    sf.write(buf, wav, sr, format="WAV")
+    return buf.getvalue()
+
+
+def _decode(chunk: bytes) -> np.ndarray:
+    data, sr = sf.read(io.BytesIO(chunk), dtype="float32")
+    assert sr == SR
+    return data
+
+
+class TestFindSplitPoint(CustomTestCase):
+    def test_picks_quietest_window(self):
+        # Loud tone with one silent 200 ms gap; the split point must land
+        # inside the gap.
+        wav = _tone_with_silences(4.0, [(2.0, 2.2)])
+        idx = find_split_point(wav, int(1.0 * SR), int(3.0 * SR))
+        self.assertGreaterEqual(idx, int(2.0 * SR))
+        self.assertLess(idx, int(2.2 * SR))
+
+    def test_uniform_energy_returns_search_start(self):
+        wav = _tone_with_silences(4.0, [])
+        idx = find_split_point(wav, int(1.0 * SR), int(3.0 * SR))
+        # All windows are equally loud; the first one wins.
+        self.assertEqual(idx, int(1.0 * SR))
+
+
+class TestSplitAudioEnergyAware(CustomTestCase):
+    def test_short_audio_single_chunk(self):
+        wav = _tone_with_silences(10.0, [])
+        chunks, offsets = split_audio_energy_aware(_wav_bytes(wav), max_clip_s=30.0)
+        self.assertEqual(len(chunks), 1)
+        self.assertEqual(offsets, [0.0])
+        self.assertEqual(len(_decode(chunks[0])), len(wav))
+
+    def test_long_audio_splits_at_silence(self):
+        # 70 s tone with silence gaps planted inside each stride's search
+        # window (the last 1 s before the 30 s boundary): the first cut must
+        # land in [29.2, 29.5], which puts the second search window at
+        # ~[58.2, 59.2] where the second gap [58.5, 58.8] lives.
+        gaps = [(29.2, 29.5), (58.5, 58.8)]
+        wav = _tone_with_silences(70.0, gaps)
+        chunks, offsets = split_audio_energy_aware(_wav_bytes(wav), max_clip_s=30.0)
+
+        self.assertEqual(len(chunks), 3)
+        self.assertEqual(offsets[0], 0.0)
+        # Each cut lands inside its silence gap, not at the blind 30 s mark.
+        for offset, (gap_start, gap_end) in zip(offsets[1:], gaps):
+            self.assertGreaterEqual(offset, gap_start)
+            self.assertLess(offset, gap_end)
+
+        decoded = [_decode(c) for c in chunks]
+        # No chunk exceeds the model's window.
+        for d in decoded:
+            self.assertLessEqual(len(d), 30 * SR)
+        # Chunks are contiguous and non-overlapping: offsets line up with
+        # cumulative chunk lengths and the total sample count is preserved.
+        cumulative = 0
+        for d, offset in zip(decoded, offsets):
+            self.assertEqual(cumulative, int(round(offset * SR)))
+            cumulative += len(d)
+        self.assertEqual(cumulative, len(wav))
+        # Concatenation reproduces the waveform (modulo PCM16 quantization).
+        stitched = np.concatenate(decoded)
+        np.testing.assert_allclose(stitched, wav, atol=2.0 / 32768)
+
+    def test_no_silence_still_makes_progress(self):
+        # A constant-energy tone has no preferred split point; the splitter
+        # must still terminate with bounded chunks covering all samples.
+        wav = _tone_with_silences(65.0, [])
+        chunks, offsets = split_audio_energy_aware(_wav_bytes(wav), max_clip_s=30.0)
+        self.assertGreaterEqual(len(chunks), 3)
+        decoded = [_decode(c) for c in chunks]
+        for d in decoded:
+            self.assertLessEqual(len(d), 30 * SR)
+        self.assertEqual(sum(len(d) for d in decoded), len(wav))
+        self.assertEqual(len(offsets), len(chunks))
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/unit/entrypoints/openai/test_serving_transcription.py

@@ -14,12 +14,20 @@ from sglang.test.test_utils import maybe_stub_sgl_kernel
 
 maybe_stub_sgl_kernel()  # must precede any import that pulls in sgl_kernel
 
+import asyncio
+import io
 import json
 import unittest
 from typing import List
-from unittest.mock import Mock
+from unittest.mock import AsyncMock, Mock, patch
 
-from sglang.srt.entrypoints.openai.protocol import TranscriptionRequest
+import numpy as np
+import soundfile as sf
+
+from sglang.srt.entrypoints.openai.protocol import (
+    TranscriptionRequest,
+    TranscriptionResponse,
+)
 from sglang.srt.entrypoints.openai.serving_transcription import (
     OpenAIServingTranscription,
 )
@@ -241,6 +249,426 @@ class TestStreamingFusedAutodetect(CustomTestCase):
         self.assertFalse(any("<|" in d for d in deltas))
 
 
+class _MockChunkTokenizerManager:
+    """Mock TM scripting one result list per dispatched request, in order."""
+
+    def __init__(self, results_per_request: List):
+        self.model_config = Mock()
+        self.model_config.hf_config = Mock()
+        self.model_config.hf_config.architectures = ["WhisperForConditionalGeneration"]
+        self.server_args = Mock(
+            incremental_streaming_output=False,
+            asr_max_concurrent_sessions=32,
+        )
+        self.request_logger = Mock(log_requests=False)
+        self.tokenizer = Mock()
+        self.requests: List[GenerateReqInput] = []
+        self.aborted: List[str] = []
+        self._results = results_per_request
+        self.active_dispatches = 0
+        self.max_active_dispatches = 0
+
+    def generate_request(self, adapted_request, raw_request):
+        idx = len(self.requests)
+        # Mimic the real generate_request assigning a rid (via
+        # normalize_batch_and_arguments) so abort-by-rid is exercisable.
+        adapted_request.rid = f"rid{idx}"
+        self.requests.append(adapted_request)
+        results = self._results[idx]
+
+        async def gen():
+            self.active_dispatches += 1
+            self.max_active_dispatches = max(
+                self.max_active_dispatches, self.active_dispatches
+            )
+            # Give concurrently scheduled generators a chance to overlap at
+            # dispatch, then record that this request reached the engine.
+            await asyncio.sleep(0)
+            self.active_dispatches -= 1
+            for r in results:
+                # A bare exception entry simulates a chunk request failing.
+                if isinstance(r, BaseException):
+                    raise r
+                yield r
+
+        return gen()
+
+    def abort_request(self, rid: str = "", abort_all: bool = False):
+        self.aborted.append(rid)
+
+    def create_abort_task(self, adapted_request):
+        return None
+
+
+def _long_wav_bytes(duration_s: float = 65.0) -> bytes:
+    """A 16 kHz tone with silence gaps inside each 30 s stride's split-search
+    window, so the energy-aware splitter cuts at ~29.2 s and ~58.5 s."""
+    sr = 16000
+    t = np.arange(int(duration_s * sr)) / sr
+    wav = (0.5 * np.sin(2 * np.pi * 440.0 * t)).astype(np.float32)
+    for gap_start, gap_end in ((29.2, 29.5), (58.5, 58.8)):
+        wav[int(gap_start * sr) : int(gap_end * sr)] = 0.0
+    buf = io.BytesIO()
+    sf.write(buf, wav, sr, format="WAV")
+    return buf.getvalue()
+
+
+class TestLongAudioChunkedNonStreaming(CustomTestCase):
+    """Audio longer than Whisper's 30 s window must be split into chunk
+    requests and the transcripts stitched in order — without chunking the
+    feature extractor silently truncates everything past 30 s."""
+
+    def _create_transcription(self, tm, audio_bytes, language="en", **kwargs):
+        serving = OpenAIServingTranscription(tm)
+        loop = get_or_create_event_loop()
+        return loop.run_until_complete(
+            serving.create_transcription(
+                audio_data=audio_bytes,
+                model="whisper",
+                language=language,
+                response_format=kwargs.pop("response_format", "json"),
+                temperature=0.0,
+                stream=False,
+                raw_request=Mock(),
+                **kwargs,
+            )
+        )
+
+    def test_long_audio_is_chunked_and_stitched(self):
+        texts = [" part one.", " part two.", " part three."]
+        tm = _MockChunkTokenizerManager(
+            [
+                [{"text": t, "meta_info": {"finish_reason": {"type": "stop"}}}]
+                for t in texts
+            ]
+        )
+        result = self._create_transcription(tm, _long_wav_bytes(65.0))
+
+        self.assertIsInstance(result, TranscriptionResponse, result)
+        # In-order plain concatenation (vLLM-parity stitching).
+        self.assertEqual(result.text, " part one. part two. part three.")
+        self.assertEqual(result.usage.seconds, 65)
+        self.assertEqual(len(tm.requests), 3)
+        self.assertEqual(tm.max_active_dispatches, 1)
+
+        # Each chunk request is independent: own audio payload, own
+        # sampling_params dict (the multimodal processor pops keys out of
+        # it per request), stream=False, audio modality.
+        params_ids = {id(req.sampling_params) for req in tm.requests}
+        self.assertEqual(len(params_ids), len(tm.requests))
+        total_samples = 0
+        for req in tm.requests:
+            self.assertFalse(req.stream)
+            self.assertEqual(req.modalities, ["audio"])
+            data, sr = sf.read(io.BytesIO(req.audio_data), dtype="float32")
+            self.assertEqual(sr, 16000)
+            self.assertLessEqual(len(data), 30 * 16000)
+            total_samples += len(data)
+        self.assertEqual(total_samples, 65 * 16000)
+
+    def test_chunk_failure_returns_error_and_stops_dispatch(self):
+        # When one chunk request fails, create_transcription must return an
+        # error response (not a partial transcript), abort the in-flight
+        # request, and leave later chunks undispatched.
+        tm = _MockChunkTokenizerManager(
+            [
+                [
+                    {
+                        "text": " part one.",
+                        "meta_info": {"finish_reason": {"type": "stop"}},
+                    }
+                ],
+                [ValueError("chunk boom")],
+                [
+                    {
+                        "text": " part three.",
+                        "meta_info": {"finish_reason": {"type": "stop"}},
+                    }
+                ],
+            ]
+        )
+        result = self._create_transcription(tm, _long_wav_bytes(65.0))
+        # Error response, not a TranscriptionResponse.
+        self.assertNotIsInstance(result, TranscriptionResponse)
+        # Sequential dispatch means the third chunk never reaches the engine.
+        self.assertEqual(len(tm.requests), 2)
+        self.assertIn(tm.requests[-1].rid, tm.aborted)
+        self.assertEqual(tm.max_active_dispatches, 1)
+
+    def test_split_failure_returns_error_without_dispatch(self):
+        tm = _MockChunkTokenizerManager([])
+        with patch(
+            "sglang.srt.entrypoints.openai.serving_transcription."
+            "split_audio_energy_aware",
+            side_effect=RuntimeError("decode failed"),
+        ):
+            result = self._create_transcription(tm, _long_wav_bytes(65.0))
+
+        self.assertEqual(result.status_code, 400)
+        self.assertIn("Failed to split audio", json.loads(result.body)["message"])
+        self.assertEqual(tm.requests, [])
+
+    def test_short_audio_stays_unchunked(self):
+        tm = _MockChunkTokenizerManager(
+            [[{"text": " short.", "meta_info": {"finish_reason": {"type": "stop"}}}]]
+        )
+        result = self._create_transcription(tm, _long_wav_bytes(10.0))
+        self.assertIsInstance(result, TranscriptionResponse, result)
+        self.assertEqual(result.text, " short.")
+        self.assertEqual(len(tm.requests), 1)
+
+    def test_chunked_fused_autodetect_first_chunk_language_wins(self):
+        # language=None → fused auto-detect per chunk. Each chunk carries
+        # its own forced prefix; the stitched text must strip all of them,
+        # and the reported language comes from the first chunk.
+        tm = _MockChunkTokenizerManager(
+            [
+                [
+                    {
+                        "text": "<|en|><|transcribe|><|notimestamps|> part one.",
+                        "meta_info": {"finish_reason": {"type": "stop"}},
+                    }
+                ],
+                [
+                    {
+                        "text": "<|fr|><|transcribe|><|notimestamps|> part two.",
+                        "meta_info": {"finish_reason": {"type": "stop"}},
+                    }
+                ],
+                [
+                    {
+                        "text": "<|en|><|transcribe|><|notimestamps|> part three.",
+                        "meta_info": {"finish_reason": {"type": "stop"}},
+                    }
+                ],
+            ]
+        )
+        result = self._create_transcription(tm, _long_wav_bytes(65.0), language=None)
+        self.assertIsInstance(result, TranscriptionResponse, result)
+        self.assertEqual(result.text, "part one. part two. part three.")
+        self.assertNotIn("<|", result.text)
+        self.assertEqual(len(tm.requests), 3)
+        # Every chunk request kept the fused regex constraint.
+        for req in tm.requests:
+            self.assertIn("regex", req.sampling_params)
+
+    def test_chunked_fused_spaceless_script_not_space_joined(self):
+        # zh/ja/th chunk texts carry no boundary whitespace; stitching must
+        # not inject an ASCII space the model never emitted.
+        tm = _MockChunkTokenizerManager(
+            [
+                [
+                    {
+                        "text": "<|zh|><|transcribe|><|notimestamps|>你好",
+                        "meta_info": {"finish_reason": {"type": "stop"}},
+                    }
+                ],
+                [
+                    {
+                        "text": "<|zh|><|transcribe|><|notimestamps|>世界",
+                        "meta_info": {"finish_reason": {"type": "stop"}},
+                    }
+                ],
+            ]
+        )
+        result = self._create_transcription(tm, _long_wav_bytes(40.0), language=None)
+        self.assertIsInstance(result, TranscriptionResponse, result)
+        self.assertEqual(result.text, "你好世界")
+        self.assertEqual(len(tm.requests), 2)
+
+    def test_chunked_fused_language_uses_first_nonempty_chunk(self):
+        tm = _MockChunkTokenizerManager(
+            [
+                [
+                    {
+                        "text": "<|fr|><|transcribe|><|notimestamps|>",
+                        "output_ids": [],
+                        "meta_info": {"finish_reason": {"type": "stop"}},
+                    }
+                ],
+                [
+                    {
+                        "text": "<|en|><|transcribe|><|notimestamps|> Hello",
+                        "output_ids": [],
+                        "meta_info": {"finish_reason": {"type": "stop"}},
+                    }
+                ],
+            ]
+        )
+        result = self._create_transcription(
+            tm,
+            _long_wav_bytes(40.0),
+            language=None,
+            response_format="verbose_json",
+        )
+
+        self.assertEqual(result.text, "Hello")
+        self.assertEqual(result.language, "en")
+
+
+class TestLongAudioChunkedStreaming(CustomTestCase):
+    """_generate_long_audio_stream: chunks transcribed sequentially, deltas
+    emitted in audio order, exactly one finish frame."""
+
+    def _run_stream(self, results_per_request, fused=False, n_chunks=2):
+        tm = _MockChunkTokenizerManager(results_per_request)
+        serving = OpenAIServingTranscription(tm)
+        request = TranscriptionRequest(model="whisper", stream=True)
+        if fused:
+            request._fused_autodetect = True
+            request._fused_ts_variant = False
+        request._audio_chunks = [b"chunk%d" % i for i in range(n_chunks)]
+        # Streaming has no segment timing, so the offsets are intentionally
+        # not set here — only the non-streaming verbose_json path reads them.
+        adapted = GenerateReqInput(
+            text="", modalities=["audio"], sampling_params={"temperature": 0.0}
+        )
+        raw_request = Mock()
+        raw_request.is_disconnected = AsyncMock(return_value=False)
+
+        async def drive():
+            frames = []
+            async for frame in serving._generate_long_audio_stream(
+                adapted, request, raw_request
+            ):
+                frames.append(frame)
+            return frames
+
+        loop = get_or_create_event_loop()
+        return tm, request, loop.run_until_complete(drive())
+
+    @staticmethod
+    def _finish_reasons(frames: List[str]) -> List[str]:
+        out = []
+        for line in frames:
+            if not line.startswith("data: ") or line.strip() == "data: [DONE]":
+                continue
+            obj = json.loads(line[len("data: ") :])
+            for choice in obj.get("choices", []):
+                if choice.get("finish_reason"):
+                    out.append(choice["finish_reason"])
+        return out
+
+    def test_chunks_streamed_in_order_with_single_finish(self):
+        tm, _, frames = self._run_stream(
+            [
+                [_chunk(" Hello"), _chunk(" Hello world", finish="stop")],
+                [_chunk(" Again"), _chunk(" Again done", finish="stop")],
+            ]
+        )
+        self.assertEqual(
+            _deltas_from_sse(frames), [" Hello", " world", " Again", " done"]
+        )
+        self.assertEqual(self._finish_reasons(frames), ["stop"])
+        self.assertEqual(frames[-1], "data: [DONE]\n\n")
+        # Both chunk requests were dispatched and streamed.
+        self.assertEqual(len(tm.requests), 2)
+        self.assertTrue(all(req.stream for req in tm.requests))
+
+    def test_disconnect_between_chunks_stops_and_aborts(self):
+        # Client disconnects after the first chunk: the second chunk is
+        # never dispatched, and the in-flight request from chunk 0 is
+        # already done so nothing is left decoding.
+        tm = _MockChunkTokenizerManager(
+            [
+                [_chunk(" Hello", finish="stop")],
+                [_chunk(" world", finish="stop")],
+            ]
+        )
+        serving = OpenAIServingTranscription(tm)
+        request = TranscriptionRequest(model="whisper", stream=True)
+        request._audio_chunks = [b"chunk0", b"chunk1"]
+        adapted = GenerateReqInput(
+            text="", modalities=["audio"], sampling_params={"temperature": 0.0}
+        )
+        raw_request = Mock()
+        # Connected for the first chunk, disconnected before the second.
+        raw_request.is_disconnected = AsyncMock(side_effect=[False, True])
+
+        async def drive():
+            return [
+                f
+                async for f in serving._generate_long_audio_stream(
+                    adapted, request, raw_request
+                )
+            ]
+
+        frames = get_or_create_event_loop().run_until_complete(drive())
+        self.assertEqual(_deltas_from_sse(frames), [" Hello"])
+        # Only the first chunk was dispatched.
+        self.assertEqual(len(tm.requests), 1)
+
+    def test_abnormal_chunk_finish_reason_not_masked(self):
+        # A non-final chunk truncated at the token cap (finish="length")
+        # must surface in the single final frame even though later chunks
+        # stop cleanly — otherwise silently missing transcript content
+        # reads as success.
+        _, _, frames = self._run_stream(
+            [
+                [_chunk(" Hello", finish="length")],
+                [_chunk(" world", finish="stop")],
+            ]
+        )
+        self.assertEqual(_deltas_from_sse(frames), [" Hello", " world"])
+        self.assertEqual(self._finish_reasons(frames), ["length"])
+
+    def test_fused_chunks_strip_prefixes_and_preserve_boundary_space(self):
+        tm, request, frames = self._run_stream(
+            [
+                [
+                    _chunk("<|fr|><|transcribe|>"),
+                    _chunk(
+                        "<|fr|><|transcribe|><|notimestamps|> Bonjour", finish="stop"
+                    ),
+                ],
+                [
+                    _chunk(
+                        "<|fr|><|transcribe|><|notimestamps|> le monde",
+                        finish="stop",
+                    )
+                ],
+            ],
+            fused=True,
+        )
+        deltas = _deltas_from_sse(frames)
+        self.assertFalse(any("<|" in d for d in deltas))
+        # No leading space at stream start; the later chunk's own leading
+        # space is the seam separator.
+        self.assertFalse(deltas[0].startswith(" "))
+        self.assertEqual("".join(deltas), "Bonjour le monde")
+        self.assertEqual(request.language, "fr")
+        self.assertEqual(self._finish_reasons(frames), ["stop"])
+
+    def test_fused_spaceless_script_chunks_not_space_joined(self):
+        # zh/ja/th transcripts carry no boundary whitespace; the seam must
+        # not inject an ASCII space the model never emitted.
+        _, request, frames = self._run_stream(
+            [
+                [_chunk("<|zh|><|transcribe|><|notimestamps|>你好", finish="stop")],
+                [_chunk("<|zh|><|transcribe|><|notimestamps|>世界", finish="stop")],
+            ],
+            fused=True,
+        )
+        self.assertEqual("".join(_deltas_from_sse(frames)), "你好世界")
+        self.assertEqual(request.language, "zh")
+
+    def test_fused_leading_silence_uses_first_nonempty_chunk_language(self):
+        _, request, frames = self._run_stream(
+            [
+                [_chunk("<|fr|><|transcribe|><|notimestamps|>", finish="stop")],
+                [
+                    _chunk(
+                        "<|en|><|transcribe|><|notimestamps|> Hello",
+                        finish="stop",
+                    )
+                ],
+            ],
+            fused=True,
+        )
+        self.assertEqual("".join(_deltas_from_sse(frames)), "Hello")
+        self.assertEqual(request.language, "en")
+
+
 class TestStreamingIncrementalOutputMode(CustomTestCase):
     """Server runs with ``incremental_streaming_output=True``.
 

test/registered/unit/entrypoints/openai/test_whisper_adapter.py

@@ -314,5 +314,74 @@ class TestWhisperBuildFusedAutodetectParams(CustomTestCase):
         SamplingParams(**params)
 
 
+class _FakeTokenizer:
+    """Decodes token ids to a deterministic placeholder string."""
+
+    eos_token_id = 50257
+
+    def decode(self, ids, skip_special_tokens=True):
+        return " " + " ".join(f"tok{i}" for i in ids)
+
+
+class TestWhisperChunkedVerboseResponse(CustomTestCase):
+    """build_verbose_response_chunked for audio split into >30 s chunks.
+
+    Whisper timestamp tokens are relative to each chunk's own 30 s window,
+    so every chunk's segments must be shifted by the chunk's start offset
+    in the original audio and segment ids must keep counting across chunks.
+    """
+
+    TS = WhisperAdapter.TIMESTAMP_BASE_TOKEN_ID  # <|0.00|>
+
+    def test_segments_offset_by_chunk_start(self):
+        request = TranscriptionRequest(
+            model="whisper", language="en", audio_duration_s=57.06
+        )
+        from sglang.srt.entrypoints.openai.protocol import TranscriptionUsage
+
+        usage = TranscriptionUsage(seconds=58)
+        rets = [
+            # chunk 0: one closed segment [0.00 → 5.00]
+            {"output_ids": [100, 101, self.TS + 250]},
+            # chunk 1: closed segment [0.00 → 2.00] + trailing unclosed text
+            {"output_ids": [200, self.TS + 100, 300]},
+        ]
+        resp = WhisperAdapter().build_verbose_response_chunked(
+            request, "你好世界", rets, [0.0, 29.3], _FakeTokenizer(), usage
+        )
+
+        self.assertEqual(resp.language, "en")
+        self.assertEqual(resp.duration, 57.06)
+        self.assertEqual(resp.text, "你好世界")
+        self.assertEqual([s.id for s in resp.segments], [0, 1, 2])
+        self.assertEqual(
+            [(s.start, s.end) for s in resp.segments],
+            [(0.0, 5.0), (29.3, 31.3), (31.3, 31.3)],
+        )
+
+    def test_single_chunk_at_zero_offset_matches_unchunked(self):
+        request = TranscriptionRequest(
+            model="whisper", language="en", audio_duration_s=10.0
+        )
+        from sglang.srt.entrypoints.openai.protocol import TranscriptionUsage
+
+        usage = TranscriptionUsage(seconds=10)
+        output_ids = [100, self.TS + 50]
+        text, segments = WhisperAdapter._parse_segments(output_ids, _FakeTokenizer())
+        chunked = WhisperAdapter().build_verbose_response_chunked(
+            request,
+            text,
+            [{"output_ids": output_ids}],
+            [0.0],
+            _FakeTokenizer(),
+            usage,
+        )
+        self.assertEqual(chunked.text, text)
+        self.assertEqual(
+            [(s.id, s.start, s.end) for s in chunked.segments],
+            [(s.id, s.start, s.end) for s in segments],
+        )
+
+
 if __name__ == "__main__":
     unittest.main()