🔥(backend) remove the S3 storage-event webhook for recordings

Recordings used to be finalized by an inbound notifications sent by
S3. This tied the recording lifecycle to bucket notifications, adding
complexity and dependency to limited S3 services.

The LiveKit egress_ended webhook, added as a fallback in aee1847, does
the same job without any object-storage dependency. Make it the only
mechanism: RecordingEventsService.handle_complete is now called on
EGRESS_COMPLETE / EGRESS_LIMIT_REACHED unconditionally, instead of only
when RECORDING_STORAGE_EVENT_ENABLE is False. Remove the storage-event
path entirely.

CHANGELOG.md

@@ -8,15 +8,10 @@ and this project adheres to
 
 ## [Unreleased]
 
-### Added
-
-- ✨(agent) support Voxtral realtime as inference engine
-
 ### Changed
 
 - 🔥(backend) remove the S3 storage-event webhook for recordings
 - ♻️(backend) always finalize recordings using the LiveKit egress_ended webhook
-- ⬆️(frontend) upgrade posthog-js from 1.409.5 to 1.414.0
 
 ## [1.29.0] - 2026-08-25
 

compose.yml

@@ -251,7 +251,6 @@ services:
     build:
       context: ./src/agents
       target: development
-    command: ["python", "multi_user_transcriber.py", "dev"]
     env_file:
       - env.d/development/multi_user_transcriber
     volumes:

env.d/development/multi_user_transcriber.dist

@@ -1,23 +1,14 @@
-AWS_S3_ENDPOINT_URL=minio:9000
-AWS_S3_ACCESS_KEY_ID=meet
-AWS_S3_SECRET_ACCESS_KEY=password
-
 LIVEKIT_URL=ws://livekit:7880
 LIVEKIT_API_KEY=devkey
 LIVEKIT_API_SECRET=secret
 
-STT_PROVIDER=voxtral-vllm # voxtral-vllm, kyutai, deepgram
+STT_PROVIDER=kyutai # kyutai, deepgram
 ENABLE_SILERO_VAD=False
 
-DEEPGRAM_API_KEY=your-deepgram-api-key
+DEEPGRAM_API_KEY=
 
-KYUTAI_STT_BASE_URL=url
-KYUTAI_API_KEY=your-kyutai-api-key
-
-VOXTRAL_VLLM_BASE_URL=wss://<host>/v1/realtime
-VOXTRAL_VLLM_MODEL=voxtral-mini-4b-realtime-2602
-VOXTRAL_VLLM_API_KEY=your-vllm-api-key
-VOXTRAL_VLLM_TARGET_STREAMING_DELAY_MS=480
+KYUTAI_STT_BASE_URL=
+KYUTAI_API_KEY=
 
 SENTRY_DSN=
 SENTRY_ENVIRONMENT=

src/agents/multi_user_transcriber.py

@@ -1,7 +1,6 @@
 """Multi user transcription agent."""
 
 import asyncio
-import contextlib
 import logging
 import os
 
@@ -26,7 +25,6 @@ from livekit.agents import (
 )
 from livekit.plugins import deepgram, silero
 
-import voxtral_vllm_stt
 from observability import configure_sentry, set_job_context
 from tasks import done_callback
 
@@ -38,18 +36,9 @@ TRANSCRIBER_AGENT_NAME = os.getenv("TRANSCRIBER_AGENT_NAME", "multi-user-transcr
 STT_PROVIDER = os.getenv("STT_PROVIDER", "deepgram")
 ENABLE_SILERO_VAD = os.getenv("ENABLE_SILERO_VAD", "true").lower() == "true"
 
-SESSION_DRAIN_TIMEOUT_S = 15.0
 
-
-def create_stt_provider(vad: silero.VAD | None = None):
-    """Create STT provider based on environment configuration.
-
-    Args:
-        vad: Shared, prewarmed VAD instance. Required in practice for
-            voxtral-vllm (no server-side endpointing): if omitted, the plugin
-            loads its own Silero model synchronously on the event loop, once
-            per participant, freezing all active sessions for the duration.
-    """
+def create_stt_provider():
+    """Create STT provider based on environment configuration."""
     if STT_PROVIDER == "deepgram":
         # Note: Not all Deepgram API parameters are supported by the LiveKit plugin
         # detect_language is NOT supported for real-time streaming
@@ -60,9 +49,6 @@ def create_stt_provider(vad: silero.VAD | None = None):
         )
     elif STT_PROVIDER == "kyutai":
         _stt_instance = kyutai.STT(base_url=os.getenv("KYUTAI_STT_BASE_URL"))
-    elif STT_PROVIDER == "voxtral-vllm":
-        # The plugin resolves base_url / model / api_key from the environment.
-        _stt_instance = voxtral_vllm_stt.STT(vad=vad)
     else:
         raise ValueError(f"Unknown STT_PROVIDER: {STT_PROVIDER}")
 
@@ -72,9 +58,9 @@ def create_stt_provider(vad: silero.VAD | None = None):
 class Transcriber(Agent):
     """Create a transcription agent for a specific participant."""
 
-    def __init__(self, *, participant_identity: str, vad: silero.VAD | None = None):
+    def __init__(self, *, participant_identity: str):
         """Init transcription agent."""
-        stt = create_stt_provider(vad=vad)
+        stt = create_stt_provider()
 
         super().__init__(
             instructions="not-needed",
@@ -90,7 +76,6 @@ class MultiUserTranscriber:
         """Init multi user transcription agent."""
         self.ctx = ctx
         self._sessions: dict[str, AgentSession] = {}
-        self._starting: dict[str, asyncio.Task] = {}
         self._tasks: set[asyncio.Task] = set()
 
     def start(self):
@@ -111,30 +96,22 @@ class MultiUserTranscriber:
 
     def on_participant_connected(self, participant: rtc.RemoteParticipant):
         """Handle new participant connection by starting transcription session."""
-        identity = participant.identity
-        if identity in self._sessions or identity in self._starting:
+        if participant.identity in self._sessions:
             return
 
-        logger.info(f"starting session for {identity}")
+        logger.info(f"starting session for {participant.identity}")
         task = asyncio.create_task(self._start_session(participant))
-        self._starting[identity] = task
         self._tasks.add(task)
-        task.add_done_callback(lambda t, i=identity: self._starting.pop(i, None))
         task.add_done_callback(
             done_callback(
                 logger,
                 self._tasks,
-                f"start transcription session for {identity}",
+                f"start transcription session for {participant.identity}",
             )
         )
 
     def on_participant_disconnected(self, participant: rtc.RemoteParticipant):
         """Handle participant disconnection by closing transcription session."""
-        if (start_task := self._starting.pop(participant.identity, None)) is not None:
-            logger.info(f"cancelling pending session start for {participant.identity}")
-            start_task.cancel()
-            return
-
         if (session := self._sessions.pop(participant.identity, None)) is None:
             return
 
@@ -150,12 +127,10 @@ class MultiUserTranscriber:
         )
 
     async def _start_session(self, participant: rtc.RemoteParticipant) -> AgentSession:
-        """Create and start transcription session for participant.
+        """Create and start transcription session for participant."""
+        if participant.identity in self._sessions:
+            return self._sessions[participant.identity]
 
-        Deduplication happens synchronously in on_participant_connected via
-        self._starting; by the time this coroutine runs, the identity is
-        already reserved.
-        """
         vad = self.ctx.proc.userdata.get("vad", None)
         session = AgentSession(vad=vad)
         room_io = RoomIO(
@@ -166,30 +141,18 @@ class MultiUserTranscriber:
                 text_input=False, audio_output=False, text_output=True
             ),
         )
-        try:
-            await room_io.start()
-            await session.start(
-                agent=Transcriber(
-                    participant_identity=participant.identity,
-                    vad=vad,
-                )
+        await room_io.start()
+        await session.start(
+            agent=Transcriber(
+                participant_identity=participant.identity,
             )
-        except BaseException:
-            with contextlib.suppress(Exception):
-                await session.aclose()
-            raise
+        )
         self._sessions[participant.identity] = session
         return session
 
     async def _close_session(self, sess: AgentSession) -> None:
         """Close and cleanup transcription session."""
-        try:
-            await asyncio.wait_for(sess.drain(), timeout=SESSION_DRAIN_TIMEOUT_S)
-        except (TimeoutError, asyncio.TimeoutError):
-            logger.warning(
-                "session drain timed out after %.0fs; forcing close",
-                SESSION_DRAIN_TIMEOUT_S,
-            )
+        await sess.drain()
         await sess.aclose()
 
 

src/agents/pyproject.toml

@@ -12,8 +12,6 @@ dependencies = [
     "protobuf==6.33.6",
     "minio==7.2.20",
     "sentry-sdk==2.66.1",
-    "websockets==17.1",
-    "httpx==0.28.1",
 ]
 
 [project.optional-dependencies]

src/agents/voxtral_vllm_stt.py

@@ -1,476 +0,0 @@
-"""LiveKit STT plugin for Voxtral Realtime served via vLLM (/v1/realtime).
-
-vLLM exposes Voxtral Realtime over a WebSocket that follows the OpenAI Realtime
-API protocol (not Mistral's proprietary realtime protocol).
-"""
-
-from __future__ import annotations
-
-import asyncio
-import base64
-import json
-import logging
-import os
-import weakref
-from collections import deque
-from dataclasses import dataclass, field
-
-import websockets
-from livekit.agents import (
-    DEFAULT_API_CONNECT_OPTIONS,
-    APIConnectionError,
-    APIConnectOptions,
-    APIStatusError,
-    stt,
-    utils,
-)
-from livekit.agents import (
-    vad as vad_module,
-)
-from livekit.agents.types import NOT_GIVEN, NotGivenOr
-from livekit.agents.utils import is_given
-
-logger = logging.getLogger("voxtral-vllm-stt")
-
-SAMPLE_RATE = 16000
-NUM_CHANNELS = 1
-CHUNK_SAMPLES = 1600  # 100 ms @ 16 kHz mono
-PREROLL_CHUNKS = 5  # keep 500 ms of audio before start of speech as detected by VAD
-
-# Reconnect policy: exponential backoff capped at MAX, give up after MAX_ATTEMPTS
-# consecutive failures (a successful handshake resets the counter).
-RECONNECT_BACKOFF_BASE_S = 0.5
-RECONNECT_BACKOFF_MAX_S = 8.0
-RECONNECT_MAX_ATTEMPTS = 5
-
-
-@dataclass
-class _STTOptions:
-    base_url: str
-    model: str
-    api_key: str | None
-    target_streaming_delay_ms: int | None
-
-
-@dataclass
-class _PendingUtterance:
-    """An utterance in flight on the shared websocket used for reconnect.
-
-    `sent_chunks` holds every chunk we have already enqueued for send on this
-    or a prior connection; on reconnect we replay them before resuming reads
-    from `queue`. vLLM concatenates `input_audio_buffer.append` events into a
-    single audio buffer per generation, so duplicates from a partial prior send
-    are harmless.
-    """
-
-    queue: asyncio.Queue[bytes | None]
-    sent_chunks: list[bytes] = field(default_factory=list)
-    ended: bool = False
-
-
-class STT(stt.STT):
-    """LiveKit STT speaking the OpenAI Realtime protocol served by vLLM."""
-
-    def __init__(
-        self,
-        *,
-        base_url: NotGivenOr[str] = NOT_GIVEN,
-        model: NotGivenOr[str] = NOT_GIVEN,
-        api_key: NotGivenOr[str] = NOT_GIVEN,
-        target_streaming_delay_ms: NotGivenOr[int] = NOT_GIVEN,
-        vad: vad_module.VAD | None = None,
-    ) -> None:
-        """Build the STT.
-
-        Args:
-            base_url: WebSocket URL of the vLLM realtime endpoint, e.g.
-                ws://example:8000/v1/realtime. Falls back to $VOXTRAL_VLLM_BASE_URL.
-            model: Model name exposed by vLLM, default
-                mistralai/Voxtral-Mini-4B-Realtime-2602.
-            api_key: Optional bearer token. Falls back to $VOXTRAL_VLLM_API_KEY.
-            target_streaming_delay_ms: Target streaming delay in ms forwarded to
-                vLLM via session.update. Falls back to
-                $VOXTRAL_VLLM_TARGET_STREAMING_DELAY_MS, else server default.
-            vad: Voice Activity Detector. If omitted, Silero VAD is loaded.
-        """
-        super().__init__(
-            capabilities=stt.STTCapabilities(streaming=True, interim_results=True)
-        )
-
-        resolved_url = (
-            base_url
-            if is_given(base_url)
-            else os.environ.get(
-                "VOXTRAL_VLLM_BASE_URL", "ws://127.0.0.1:8000/v1/realtime"
-            )
-        )
-        resolved_model = (
-            model
-            if is_given(model)
-            else os.environ.get(
-                "VOXTRAL_VLLM_MODEL", "mistralai/Voxtral-Mini-4B-Realtime-2602"
-            )
-        )
-        resolved_key = (
-            api_key if is_given(api_key) else os.environ.get("VOXTRAL_VLLM_API_KEY")
-        )
-        resolved_delay = (
-            target_streaming_delay_ms
-            if is_given(target_streaming_delay_ms)
-            else (
-                int(os.environ["VOXTRAL_VLLM_TARGET_STREAMING_DELAY_MS"])
-                if os.environ.get("VOXTRAL_VLLM_TARGET_STREAMING_DELAY_MS")
-                else None
-            )
-        )
-
-        if vad is None:
-            try:
-                from livekit.plugins.silero import VAD as SileroVAD  # noqa: PLC0415
-            except ImportError as exc:
-                raise ImportError(
-                    "livekit-plugins-silero is required for vLLM Voxtral realtime "
-                    "(no server-side endpointing)."
-                ) from exc
-            vad = SileroVAD.load()
-        self._vad = vad
-
-        self._opts = _STTOptions(
-            base_url=resolved_url,
-            model=resolved_model,
-            api_key=resolved_key,
-            target_streaming_delay_ms=resolved_delay,
-        )
-        self._streams: weakref.WeakSet[SpeechStream] = weakref.WeakSet()
-
-    @property
-    def model(self) -> str:
-        """Return the configured vLLM model name."""
-        return self._opts.model
-
-    @property
-    def provider(self) -> str:
-        """Return the provider identifier."""
-        return "vllm-voxtral-realtime"
-
-    async def _recognize_impl(self, *_args, **_kwargs) -> stt.SpeechEvent:
-        raise NotImplementedError(
-            "vLLM Voxtral Realtime STT only supports streaming recognition."
-        )
-
-    def stream(
-        self,
-        *,
-        conn_options: APIConnectOptions = DEFAULT_API_CONNECT_OPTIONS,
-    ) -> SpeechStream:
-        """Open a new streaming recognition stream."""
-        s = SpeechStream(
-            stt=self,
-            opts=self._opts,
-            vad_instance=self._vad,
-            conn_options=conn_options,
-        )
-        self._streams.add(s)
-        return s
-
-
-class SpeechStream(stt.RecognizeStream):
-    """Voxtral realtime handler."""
-
-    def __init__(
-        self,
-        *,
-        stt: STT,
-        opts: _STTOptions,
-        vad_instance: vad_module.VAD,
-        conn_options: APIConnectOptions,
-    ) -> None:
-        """Init the speech stream."""
-        super().__init__(stt=stt, conn_options=conn_options, sample_rate=SAMPLE_RATE)
-        self._opts = opts
-        self._vad = vad_instance
-        self._utterance_q: asyncio.Queue[bytes | None] | None = None
-        self._speaking = False
-        self._preroll: deque[bytes] = deque(maxlen=PREROLL_CHUNKS)
-        # Voxtral realtime is strictly sequential: only one generation runs at a
-        # time, and a new `commit` is ignored while the previous one is still
-        # producing. We queue per-utterance audio buffers here and let the
-        # pipeline process them one by one on the shared websocket.
-        self._utterance_chan: asyncio.Queue[asyncio.Queue[bytes | None] | None] = (
-            asyncio.Queue()
-        )
-
-    @utils.log_exceptions(logger=logger)
-    async def _run(self) -> None:
-        vad_stream = self._vad.stream()
-
-        bstream = utils.audio.AudioByteStream(
-            sample_rate=SAMPLE_RATE,
-            num_channels=NUM_CHANNELS,
-            samples_per_channel=CHUNK_SAMPLES,
-        )
-
-        async def input_task() -> None:
-            async for data in self._input_ch:
-                if isinstance(data, self._FlushSentinel):
-                    for frame in bstream.flush():
-                        self._handle_chunk(frame.data.tobytes())
-                    continue
-
-                vad_stream.push_frame(data)
-                for frame in bstream.write(data.data.tobytes()):
-                    self._handle_chunk(frame.data.tobytes())
-
-            vad_stream.end_input()
-
-        async def vad_task() -> None:
-            async for ev in vad_stream:
-                if ev.type == vad_module.VADEventType.START_OF_SPEECH:
-                    self._on_start_of_speech()
-                elif ev.type == vad_module.VADEventType.END_OF_SPEECH:
-                    self._on_end_of_speech()
-
-        pipeline_t = asyncio.create_task(self._utterance_pipeline())
-        try:
-            await asyncio.gather(input_task(), vad_task())
-            # signal end-of-stream; pipeline finishes pending utterances first
-            self._utterance_chan.put_nowait(None)
-            await pipeline_t
-        except (APIStatusError, APIConnectionError, asyncio.CancelledError):
-            raise
-        except Exception as exc:
-            logger.exception("vLLM realtime stream failed")
-            raise APIConnectionError() from exc
-        finally:
-            if not pipeline_t.done():
-                pipeline_t.cancel()
-            try:
-                await pipeline_t
-            except asyncio.CancelledError:
-                # CancelledError is the expected flow on cancel()
-                pass
-            except Exception:
-                logger.exception("utterance pipeline failed during finalize")
-            await vad_stream.aclose()
-
-    def _handle_chunk(self, chunk: bytes) -> None:
-        self._preroll.append(chunk)
-        if self._speaking and self._utterance_q is not None:
-            self._utterance_q.put_nowait(chunk)
-
-    def _on_start_of_speech(self) -> None:
-        if self._speaking:
-            return
-        self._speaking = True
-        q: asyncio.Queue[bytes | None] = asyncio.Queue()
-        for chunk in self._preroll:
-            q.put_nowait(chunk)
-        self._utterance_q = q
-        self._utterance_chan.put_nowait(q)
-        self._event_ch.send_nowait(
-            stt.SpeechEvent(type=stt.SpeechEventType.START_OF_SPEECH)
-        )
-
-    def _on_end_of_speech(self) -> None:
-        if not self._speaking:
-            return
-        self._speaking = False
-        if self._utterance_q is not None:
-            self._utterance_q.put_nowait(None)
-        self._utterance_q = None
-        self._event_ch.send_nowait(
-            stt.SpeechEvent(type=stt.SpeechEventType.END_OF_SPEECH)
-        )
-
-    async def _handshake(self, ws: websockets.ClientConnection) -> str:
-        created = json.loads(await ws.recv())
-        if created.get("type") != "session.created":
-            raise APIStatusError(
-                f"expected session.created, got {created}",
-                status_code=500,
-                body=created,
-            )
-        session_update: dict = {"type": "session.update", "model": self._opts.model}
-        if self._opts.target_streaming_delay_ms is not None:
-            session_update["target_streaming_delay_ms"] = (
-                self._opts.target_streaming_delay_ms
-            )
-        await ws.send(json.dumps(session_update))
-        return created.get("id", "")
-
-    def _auth_headers(self) -> dict[str, str]:
-        if self._opts.api_key:
-            return {"Authorization": f"Bearer {self._opts.api_key}"}
-        return {}
-
-    async def _utterance_pipeline(self) -> None:
-        # Owns the websocket lifecycle. On drop, reopens and resumes the
-        # in-flight utterance (if any) by replaying its already-sent chunks.
-        pending: _PendingUtterance | None = None
-        attempt = 0
-        while True:
-            try:
-                async with websockets.connect(
-                    self._opts.base_url,
-                    additional_headers=self._auth_headers(),
-                    open_timeout=self._conn_options.timeout,
-                ) as ws:
-                    request_id = await self._handshake(ws)
-                    attempt = 0
-                    while True:
-                        if pending is None:
-                            q = await self._utterance_chan.get()
-                            if q is None:
-                                return
-                            pending = _PendingUtterance(queue=q)
-                        await self._process_utterance(ws, pending, request_id)
-                        pending = None
-            except (websockets.WebSocketException, OSError, TimeoutError) as exc:
-                attempt += 1
-                if attempt > RECONNECT_MAX_ATTEMPTS:
-                    logger.exception(
-                        "vLLM realtime: giving up after %d reconnect attempts",
-                        RECONNECT_MAX_ATTEMPTS,
-                    )
-                    raise APIConnectionError() from exc
-                backoff = min(
-                    RECONNECT_BACKOFF_BASE_S * (2 ** (attempt - 1)),
-                    RECONNECT_BACKOFF_MAX_S,
-                )
-                if pending is None:
-                    logger.warning(
-                        "vLLM WS connection lost between utterances "
-                        "(attempt %d/%d): %s; retrying in %.1fs",
-                        attempt,
-                        RECONNECT_MAX_ATTEMPTS,
-                        exc,
-                        backoff,
-                    )
-                else:
-                    logger.warning(
-                        "vLLM WS dropped mid-utterance (%d chunks buffered, "
-                        "ended=%s, attempt %d/%d): %s; retrying in %.1fs",
-                        len(pending.sent_chunks),
-                        pending.ended,
-                        attempt,
-                        RECONNECT_MAX_ATTEMPTS,
-                        exc,
-                        backoff,
-                    )
-                await asyncio.sleep(backoff)
-
-    async def _process_utterance(
-        self,
-        ws: websockets.ClientConnection,
-        pending: _PendingUtterance,
-        request_id: str,
-    ) -> None:
-        # Start a fresh generation. Safe to send here: the previous utterance's
-        # transcription.done has already been received (we await it below), so
-        # the server-side generation_task is done and won't ignore this commit.
-        await ws.send(json.dumps({"type": "input_audio_buffer.commit"}))
-        send_t = asyncio.create_task(self._send_audio(ws, pending))
-        try:
-            await self._receive_one_transcription(ws, request_id)
-        finally:
-            if not send_t.done():
-                send_t.cancel()
-            try:
-                await send_t
-            except (asyncio.CancelledError, websockets.WebSocketException):
-                pass
-            except Exception:
-                logger.exception("send-audio task failed during finalize")
-
-    @staticmethod
-    async def _send_audio(
-        ws: websockets.ClientConnection, pending: _PendingUtterance
-    ) -> None:
-        # Replay anything already sent on a previous (now-dead) connection.
-        # sent_chunks is appended before send, so a chunk that failed to send
-        # last time is still present and gets retried here.
-        for chunk in pending.sent_chunks:
-            await ws.send(
-                json.dumps(
-                    {
-                        "type": "input_audio_buffer.append",
-                        "audio": base64.b64encode(chunk).decode("ascii"),
-                    }
-                )
-            )
-        if pending.ended:
-            await ws.send(
-                json.dumps({"type": "input_audio_buffer.commit", "final": True})
-            )
-            return
-        while True:
-            chunk = await pending.queue.get()
-            if chunk is None:
-                pending.ended = True
-                await ws.send(
-                    json.dumps({"type": "input_audio_buffer.commit", "final": True})
-                )
-                return
-            pending.sent_chunks.append(chunk)
-            await ws.send(
-                json.dumps(
-                    {
-                        "type": "input_audio_buffer.append",
-                        "audio": base64.b64encode(chunk).decode("ascii"),
-                    }
-                )
-            )
-
-    async def _receive_one_transcription(
-        self, ws: websockets.ClientConnection, request_id: str
-    ) -> None:
-        # Use recv() rather than `async for`: the latter swallows
-        # ConnectionClosed on close-mid-iteration, which would let a dropped
-        # WS look like a clean "no transcription" return.
-        current_text = ""
-        while True:
-            raw = await ws.recv()
-            data = json.loads(raw)
-            event_type = data.get("type")
-
-            if event_type == "transcription.delta":
-                delta = data.get("delta", "")
-                if not delta:
-                    continue
-                current_text += delta
-                self._event_ch.send_nowait(
-                    stt.SpeechEvent(
-                        type=stt.SpeechEventType.INTERIM_TRANSCRIPT,
-                        request_id=request_id,
-                        alternatives=[stt.SpeechData(text=current_text, language="")],
-                    )
-                )
-            elif event_type == "transcription.done":
-                final_text = data.get("text") or current_text
-                self._event_ch.send_nowait(
-                    stt.SpeechEvent(
-                        type=stt.SpeechEventType.FINAL_TRANSCRIPT,
-                        request_id=request_id,
-                        alternatives=[stt.SpeechData(text=final_text, language="")],
-                    )
-                )
-                usage = data.get("usage") or {}
-                self._event_ch.send_nowait(
-                    stt.SpeechEvent(
-                        type=stt.SpeechEventType.RECOGNITION_USAGE,
-                        request_id=request_id,
-                        recognition_usage=stt.RecognitionUsage(
-                            audio_duration=float(
-                                usage.get("audio_seconds")
-                                or usage.get("prompt_audio_seconds")
-                                or 0
-                            ),
-                            input_tokens=int(usage.get("prompt_tokens") or 0),
-                            output_tokens=int(usage.get("completion_tokens") or 0),
-                        ),
-                    )
-                )
-                return
-            elif event_type == "error":
-                err = data.get("error")
-                raise APIStatusError(str(err), status_code=500, body=data)

src/frontend/package-lock.json

@@ -29,7 +29,7 @@
         "i18next-parser": "9.4.0",
         "i18next-resources-to-backend": "1.2.3",
         "livekit-client": "2.21.0",
-        "posthog-js": "1.414.0",
+        "posthog-js": "1.409.5",
         "react": "18.3.1",
         "react-aria": "3.50.0",
         "react-aria-components": "1.19.0",
@@ -1720,28 +1720,28 @@
       "license": "MIT"
     },
     "node_modules/@posthog/browser-common": {
-      "version": "0.4.0",
-      "resolved": "https://registry.npmjs.org/@posthog/browser-common/-/browser-common-0.4.0.tgz",
-      "integrity": "sha512-W9DCGVks15docUMPvJ2nd8NS16Gn74bsGWuaeg31beEKFSjdW8wvnQ1ETY6WSql5pYxZb3GdJmEUZVVstKSrBQ==",
+      "version": "0.3.1",
+      "resolved": "https://registry.npmjs.org/@posthog/browser-common/-/browser-common-0.3.1.tgz",
+      "integrity": "sha512-1nhMVY1wnHADTg8tR9yvm+lPAz5ROxznfQlBtLzB2FFo4lp/LU8lk9KyFPsARDkBxCfYachsJfvRjocL1G/AJQ==",
       "license": "MIT",
       "dependencies": {
-        "@posthog/core": "^1.46.8",
-        "@posthog/types": "^1.402.0"
+        "@posthog/core": "^1.46.0",
+        "@posthog/types": "^1.399.0"
       }
     },
     "node_modules/@posthog/core": {
-      "version": "1.49.1",
-      "resolved": "https://registry.npmjs.org/@posthog/core/-/core-1.49.1.tgz",
-      "integrity": "sha512-jdZh85tG56OXLH881CVwBZyiXCPPaZasfYeWwm9kVUvxC/Rb+lz7wYN9GuqSEmNPJgVnK4v8wS0bCaFc3OmVEA==",
+      "version": "1.48.3",
+      "resolved": "https://registry.npmjs.org/@posthog/core/-/core-1.48.3.tgz",
+      "integrity": "sha512-kwVDVvwtCTXctApA2tpnwDjDDan8LrkwCW1Wv6PABaVs/s5ahbQ9W3pvdXcqr71HCuqukSA1jp7MySczebubGg==",
       "license": "MIT",
       "dependencies": {
-        "@posthog/types": "^1.407.0"
+        "@posthog/types": "^1.405.0"
       }
     },
     "node_modules/@posthog/types": {
-      "version": "1.407.1",
-      "resolved": "https://registry.npmjs.org/@posthog/types/-/types-1.407.1.tgz",
-      "integrity": "sha512-WhbkXPC2rgylXqmxHqv70ffI3k+KxyR6s7DBIfr5NvIqHkxp6v0pk31D/jbz0DNVbzwkLjyll2pxr4FNbJiYzg==",
+      "version": "1.405.0",
+      "resolved": "https://registry.npmjs.org/@posthog/types/-/types-1.405.0.tgz",
+      "integrity": "sha512-4rZ/taVXKQxs9Jrf7ZjlCRgrOSL69oKAgIWJQa5kRNJ6wll1UANbrJTSY+Su1e88LIG4zZVjKKKjyQHCkHdHcw==",
       "license": "MIT"
     },
     "node_modules/@react-aria/overlays": {
@@ -4794,10 +4794,13 @@
       }
     },
     "node_modules/dompurify": {
-      "version": "3.4.14",
-      "resolved": "https://registry.npmjs.org/dompurify/-/dompurify-3.4.14.tgz",
-      "integrity": "sha512-dVoH9z+MY+C9IilgGCk3YfFqjLi3fChm2OiKJMzh6axrJ5qwxqWaZamgmHrpv22CN/KdbZJuGEGgfQoL00LTdg==",
+      "version": "3.3.2",
+      "resolved": "https://registry.npmjs.org/dompurify/-/dompurify-3.3.2.tgz",
+      "integrity": "sha512-6obghkliLdmKa56xdbLOpUZ43pAR6xFy1uOrxBaIDjT+yaRuuybLjGS9eVBoSR/UPU5fq3OXClEHLJNGvbxKpQ==",
       "license": "(MPL-2.0 OR Apache-2.0)",
+      "engines": {
+        "node": ">=20"
+      },
       "optionalDependencies": {
         "@types/trusted-types": "^2.0.7"
       }
@@ -9131,21 +9134,20 @@
       "license": "MIT"
     },
     "node_modules/posthog-js": {
-      "version": "1.414.0",
-      "resolved": "https://registry.npmjs.org/posthog-js/-/posthog-js-1.414.0.tgz",
-      "integrity": "sha512-dtZd4asdskr8lNyltAEX6zyn48uO1pO0EMvx6AXJU65PFhu6yn2LPbKtQcyLysjcN57FJPNT9QYL6St5SBJHqw==",
+      "version": "1.409.5",
+      "resolved": "https://registry.npmjs.org/posthog-js/-/posthog-js-1.409.5.tgz",
+      "integrity": "sha512-s1iJz+vq0YAluUD4OwLjUFIJt/9pqiiB6MITvHWIS16a7pqTJqbSLXsd5/8kJcIgfrOSZHe+mdYAXLYcJCS4LQ==",
       "license": "(Apache-2.0 AND MIT)",
       "dependencies": {
-        "@posthog/browser-common": "^0.4.0",
-        "@posthog/core": "^1.46.9",
-        "@posthog/types": "^1.402.2",
+        "@posthog/browser-common": "^0.3.1",
+        "@posthog/core": "^1.46.1",
+        "@posthog/types": "^1.399.0",
         "core-js": "^3.49.0",
-        "dompurify": "^3.4.12",
+        "dompurify": "^3.3.2",
         "fflate": "^0.4.8",
         "preact": "^10.29.3",
         "query-selector-shadow-dom": "^1.0.1",
-        "web-vitals": "^5.3.0",
-        "web-vitals-soft-navs": "npm:web-vitals@6.0.0"
+        "web-vitals": "^5.3.0"
       }
     },
     "node_modules/powershell-utils": {
@@ -11657,13 +11659,6 @@
       "integrity": "sha512-q6LWsLatGYZp5VGBIOvbTj6JBV2nOmC8KvWztXBmwJcfFAzhwKwbOxhUH306XY3CcaZDUlSmSuNPBsCn0bFu+g==",
       "license": "Apache-2.0"
     },
-    "node_modules/web-vitals-soft-navs": {
-      "name": "web-vitals",
-      "version": "6.0.0",
-      "resolved": "https://registry.npmjs.org/web-vitals/-/web-vitals-6.0.0.tgz",
-      "integrity": "sha512-Guaibvy/+uNtL6Bsu4jmMJGzuSl91oeRH5iO9pPRbYftnFUr3yqT1TUNX/OE4o9HexuEMU3Kb/Wg7iKhlffZUA==",
-      "license": "Apache-2.0"
-    },
     "node_modules/webrtc-adapter": {
       "version": "9.0.6",
       "resolved": "https://registry.npmjs.org/webrtc-adapter/-/webrtc-adapter-9.0.6.tgz",

src/frontend/package.json

@@ -36,7 +36,7 @@
     "i18next-parser": "9.4.0",
     "i18next-resources-to-backend": "1.2.3",
     "livekit-client": "2.21.0",
-    "posthog-js": "1.414.0",
+    "posthog-js": "1.409.5",
     "react": "18.3.1",
     "react-aria": "3.50.0",
     "react-aria-components": "1.19.0",

src/frontend/src/features/subtitle/component/Subtitles.tsx

@@ -75,15 +75,15 @@ const useTranscriptionState = () => {
     const segment = segments[0]
 
     setTranscriptionSegments((prevSegments) => {
-      const existingIndex = prevSegments.findIndex(
-        (s: TranscriptionSegmentWithParticipant) => s.id === segment.id
-      )
-      if (existingIndex === -1) {
-        return [...prevSegments, { participant, ...segment }]
-      }
-      const next = prevSegments.slice()
-      next[existingIndex] = { ...next[existingIndex], ...segment }
-      return next
+      const existingSegmentIds = new Set(prevSegments.map((s) => s.id))
+      if (existingSegmentIds.has(segment.id)) return prevSegments
+      return [
+        ...prevSegments,
+        {
+          participant: participant,
+          ...segment,
+        },
+      ]
     })
   }