cleanup

src/agents/multi_user_transcriber.py

@@ -49,9 +49,7 @@ def create_stt_provider():
     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
-        # itself (with sensible defaults). Passing os.getenv(...) directly would
-        # forward an explicit None for unset vars, which defeats that fallback.
+        # The plugin resolves base_url / model / api_key from the environment.
         _stt_instance = voxtral_vllm_stt.STT()
     else:
         raise ValueError(f"Unknown STT_PROVIDER: {STT_PROVIDER}")

src/agents/pyproject.toml

@@ -4,7 +4,7 @@ name = "agents"
 version = "1.15.0"
 requires-python = ">=3.12"
 dependencies = [
-    "livekit-agents>=1.4.5",
+    "livekit-agents==1.4.5",
     "livekit-plugins-deepgram==1.4.5",
     "livekit-plugins-silero==1.4.5",
     "livekit-plugins-kyutai-lasuite==0.0.6",

src/agents/voxtral_vllm_stt.py

@@ -49,6 +49,7 @@ class _STTOptions:
     base_url: str
     model: str
     api_key: str | None
+    target_streaming_delay_ms: int | None
 
 
 @dataclass
@@ -76,6 +77,7 @@ class STT(stt.STT):
         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.
@@ -86,6 +88,9 @@ class STT(stt.STT):
             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__(
@@ -109,6 +114,15 @@ class STT(stt.STT):
         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:

src/frontend/src/features/rooms/livekit/processors/RnnNoiseProcessor.ts

@@ -47,8 +47,6 @@ export class RnnNoiseProcessor implements AudioProcessorInterface {
     )
 
     this.destinationNode = audioContext.createMediaStreamDestination()
-    // this.destinationNode.channelCount = 1
-    // this.destinationNode.channelCountMode = 'explicit'
 
     // Connect the audio processing chain
     this.sourceNode

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

@@ -65,8 +65,6 @@ const useTranscriptionState = () => {
   ) => {
     console.log(participant, segments)
 
-
-
     if (!participant || segments.length === 0) return
 
     if (segments.length > 1) {