Working !

Three bugs prevented voice input from working:

1. ResampleTo16000() treated NumFrames as total samples, dividing by
   channel count again — losing half the audio data with stereo input.
   The corrupted audio was unrecognizable to ElevenLabs VAD/STT.

2. Sent nonexistent "client_vad" turn mode in session init. The API has
   no turn.mode field; replaced with turn_timeout parameter.

3. Sent user_activity with every audio chunk, which resets the turn
   timeout timer and prevents the server from taking its turn.

Also: send audio chunks as compact JSON, add message type debug logging,
send conversation_initiation_client_data on connect.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

.claude/MEMORY.md

@@ -43,20 +43,30 @@
 
 ## Plugin Status
 - **PS_AI_Agent_ElevenLabs**: compiles cleanly on UE 5.5 Win64 (verified 2026-02-19)
-- v1.1.0 — all 3 protocol bugs fixed (transcript fields, pong format, client turn mode)
+- v1.5.0 — mic audio chunk size fixed: WASAPI 5ms callbacks accumulated to 100ms before sending
+- v1.4.0 — push-to-talk fully fixed: bAutoStartListening now ignored in Client turn mode
 - Binary WS frame handling implemented (ElevenLabs sends ALL frames as binary, not text)
 - First-byte discrimination: `{` = JSON control message, else = raw PCM audio
 - `SendTextMessage()` added to both WebSocketProxy and ConversationalAgentComponent
-- Connection confirmed working end-to-end; audio receive path functional
+- `conversation_initiation_client_data` now sent immediately on WS connect (required for mic + latency)
+
+## Audio Chunk Size — CRITICAL
+- WASAPI fires mic callbacks every ~5ms → **158 bytes** at 16kHz 16-bit mono
+- ElevenLabs VAD/STT requires **≥3200 bytes (100ms)** per chunk; smaller chunks are silently ignored
+- Fix: `MicAccumulationBuffer` in component accumulates chunks; sends only when `>= MicChunkMinBytes` (3200)
+- `StopListening()` flushes remainder so final partial chunk is never dropped before end-of-turn
 
 ## ElevenLabs WebSocket Protocol Notes
-- **ALL frames are binary** — `OnRawMessage` handles everything; `OnMessage` (text) never fires
+- **ALL frames are binary** — bind ONLY `OnRawMessage`; NEVER bind `OnMessage` (text) — UE fires both for same frame → double audio bug
 - Binary frame discrimination: peek byte[0] → `'{'` (0x7B) = JSON, else = raw PCM audio
 - Fragment reassembly: accumulate into `BinaryFrameBuffer` until `BytesRemaining == 0`
 - Pong: `{"type":"pong","event_id":N}` — `event_id` is **top-level**, NOT nested
 - Transcript: type=`user_transcript`, key=`user_transcription_event`, field=`user_transcript`
-- Client turn mode: `{"type":"user_activity"}` to signal speaking; no explicit end message
+- Client turn mode (`client_vad`): send `user_activity` **with every audio chunk** (not just once) — server needs continuous signal to know user is speaking; stopping chunks = silence detected = agent responds
 - Text input: `{"type":"user_message","text":"..."}` — agent replies with audio + text
+- **MUST send `conversation_initiation_client_data` immediately after WS connect** — without it, server won't process client audio (mic appears dead)
+- `conversation_initiation_client_data` payload: `conversation_config_override.agent.turn.mode`, `conversation_config_override.tts.optimize_streaming_latency`, `custom_llm_extra_body.enable_intermediate_response`
+- `enable_intermediate_response: true` in `custom_llm_extra_body` reduces time-to-first-audio latency
 
 ## API Keys / Secrets
 - ElevenLabs API key is set in **Project Settings → Plugins → ElevenLabs AI Agent** in the Editor

.claude/session_log_2026-02-19.md

@@ -189,9 +189,160 @@ Commit: `99017f4`
 
 ---
 
+---
+
+## Session 3 — 2026-02-19 (bug fixes from live testing)
+
+### 16. Three Runtime Bugs Fixed (v1.2.0)
+
+User reported after live testing:
+1. **AI speaks twice** — every audio response played double
+2. **Cannot speak** — mic capture didn't reach ElevenLabs
+3. **Latency** — requested `enable_intermediate_response: true`
+
+**Bug 1 Root Cause — Double Audio:**
+UE's libwebsockets backend fires **both** `OnMessage()` (text callback) **and** `OnRawMessage()` (binary callback) for the same incoming frame.
+We had bound both `WebSocket->OnMessage()` and `WebSocket->OnRawMessage()` in `Connect()`.
+Result: every audio frame was decoded and enqueued twice → played twice.
+
+Fix: **Remove `OnMessage` binding entirely.** `OnRawMessage` now handles all frames (JSON control messages peeked via first byte, raw PCM otherwise).
+
+**Bug 2 Root Cause — Mic Silent:**
+ElevenLabs requires a `conversation_initiation_client_data` message sent **immediately** after the WebSocket handshake completes. Without it, the server never enters a state where it will accept and process client audio chunks. This is a required session negotiation step, not optional.
+
+Fix: Send `conversation_initiation_client_data` in `OnWsConnected()` before any other message.
+
+**Bug 2 Secondary — Delegate Stacking:**
+`StartListening()` called `Mic->OnAudioCaptured.AddUObject(this, ...)` without first removing existing bindings. If called more than once (e.g. after reconnect), delegates stack up and audio is sent multiple times per frame.
+
+Fix: Add `Mic->OnAudioCaptured.RemoveAll(this)` before `AddUObject` in `StartListening()`.
+
+**Bug 3 — Latency:**
+Added `"enable_intermediate_response": true` inside `custom_llm_extra_body` of the `conversation_initiation_client_data` message. Also added `optimize_streaming_latency: 3` in `conversation_config_override.tts`.
+
+**Files changed:**
+- `ElevenLabsWebSocketProxy.cpp`:
+  - `Connect()`: removed `OnMessage` binding
+  - `OnWsConnected()`: now sends full `conversation_initiation_client_data` JSON
+- `ElevenLabsConversationalAgentComponent.cpp`:
+  - `StartListening()`: added `RemoveAll` guard before delegate binding
+
+---
+
+---
+
+## Session 4 — 2026-02-19 (mic still silent — push-to-talk deeper investigation)
+
+### 17. Two More Bugs Found and Fixed (v1.3.0)
+
+User confirmed Bug 1 (double audio) was fixed. Bug 2 (cannot speak) persisted.
+
+**Analysis of log:**
+- Blueprint is correct: T Pressed → StartListening, T Released → StopListening (proper push-to-talk)
+- Mic opens and closes correctly — audio capture IS happening
+- Server never responds to mic input → audio reaching ElevenLabs but being ignored
+
+**Bug A — TurnMode mismatch in conversation_initiation_client_data:**
+`OnWsConnected()` hardcoded `"mode": "server_vad"` in the init message regardless of the
+component's `TurnMode` setting. User's Blueprint uses Client turn mode (push-to-talk),
+so the server was configured for server_vad while the client sent client_vad audio signals.
+
+Fix: Read `TurnMode` field on the proxy (set from the component before `Connect()`).
+Translate `EElevenLabsTurnMode::Client` → `"client_vad"`, Server → `"server_vad"`.
+
+**Bug B — user_activity never sent continuously:**
+In client VAD mode, ElevenLabs requires `user_activity` to be sent **continuously**
+alongside every audio chunk to keep the server's VAD aware the user is speaking.
+`SendUserTurnStart()` sent it once on key press, but never again during speech.
+Server-side, without continuous `user_activity`, the server treated the audio as noise.
+
+Fix: In `SendAudioChunk()`, automatically send `user_activity` before each audio chunk
+when `TurnMode == Client`. This keeps the signal continuous for the full duration of speech.
+When the user releases T, `StopListening()` stops the mic → audio stops → `user_activity`
+stops → server detects silence and triggers the agent response.
+
+**Bug C — TurnMode not propagated to proxy:**
+`UElevenLabsConversationalAgentComponent` never told the proxy what TurnMode to use.
+Added `WebSocketProxy->TurnMode = TurnMode` before `Connect()` in `StartConversation()`.
+
+**Files changed:**
+- `ElevenLabsWebSocketProxy.h`: added `public TurnMode` field
+- `ElevenLabsWebSocketProxy.cpp`:
+  - `OnWsConnected()`: use `TurnMode` to set correct mode string in init message
+  - `SendAudioChunk()`: auto-send `user_activity` before each chunk in Client mode
+- `ElevenLabsConversationalAgentComponent.cpp`:
+  - `StartConversation()`: set `WebSocketProxy->TurnMode = TurnMode` before `Connect()`
+
+---
+
+---
+
+## Session 5 — 2026-02-19 (still can't speak — bAutoStartListening conflict)
+
+### 18. Root Cause Found and Fixed (v1.4.0)
+
+Log analysis revealed the true root cause:
+
+**Exact sequence:**
+```
+OnConnected → bAutoStartListening=true → StartListening() → bIsListening=true, mic opens
+OnAgentStoppedSpeaking → Blueprint calls StartListening() → bIsListening guard → no-op (already open)
+User presses T → StartListening() → bIsListening guard → no-op
+User releases T → StopListening() → bIsListening=false, mic CLOSES
+User presses T → StartListening() → NOW opens mic (was closed)
+User releases T → StopListening() → mic closes — but ElevenLabs never got audio
+```
+
+**Root cause:** `bAutoStartListening = true` opens the mic on connect and sets `bIsListening = true`.
+In Client/push-to-talk mode, every T-press hits the `bIsListening` guard and does nothing.
+Every T-release closes the auto-started mic. The mic was never open during actual speech.
+
+**Fix:** `HandleConnected()` now only calls `StartListening()` when `TurnMode == Server`.
+In Client mode, `bAutoStartListening` is ignored — the user controls listening via T key.
+
+**File changed:**
+- `ElevenLabsConversationalAgentComponent.cpp`:
+  - `HandleConnected()`: guard `bAutoStartListening` with `TurnMode == Server` check
+
+---
+
+---
+
+## Session 6 — 2026-02-19 (audio chunk size fix)
+
+### 19. Mic Audio Chunk Accumulation (v1.5.0)
+
+**Root cause (from diagnostic log in Session 5):**
+Log showed hundreds of `SendAudioChunk: 158 bytes (TurnMode=Client)` lines with zero server responses.
+- 158 bytes = 79 samples = ~5ms of audio at 16kHz 16-bit mono
+- WASAPI (Windows Audio Session API) fires the `FAudioCapture` callback at its internal buffer period (~5ms)
+- ElevenLabs requires a minimum chunk size for its VAD and STT to operate (~100ms / 3200 bytes)
+- Tiny 5ms fragments arrived at the server but were silently ignored → agent never responded
+
+**Fix applied:**
+Added `MicAccumulationBuffer TArray<uint8>` to `UElevenLabsConversationalAgentComponent`.
+`OnMicrophoneDataCaptured()` appends each callback's converted bytes and only calls `SendAudioChunk`
+when `>= MicChunkMinBytes` (3200 bytes = 100ms) have accumulated.
+
+`StopListening()` flushes any remaining bytes in the buffer before sending `SendUserTurnEnd()`,
+so the last partial chunk of speech is never dropped.
+
+`HandleDisconnected()` clears the buffer to prevent stale data on reconnect.
+
+**Files changed:**
+- `ElevenLabsConversationalAgentComponent.h`: added `MicAccumulationBuffer` + `MicChunkMinBytes = 3200`
+- `ElevenLabsConversationalAgentComponent.cpp`:
+  - `OnMicrophoneDataCaptured()`: accumulate → send when threshold reached
+  - `StopListening()`: flush remainder before end-of-turn signal
+  - `HandleDisconnected()`: clear accumulation buffer
+
+Commit: `91cf5b1`
+
+---
+
 ## Next Steps (not done yet)
 
-- [ ] Verify mic audio actually reaches ElevenLabs (enable Verbose Logging, test in Editor)
+- [ ] Test v1.5.0 in Editor — verify push-to-talk mic works end-to-end (should be the final fix)
 - [ ] Test `USoundWaveProcedural` underflow behaviour in practice (check for audio glitches)
 - [ ] Test `SendTextMessage` end-to-end in Blueprint
 - [ ] Add lip-sync support (future)

Unreal/PS_AI_Agent/Plugins/PS_AI_Agent_ElevenLabs/Config/FilterPlugin.ini

@@ -0,0 +1,8 @@
+[FilterPlugin]
+; This section lists additional files which will be packaged along with your plugin. Paths should be listed relative to the root plugin directory, and
+; may include "...", "*", and "?" wildcards to match directories, files, and individual characters respectively.
+;
+; Examples:
+;    /README.txt
+;    /Extras/...
+;    /Binaries/ThirdParty/*.dll

Unreal/PS_AI_Agent/Plugins/PS_AI_Agent_ElevenLabs/Source/PS_AI_Agent_ElevenLabs/Private/ElevenLabsConversationalAgentComponent.cpp

@@ -86,6 +86,10 @@ void UElevenLabsConversationalAgentComponent::StartConversation()
 			&UElevenLabsConversationalAgentComponent::HandleInterrupted);
 	}
 
+	// Pass configuration to the proxy before connecting.
+	WebSocketProxy->TurnMode = TurnMode;
+	WebSocketProxy->bSpeculativeTurn = bSpeculativeTurn;
+
 	WebSocketProxy->Connect(AgentID);
 }
 
@@ -128,6 +132,9 @@ void UElevenLabsConversationalAgentComponent::StartListening()
 		Mic->RegisterComponent();
 	}
 
+	// Always remove existing binding first to prevent duplicate delegates stacking
+	// up if StartListening is called more than once without a matching StopListening.
+	Mic->OnAudioCaptured.RemoveAll(this);
 	Mic->OnAudioCaptured.AddUObject(this,
 		&UElevenLabsConversationalAgentComponent::OnMicrophoneDataCaptured);
 	Mic->StartCapture();
@@ -147,6 +154,15 @@ void UElevenLabsConversationalAgentComponent::StopListening()
 		Mic->OnAudioCaptured.RemoveAll(this);
 	}
 
+	// Flush any partially-accumulated mic audio before signalling end-of-turn.
+	// This ensures the final words aren't discarded just because the last callback
+	// didn't push the buffer over the MicChunkMinBytes threshold.
+	if (MicAccumulationBuffer.Num() > 0 && WebSocketProxy && IsConnected())
+	{
+		WebSocketProxy->SendAudioChunk(MicAccumulationBuffer);
+	}
+	MicAccumulationBuffer.Reset();
+
 	if (WebSocketProxy && TurnMode == EElevenLabsTurnMode::Client)
 	{
 		WebSocketProxy->SendUserTurnEnd();
@@ -193,7 +209,12 @@ void UElevenLabsConversationalAgentComponent::HandleConnected(const FElevenLabsC
 	UE_LOG(LogElevenLabsAgent, Log, TEXT("Agent connected. ConversationID=%s"), *Info.ConversationID);
 	OnAgentConnected.Broadcast(Info);
 
-	if (bAutoStartListening)
+	// In Client turn mode (push-to-talk), the user controls listening manually via
+	// StartListening()/StopListening(). Auto-starting would leave the mic open
+	// permanently and interfere with push-to-talk — the T-release StopListening()
+	// would close the mic that auto-start opened, leaving the user unable to speak.
+	// Only auto-start in Server VAD mode where the mic stays open the whole session.
+	if (bAutoStartListening && TurnMode == EElevenLabsTurnMode::Server)
 	{
 		StartListening();
 	}
@@ -204,6 +225,7 @@ void UElevenLabsConversationalAgentComponent::HandleDisconnected(int32 StatusCod
 	UE_LOG(LogElevenLabsAgent, Log, TEXT("Agent disconnected. Code=%d Reason=%s"), StatusCode, *Reason);
 	bIsListening = false;
 	bAgentSpeaking = false;
+	MicAccumulationBuffer.Reset();
 	OnAgentDisconnected.Broadcast(StatusCode, Reason);
 }
 
@@ -220,12 +242,18 @@ void UElevenLabsConversationalAgentComponent::HandleAudioReceived(const TArray<u
 
 void UElevenLabsConversationalAgentComponent::HandleTranscript(const FElevenLabsTranscriptSegment& Segment)
 {
-	OnAgentTranscript.Broadcast(Segment);
+	if (bEnableUserTranscript)
+	{
+		OnAgentTranscript.Broadcast(Segment);
+	}
 }
 
 void UElevenLabsConversationalAgentComponent::HandleAgentResponse(const FString& ResponseText)
 {
-	OnAgentTextResponse.Broadcast(ResponseText);
+	if (bEnableAgentTextResponse)
+	{
+		OnAgentTextResponse.Broadcast(ResponseText);
+	}
 }
 
 void UElevenLabsConversationalAgentComponent::HandleInterrupted()
@@ -321,8 +349,18 @@ void UElevenLabsConversationalAgentComponent::OnMicrophoneDataCaptured(const TAr
 {
 	if (!IsConnected() || !bIsListening) return;
 
+	// Convert this callback's samples to int16 bytes and accumulate.
+	// WASAPI fires every ~5ms (158 bytes at 16kHz). ElevenLabs needs ≥100ms
+	// (3200 bytes) per chunk for reliable VAD and STT. We hold bytes here
+	// until we have enough, then send the whole batch in one WebSocket frame.
 	TArray<uint8> PCMBytes = FloatPCMToInt16Bytes(FloatPCM);
-	WebSocketProxy->SendAudioChunk(PCMBytes);
+	MicAccumulationBuffer.Append(PCMBytes);
+
+	if (MicAccumulationBuffer.Num() >= MicChunkMinBytes)
+	{
+		WebSocketProxy->SendAudioChunk(MicAccumulationBuffer);
+		MicAccumulationBuffer.Reset();
+	}
 }
 
 TArray<uint8> UElevenLabsConversationalAgentComponent::FloatPCMToInt16Bytes(const TArray<float>& FloatPCM)

Unreal/PS_AI_Agent/Plugins/PS_AI_Agent_ElevenLabs/Source/PS_AI_Agent_ElevenLabs/Private/ElevenLabsMicrophoneCaptureComponent.cpp

@@ -119,20 +119,22 @@ void UElevenLabsMicrophoneCaptureComponent::OnAudioGenerate(
 // Resampling
 // ─────────────────────────────────────────────────────────────────────────────
 TArray<float> UElevenLabsMicrophoneCaptureComponent::ResampleTo16000(
-	const float* InAudio, int32 NumSamples,
+	const float* InAudio, int32 NumFrames,
 	int32 InChannels, int32 InSampleRate)
 {
 	const int32 TargetRate = ElevenLabsAudio::SampleRate; // 16000
 
 	// --- Step 1: Downmix to mono ---
+	// NOTE: NumFrames is the number of audio frames (not total samples).
+	// Each frame contains InChannels samples (e.g. 2 for stereo).
+	// The raw buffer has NumFrames * InChannels total float values.
 	TArray<float> Mono;
 	if (InChannels == 1)
 	{
-		Mono = TArray<float>(InAudio, NumSamples);
+		Mono = TArray<float>(InAudio, NumFrames);
 	}
 	else
 	{
-		const int32 NumFrames = NumSamples / InChannels;
 		Mono.Reserve(NumFrames);
 		for (int32 i = 0; i < NumFrames; i++)
 		{

Unreal/PS_AI_Agent/Plugins/PS_AI_Agent_ElevenLabs/Source/PS_AI_Agent_ElevenLabs/Private/ElevenLabsWebSocketProxy.cpp

@@ -72,7 +72,11 @@ void UElevenLabsWebSocketProxy::Connect(const FString& AgentIDOverride, const FS
 	WebSocket->OnConnected().AddUObject(this, &UElevenLabsWebSocketProxy::OnWsConnected);
 	WebSocket->OnConnectionError().AddUObject(this, &UElevenLabsWebSocketProxy::OnWsConnectionError);
 	WebSocket->OnClosed().AddUObject(this, &UElevenLabsWebSocketProxy::OnWsClosed);
-	WebSocket->OnMessage().AddUObject(this, &UElevenLabsWebSocketProxy::OnWsMessage);
+	// NOTE: We bind ONLY OnRawMessage (binary frames), NOT OnMessage (text frames).
+	// UE's WebSocket implementation fires BOTH callbacks for the same frame when using
+	// the libwebsockets backend — binding both causes every audio packet to be decoded
+	// and played twice. OnRawMessage handles all frame types: raw binary audio AND
+	// text-framed JSON (detected by peeking first byte for '{').
 	WebSocket->OnRawMessage().AddUObject(this, &UElevenLabsWebSocketProxy::OnWsBinaryMessage);
 
 	WebSocket->Connect();
@@ -94,36 +98,58 @@ void UElevenLabsWebSocketProxy::SendAudioChunk(const TArray<uint8>& PCMData)
 {
 	if (!IsConnected())
 	{
-		UE_LOG(LogElevenLabsWS, Warning, TEXT("SendAudioChunk: not connected."));
+		UE_LOG(LogElevenLabsWS, Warning, TEXT("SendAudioChunk: not connected (state=%d). Audio dropped."),
+			(int32)ConnectionState);
 		return;
 	}
 	if (PCMData.Num() == 0) return;
 
+	UE_LOG(LogElevenLabsWS, Log, TEXT("SendAudioChunk: %d bytes (PCM int16 LE @ 16kHz mono)"), PCMData.Num());
+
+	// Track when the last audio chunk was sent for latency measurement.
+	LastAudioChunkSentTime = FPlatformTime::Seconds();
+
 	// ElevenLabs expects: { "user_audio_chunk": "<base64 PCM>" }
+	// The server's VAD detects silence to determine end-of-turn.
+	// Do NOT send user_activity here — it resets the turn timeout timer
+	// and would prevent the server from taking the turn after the user stops speaking.
 	const FString Base64Audio = FBase64::Encode(PCMData.GetData(), PCMData.Num());
 
-	TSharedPtr<FJsonObject> Msg = MakeShareable(new FJsonObject());
-	Msg->SetStringField(ElevenLabsMessageType::AudioChunk, Base64Audio);
-	SendJsonMessage(Msg);
+	// Send as compact JSON (no pretty-printing) directly, bypassing SendJsonMessage
+	// to avoid the pretty-printed writer and to keep the payload minimal.
+	const FString AudioJson = FString::Printf(TEXT("{\"user_audio_chunk\":\"%s\"}"), *Base64Audio);
+
+	// Log first chunk fully for debugging
+	static int32 AudioChunksSent = 0;
+	AudioChunksSent++;
+	if (AudioChunksSent <= 2)
+	{
+		UE_LOG(LogElevenLabsWS, Log, TEXT("  Audio JSON (first 200 chars): %.200s"), *AudioJson);
+	}
+
+	if (WebSocket.IsValid() && WebSocket->IsConnected())
+	{
+		WebSocket->Send(AudioJson);
+	}
 }
 
 void UElevenLabsWebSocketProxy::SendUserTurnStart()
 {
-	// In client turn mode, signal that the user is active/speaking.
-	// API message: { "type": "user_activity" }
-	if (!IsConnected()) return;
-	TSharedPtr<FJsonObject> Msg = MakeShareable(new FJsonObject());
-	Msg->SetStringField(TEXT("type"), ElevenLabsMessageType::UserActivity);
-	SendJsonMessage(Msg);
+	// No-op: the ElevenLabs API does not require a "start speaking" signal.
+	// The server's VAD detects speech from the audio chunks we send.
+	// user_activity is a keep-alive/timeout-reset message and should NOT be
+	// sent here — it would delay the agent's turn after the user stops.
+	UE_LOG(LogElevenLabsWS, Log, TEXT("User turn started (audio chunks will follow)."));
 }
 
 void UElevenLabsWebSocketProxy::SendUserTurnEnd()
 {
-	// In client turn mode, stopping user_activity signals end of user turn.
-	// The API uses user_activity for ongoing speech; simply stop sending it.
-	// No explicit end message is required — silence is detected server-side.
-	// We still log for debug visibility.
-	UE_LOG(LogElevenLabsWS, Log, TEXT("User turn ended (client mode) — stopped sending user_activity."));
+	// No explicit "end turn" message exists in the ElevenLabs API.
+	// The server detects end-of-speech via VAD when we stop sending audio chunks.
+	UserTurnEndTime = FPlatformTime::Seconds();
+	bWaitingForResponse = true;
+	bFirstAudioResponseLogged = false;
+	UE_LOG(LogElevenLabsWS, Log, TEXT("User turn ended — stopped sending audio chunks. Server VAD will detect silence."));
 }
 
 void UElevenLabsWebSocketProxy::SendTextMessage(const FString& Text)
@@ -155,8 +181,79 @@ void UElevenLabsWebSocketProxy::SendInterrupt()
 // ─────────────────────────────────────────────────────────────────────────────
 void UElevenLabsWebSocketProxy::OnWsConnected()
 {
-	UE_LOG(LogElevenLabsWS, Log, TEXT("WebSocket connected. Waiting for conversation_initiation_metadata..."));
-	// State stays Connecting until we receive the initiation metadata from the server.
+	UE_LOG(LogElevenLabsWS, Log, TEXT("WebSocket connected. Sending conversation_initiation_client_data..."));
+	// State stays Connecting until we receive conversation_initiation_metadata from the server.
+
+	// ElevenLabs requires this message immediately after the WebSocket handshake to
+	// negotiate the session configuration. Without it, the server won't accept audio
+	// from the client (microphone stays silent from server perspective) and default
+	// settings are used (higher latency, no intermediate responses).
+	//
+	// Structure:
+	// {
+	//   "type": "conversation_initiation_client_data",
+	//   "conversation_config_override": {
+	//     "agent": {
+	//       "turn": { "turn_timeout": 3, "speculative_turn": true }
+	//     },
+	//     "tts": {
+	//       "optimize_streaming_latency": 3
+	//     }
+	//   },
+	//   "custom_llm_extra_body": {
+	//     "enable_intermediate_response": true
+	//   }
+	// }
+
+	// Configure turn-taking behaviour.
+	// The ElevenLabs API does NOT have a turn.mode field.
+	// Turn-taking is controlled by the server's VAD and the turn_* parameters.
+	// In push-to-talk (Client mode) the user controls the mic; the server still
+	// uses its VAD to detect the end of speech from the audio chunks it receives.
+	TSharedPtr<FJsonObject> TurnObj = MakeShareable(new FJsonObject());
+	// Lower turn_timeout so the agent responds faster after the user stops speaking.
+	// Default is 7s. In push-to-talk (Client mode), the user explicitly signals
+	// end-of-turn by releasing the key, so we can use a very short timeout (1s).
+	if (TurnMode == EElevenLabsTurnMode::Client)
+	{
+		TurnObj->SetNumberField(TEXT("turn_timeout"), 1);
+	}
+	// Speculative turn: start LLM generation during silence before the VAD is
+	// fully confident the user finished speaking. Reduces latency by 200-500ms.
+	if (bSpeculativeTurn)
+	{
+		TurnObj->SetBoolField(TEXT("speculative_turn"), true);
+	}
+
+	TSharedPtr<FJsonObject> AgentObj = MakeShareable(new FJsonObject());
+	AgentObj->SetObjectField(TEXT("turn"), TurnObj);
+
+	TSharedPtr<FJsonObject> TtsObj = MakeShareable(new FJsonObject());
+	TtsObj->SetNumberField(TEXT("optimize_streaming_latency"), 3);
+
+	TSharedPtr<FJsonObject> ConversationConfigOverride = MakeShareable(new FJsonObject());
+	ConversationConfigOverride->SetObjectField(TEXT("agent"), AgentObj);
+	ConversationConfigOverride->SetObjectField(TEXT("tts"), TtsObj);
+
+	// enable_intermediate_response reduces time-to-first-audio by allowing the agent
+	// to start speaking before it has finished generating the full response.
+	TSharedPtr<FJsonObject> CustomLlmExtraBody = MakeShareable(new FJsonObject());
+	CustomLlmExtraBody->SetBoolField(TEXT("enable_intermediate_response"), true);
+
+	TSharedPtr<FJsonObject> InitMsg = MakeShareable(new FJsonObject());
+	InitMsg->SetStringField(TEXT("type"), ElevenLabsMessageType::ConversationClientData);
+	InitMsg->SetObjectField(TEXT("conversation_config_override"), ConversationConfigOverride);
+	InitMsg->SetObjectField(TEXT("custom_llm_extra_body"), CustomLlmExtraBody);
+
+	// NOTE: We bypass SendJsonMessage() here intentionally.
+	// SendJsonMessage() guards on WebSocket->IsConnected(), but OnWsConnected fires
+	// during the handshake before IsConnected() returns true in some UE WS backends.
+	// We know the socket is open at this point — send directly.
+	FString InitJson;
+	TSharedRef<TJsonWriter<>> InitWriter = TJsonWriterFactory<>::Create(&InitJson);
+	FJsonSerializer::Serialize(InitMsg.ToSharedRef(), InitWriter);
+	UE_LOG(LogElevenLabsWS, Log, TEXT("Sending initiation: %s"), *InitJson);
+	WebSocket->Send(InitJson);
 }
 
 void UElevenLabsWebSocketProxy::OnWsConnectionError(const FString& Error)
@@ -200,20 +297,53 @@ void UElevenLabsWebSocketProxy::OnWsMessage(const FString& Message)
 		return;
 	}
 
+	// Log every message type received from the server for debugging.
+	UE_LOG(LogElevenLabsWS, Log, TEXT("Received message type: %s"), *MsgType);
+
 	if (MsgType == ElevenLabsMessageType::ConversationInitiation)
 	{
 		HandleConversationInitiation(Root);
 	}
 	else if (MsgType == ElevenLabsMessageType::AudioResponse)
 	{
+		// Log time-to-first-audio: latency between end of user turn and first agent audio.
+		if (bWaitingForResponse && !bFirstAudioResponseLogged)
+		{
+			const double Now = FPlatformTime::Seconds();
+			const double LatencyFromTurnEnd = (Now - UserTurnEndTime) * 1000.0;
+			const double LatencyFromLastChunk = (Now - LastAudioChunkSentTime) * 1000.0;
+			UE_LOG(LogElevenLabsWS, Warning,
+				TEXT("[LATENCY] Time-to-first-audio: %.0f ms (from turn end), %.0f ms (from last chunk sent)"),
+				LatencyFromTurnEnd, LatencyFromLastChunk);
+			bFirstAudioResponseLogged = true;
+		}
 		HandleAudioResponse(Root);
 	}
 	else if (MsgType == ElevenLabsMessageType::UserTranscript)
 	{
+		// Log transcription latency.
+		if (bWaitingForResponse)
+		{
+			const double Now = FPlatformTime::Seconds();
+			const double LatencyFromTurnEnd = (Now - UserTurnEndTime) * 1000.0;
+			UE_LOG(LogElevenLabsWS, Warning,
+				TEXT("[LATENCY] User transcript received: %.0f ms after turn end"),
+				LatencyFromTurnEnd);
+			bWaitingForResponse = false;
+		}
 		HandleTranscript(Root);
 	}
 	else if (MsgType == ElevenLabsMessageType::AgentResponse)
 	{
+		// Log agent text response latency.
+		if (UserTurnEndTime > 0.0)
+		{
+			const double Now = FPlatformTime::Seconds();
+			const double LatencyFromTurnEnd = (Now - UserTurnEndTime) * 1000.0;
+			UE_LOG(LogElevenLabsWS, Warning,
+				TEXT("[LATENCY] Agent text response: %.0f ms after turn end"),
+				LatencyFromTurnEnd);
+		}
 		HandleAgentResponse(Root);
 	}
 	else if (MsgType == ElevenLabsMessageType::AgentResponseCorrection)

Unreal/PS_AI_Agent/Plugins/PS_AI_Agent_ElevenLabs/Source/PS_AI_Agent_ElevenLabs/Public/ElevenLabsConversationalAgentComponent.h

@@ -80,6 +80,29 @@ public:
 	UPROPERTY(EditAnywhere, BlueprintReadWrite, Category = "ElevenLabs")
 	bool bAutoStartListening = true;
 
+	/**
+	 * Enable speculative turn: the LLM starts generating a response during
+	 * silence before the VAD is fully confident the user has finished speaking.
+	 * Reduces latency by 200-500ms but may occasionally produce premature responses.
+	 */
+	UPROPERTY(EditAnywhere, BlueprintReadWrite, Category = "ElevenLabs|Latency")
+	bool bSpeculativeTurn = true;
+
+	/**
+	 * Forward user speech transcripts (user_transcript events) to the
+	 * OnAgentTranscript delegate. Disable to reduce overhead if you don't
+	 * need to display what the user said.
+	 */
+	UPROPERTY(EditAnywhere, BlueprintReadWrite, Category = "ElevenLabs|Events")
+	bool bEnableUserTranscript = true;
+
+	/**
+	 * Forward agent text responses (agent_response events) to the
+	 * OnAgentTextResponse delegate. Disable if you only need audio output.
+	 */
+	UPROPERTY(EditAnywhere, BlueprintReadWrite, Category = "ElevenLabs|Events")
+	bool bEnableAgentTextResponse = true;
+
 	// ── Events ────────────────────────────────────────────────────────────────
 
 	UPROPERTY(BlueprintAssignable, Category = "ElevenLabs|Events")
@@ -230,4 +253,11 @@ private:
 	// consider the agent done speaking.
 	int32 SilentTickCount = 0;
 	static constexpr int32 SilenceThresholdTicks = 30; // ~0.5s at 60fps
+
+	// ── Microphone accumulation ───────────────────────────────────────────────
+	// WASAPI fires callbacks every ~5ms (158 bytes at 16kHz 16-bit mono).
+	// ElevenLabs needs at least ~100ms (3200 bytes) per chunk for reliable VAD/STT.
+	// We accumulate here and only call SendAudioChunk once enough bytes are ready.
+	TArray<uint8> MicAccumulationBuffer;
+	static constexpr int32 MicChunkMinBytes = 3200; // 100ms @ 16kHz 16-bit mono
 };

Unreal/PS_AI_Agent/Plugins/PS_AI_Agent_ElevenLabs/Source/PS_AI_Agent_ElevenLabs/Public/ElevenLabsWebSocketProxy.h

@@ -183,4 +183,22 @@ private:
 	// Accumulation buffer for multi-fragment binary WebSocket frames.
 	// ElevenLabs sends JSON as binary frames; large messages arrive in fragments.
 	TArray<uint8> BinaryFrameBuffer;
+
+	// ── Latency tracking ─────────────────────────────────────────────────────
+	// Timestamp of the last audio chunk sent (user speech).
+	double LastAudioChunkSentTime = 0.0;
+	// Timestamp when user turn ended (StopListening).
+	double UserTurnEndTime = 0.0;
+	// Whether we are waiting for the first response after user stopped speaking.
+	bool bWaitingForResponse = false;
+	// Whether we already logged the first audio response latency for this turn.
+	bool bFirstAudioResponseLogged = false;
+
+public:
+	// Set by UElevenLabsConversationalAgentComponent before calling Connect().
+	// Controls turn_timeout in conversation_initiation_client_data.
+	EElevenLabsTurnMode TurnMode = EElevenLabsTurnMode::Server;
+
+	// Speculative turn: start LLM generation during silence before full turn confidence.
+	bool bSpeculativeTurn = true;
 };