Pre-download the spacy model that kokoro needs at runtime

Dockerfile

@@ -35,6 +35,9 @@ RUN python3.11 -m pip install --no-cache-dir \
 COPY requirements.txt .
 RUN python3.11 -m pip install --no-cache-dir -r requirements.txt
 
+# Pre-download the spacy model that kokoro needs at runtime
+RUN python3.11 -m spacy download en_core_web_sm
+
 COPY . .
 
 EXPOSE 8000

config.yml

@@ -0,0 +1,14 @@
+# LLM backend: "local" or "lmstudio"
+llm:
+  backend: local        # change to "lmstudio" to use LM Studio instead
+  max_cache_tokens: 4096  # max KV-cache size per session (tokens); 0 to disable caching
+
+  system_prompt: >-
+    You are a helpful voice assistant. 
+    Keep your responses extremely concise but natural for spoken conversation.
+    Do not use markdown, bullet points, code blocks, emojis, or any formatting that doesn't work in speech.
+
+  # Settings used only when backend = "lmstudio"
+  lmstudio:
+    url: http://host.docker.internal:1234   # host.docker.internal resolves to your PC from inside Docker
+    model: ""           # leave empty to use whatever model LM Studio has loaded

docker-compose.yml

@@ -6,6 +6,11 @@ services:
     volumes:
       # Cache models on the host so they survive container rebuilds
       - huggingface-cache:/cache/huggingface
+      # Mount source so you can edit code/config without rebuilding the image
+      - ./config.yml:/app/config.yml:ro
+      - ./server:/app/server:ro
+      - ./static:/app/static:ro
+      - ./run.py:/app/run.py:ro
     deploy:
       resources:
         reservations:

requirements.txt

@@ -13,3 +13,4 @@ numpy
 soundfile
 scipy
 python-multipart
+pyyaml

server/config.py

@@ -0,0 +1,12 @@
+import pathlib
+import yaml
+
+_CONFIG_PATH = pathlib.Path(__file__).parent.parent / "config.yml"
+
+
+def load_config() -> dict:
+    with open(_CONFIG_PATH) as f:
+        return yaml.safe_load(f)
+
+
+config = load_config()

server/llm.py

@@ -1,32 +1,38 @@
+import copy
+import dataclasses
 import logging
 import threading
 from typing import AsyncIterator
 
 import torch
-from transformers import AutoModelForCausalLM, AutoTokenizer
+from transformers import AutoModelForCausalLM, AutoTokenizer, DynamicCache
 
 from server.audio_utils import split_sentences
 
 log = logging.getLogger(__name__)
 
+@dataclasses.dataclass
+class KVCacheState:
+    """Per-session KV-cache persisted across generate() calls."""
+    past_key_values: DynamicCache | None
+    cached_token_count: int
+    cached_messages: list[dict]  # snapshot of messages when cache was built
+
 
 class LLMEngine:
-    """Wraps Qwen3 for conversation generation."""
+    """Wraps Qwen3 for conversation generation with persistent KV-cache."""
 
-    SYSTEM_PROMPT = (
-        "You are a helpful voice assistant. Keep your responses concise and natural "
-        "for spoken conversation. Respond in 1-3 short sentences. "
-        "Do not use markdown, bullet points, code blocks, emojis, or any "
-        "formatting that doesn't work in speech."
-    )
-
-    def __init__(self, model: AutoModelForCausalLM, tokenizer: AutoTokenizer):
+    def __init__(self, model: AutoModelForCausalLM, tokenizer: AutoTokenizer, system_prompt: str,
+                 max_cache_tokens: int = 4096):
         self.model = model
         self.tokenizer = tokenizer
+        self.system_prompt = system_prompt
+        self.max_cache_tokens = max_cache_tokens
+        self._generate_lock = threading.Lock()
 
     def _build_inputs(self, messages: list[dict]):
         """Build input token ids using the model's chat template."""
-        chat_messages = [{"role": "system", "content": self.SYSTEM_PROMPT}]
+        chat_messages = [{"role": "system", "content": self.system_prompt}]
         for msg in messages:
             chat_messages.append({"role": msg["role"], "content": msg["content"]})
 
@@ -38,26 +44,145 @@ class LLMEngine:
         )
         return self.tokenizer(text, return_tensors="pt").to(self.model.device)
 
-    def generate(self, messages: list[dict], max_new_tokens: int = 256) -> str:
-        """Generate a complete response (blocking)."""
-        inputs = self._build_inputs(messages)
-        input_len = inputs["input_ids"].shape[1]
+    def _validate_cache(self, messages: list[dict], cache_state: KVCacheState | None) -> DynamicCache | None:
+        """Return past_key_values if the cache is valid for the given messages, else None."""
+        if cache_state is None or cache_state.past_key_values is None:
+            return None
+        if self.max_cache_tokens and cache_state.cached_token_count > self.max_cache_tokens:
+            log.info("KV-cache exceeds max size, discarding.")
+            return None
+        cached = cache_state.cached_messages
+        # The current messages must start with the cached messages as a prefix
+        if len(cached) > len(messages):
+            return None
+        for cached_msg, current_msg in zip(cached, messages):
+            if cached_msg["role"] != current_msg["role"] or cached_msg["content"] != current_msg["content"]:
+                return None
+        return cache_state.past_key_values
 
-        with torch.no_grad():
-            output_ids = self.model.generate(
-                **inputs,
-                max_new_tokens=max_new_tokens,
-                temperature=0.7,
-                top_p=0.9,
-                do_sample=True,
-                repetition_penalty=1.2,
+    def generate(
+        self,
+        messages: list[dict],
+        max_new_tokens: int = 256,
+        cache_state: KVCacheState | None = None,
+    ) -> tuple[str, KVCacheState]:
+        """Generate a complete response (blocking). Returns (response, updated_cache_state)."""
+        with self._generate_lock:
+            inputs = self._build_inputs(messages)
+            input_ids = inputs["input_ids"]
+            input_len = input_ids.shape[1]
+
+            past_kv = self._validate_cache(messages, cache_state)
+            cached_len = cache_state.cached_token_count if past_kv is not None else 0
+
+            log.info(
+                f"KV-cache: {cached_len}/{input_len} tokens cached, "
+                f"processing {input_len - cached_len} new tokens"
             )
 
-        # Decode only the generated tokens (skip prompt)
-        new_ids = output_ids[0][input_len:]
-        response = self.tokenizer.decode(new_ids, skip_special_tokens=True).strip()
-        log.info(f"LLM response: {response}")
-        return response
+            with torch.no_grad():
+                outputs = self.model.generate(
+                    input_ids=input_ids,
+                    attention_mask=inputs.get("attention_mask"),
+                    past_key_values=past_kv,
+                    max_new_tokens=max_new_tokens,
+                    temperature=0.7,
+                    top_p=0.9,
+                    do_sample=True,
+                    repetition_penalty=1.2,
+                    return_dict_in_generate=True,
+                    use_cache=True,
+                )
+
+            # Decode only the generated tokens (skip prompt)
+            new_ids = outputs.sequences[0][input_len:]
+            response = self.tokenizer.decode(new_ids, skip_special_tokens=True).strip()
+            log.info(f"LLM response: {response}")
+
+            # Build updated cache state: messages now include the assistant response
+            new_messages = copy.deepcopy(messages) + [{"role": "assistant", "content": response}]
+            new_cache = KVCacheState(
+                past_key_values=outputs.past_key_values,
+                cached_token_count=outputs.sequences.shape[1],
+                cached_messages=new_messages,
+            )
+            return response, new_cache
+
+    def trim_cache(self, cache_state: KVCacheState, messages: list[dict]) -> KVCacheState | None:
+        """Trim cache to match the actual conversation history (e.g. after barge-in)."""
+        if cache_state is None or cache_state.past_key_values is None:
+            return None
+        inputs = self._build_inputs(messages)
+        target_len = inputs["input_ids"].shape[1]
+        if target_len >= cache_state.cached_token_count:
+            return cache_state
+        cache_state.past_key_values.crop(target_len)
+        cache_state.cached_token_count = target_len
+        cache_state.cached_messages = copy.deepcopy(messages)
+        return cache_state
+
+    async def generate_sentences(
+        self,
+        messages: list[dict],
+        cancel_event: threading.Event | None = None,
+        cache_state: KVCacheState | None = None,
+    ) -> AsyncIterator[str]:
+        """Generate response and yield it sentence by sentence for TTS pipelining."""
+        import asyncio
+
+        response = await asyncio.to_thread(self.generate, messages, 256, cache_state)
+
+        if cancel_event and cancel_event.is_set():
+            return
+
+        # Split into sentences and yield each
+        sentences, remainder = split_sentences(response)
+        for sentence in sentences:
+            if cancel_event and cancel_event.is_set():
+                return
+            yield sentence
+
+        if remainder:
+            yield remainder
+
+
+class LMStudioEngine:
+    """LLM engine that delegates to an LM Studio server via its OpenAI-compatible API."""
+
+    def __init__(self, base_url: str, model: str, system_prompt: str):
+        self.base_url = base_url.rstrip("/")
+        self.model = model
+        self.system_prompt = system_prompt
+
+    def generate(
+        self,
+        messages: list[dict],
+        max_new_tokens: int = 256,
+        cache_state: KVCacheState | None = None,
+    ) -> tuple[str, None]:
+        import requests
+
+        payload_messages = [{"role": "system", "content": self.system_prompt}]
+        payload_messages.extend(messages)
+
+        body: dict = {
+            "messages": payload_messages,
+            "max_tokens": max_new_tokens,
+            "temperature": 0.7,
+            "stream": False,
+        }
+        if self.model:
+            body["model"] = self.model
+
+        resp = requests.post(
+            f"{self.base_url}/v1/chat/completions",
+            json=body,
+            timeout=30,
+        )
+        resp.raise_for_status()
+        response = resp.json()["choices"][0]["message"]["content"].strip()
+        log.info(f"LM Studio response: {response}")
+        return response, None
 
     async def generate_sentences(
         self,
@@ -72,7 +197,6 @@ class LLMEngine:
         if cancel_event and cancel_event.is_set():
             return
 
-        # Split into sentences and yield each
         sentences, remainder = split_sentences(response)
         for sentence in sentences:
             if cancel_event and cancel_event.is_set():

server/main.py

@@ -77,7 +77,9 @@ async def websocket_chat(ws: WebSocket):
                     continue
 
                 if msg.get("type") == "interrupt":
-                    await session.interrupt()
+                    await session.interrupt(
+                        last_chunk_id=msg.get("last_chunk_id")
+                    )
 
     except WebSocketDisconnect:
         log.info("WebSocket client disconnected.")

server/models.py

@@ -46,7 +46,7 @@ class ModelManager:
         from server.vad import SileroVADOnnx
 
         model_path = hf_hub_download(
-            repo_id="onnx-community/silero-vad", filename="silero_vad.onnx"
+            repo_id="onnx-community/silero-vad", filename="onnx/model.onnx"
         )
         self.vad_model = SileroVADOnnx(model_path)
         log.info("Silero VAD loaded (ONNX, CPU).")
@@ -66,19 +66,35 @@ class ModelManager:
         log.info("Qwen3-ASR-0.6B loaded.")
 
     def _load_llm(self):
-        log.info("Loading Qwen3-4B (GPTQ 4-bit)...")
-        from transformers import AutoModelForCausalLM, AutoTokenizer
+        from server.config import config
 
-        model_name = "Qwen/Qwen3.5-0.8B"
+        llm_config = config.get("llm", {})
+        backend = llm_config.get("backend", "local")
+        system_prompt = llm_config.get("system_prompt", "You are a helpful assistant.")
 
-        tokenizer = AutoTokenizer.from_pretrained(model_name)
-        device = get_device()
-        model = AutoModelForCausalLM.from_pretrained(
-            model_name,
-            device_map=device,
-        )
-        self.llm_engine = LLMEngine(model, tokenizer)
-        log.info("Qwen3-4B-GPTQ-Int4 loaded (~2.5GB VRAM).")
+        if backend == "lmstudio":
+            from server.llm import LMStudioEngine
+
+            lms = llm_config.get("lmstudio", {})
+            url = lms.get("url", "http://host.docker.internal:1234")
+            model = lms.get("model", "") or ""
+            log.info(f"Using LM Studio backend at {url} (model={model or 'server default'})")
+            self.llm_engine = LMStudioEngine(url, model, system_prompt)
+        else:
+            log.info("Loading Qwen3-4B (GPTQ 4-bit)...")
+            from transformers import AutoModelForCausalLM, AutoTokenizer
+
+            model_name = "Qwen/Qwen3.5-0.8B"
+            
+            tokenizer = AutoTokenizer.from_pretrained(model_name)
+            device = get_device()
+            model = AutoModelForCausalLM.from_pretrained(
+                model_name,
+                device_map=device,
+            )
+            max_cache_tokens = llm_config.get("max_cache_tokens", 4096)
+            self.llm_engine = LLMEngine(model, tokenizer, system_prompt, max_cache_tokens)
+            log.info("Qwen3-4B-GPTQ-Int4 loaded (~2.5GB VRAM).")
 
     def _load_tts(self):
         log.info("Loading Kokoro TTS...")

server/pipeline.py

@@ -1,11 +1,13 @@
 import asyncio
 import logging
 import queue
+import re
 import threading
 
 import numpy as np
 
 from server.audio_utils import float32_to_pcm_bytes
+from server.llm import KVCacheState
 from server.models import ModelManager
 from server.vad import StreamingVAD
 
@@ -13,6 +15,56 @@ log = logging.getLogger(__name__)
 
 _SENTINEL = None
 
+# Regex: split after sentence-ending punctuation followed by whitespace
+_SENTENCE_RE = re.compile(r'(?<=[.!?])\s+')
+# Regex: split after clause-level punctuation followed by whitespace
+_CLAUSE_RE = re.compile(r'(?<=[,;:\u2014])\s+')
+
+MAX_SEGMENT_WORDS = 20
+MIN_SEGMENT_WORDS = 4
+
+
+def _split_into_segments(text: str) -> list[str]:
+    """Split text into small TTS-friendly segments for fine-grained streaming.
+
+    Splits on sentence boundaries first, then breaks long sentences at clause
+    boundaries (commas, semicolons, colons, em-dashes).  Avoids tiny fragments
+    by merging short pieces with their neighbours.
+    """
+    sentences = _SENTENCE_RE.split(text.strip())
+    segments: list[str] = []
+    for sent in sentences:
+        if len(sent.split()) <= MAX_SEGMENT_WORDS:
+            segments.append(sent)
+        else:
+            # Split long sentences at clause boundaries
+            clauses = _CLAUSE_RE.split(sent)
+            current = ""
+            for clause in clauses:
+                combined = (current + " " + clause) if current else clause
+                if current and len(combined.split()) > MAX_SEGMENT_WORDS:
+                    segments.append(current)
+                    current = clause
+                else:
+                    current = combined
+            if current:
+                segments.append(current)
+
+    # Merge any tiny fragments into their neighbour
+    merged: list[str] = []
+    for seg in segments:
+        if not seg.strip():
+            continue
+        if merged and len(merged[-1].split()) < MIN_SEGMENT_WORDS:
+            merged[-1] = merged[-1] + " " + seg
+        else:
+            merged.append(seg)
+    # Also merge a trailing runt
+    if len(merged) > 1 and len(merged[-1].split()) < MIN_SEGMENT_WORDS:
+        merged[-2] = merged[-2] + " " + merged[-1]
+        merged.pop()
+    return merged
+
 
 class ConversationSession:
     """Manages a single client's voice conversation pipeline.
@@ -27,6 +79,7 @@ class ConversationSession:
 
         self.vad: StreamingVAD = models.create_vad()
         self.conversation_history: list[dict] = []
+        self.kv_cache_state: KVCacheState | None = None
         self.cancel_event = threading.Event()
         self.is_responding = False
         self._response_task: asyncio.Task | None = None
@@ -38,6 +91,7 @@ class ConversationSession:
         self.cancel_event.set()
         if self._response_task and not self._response_task.done():
             self._response_task.cancel()
+        self.kv_cache_state = None
 
     async def handle_audio_chunk(self, chunk_16k: np.ndarray):
         utterance = self.vad.process_chunk(chunk_16k)
@@ -50,15 +104,17 @@ class ConversationSession:
         elif self.vad.is_speaking and self.is_responding:
             await self._interrupt()
 
-    async def interrupt(self):
+    async def interrupt(self, last_chunk_id: int | None = None):
         """Public interrupt method for WebSocket text messages."""
         if self.is_responding:
-            await self._interrupt()
+            await self._interrupt(last_chunk_id=last_chunk_id)
 
-    async def _interrupt(self):
+    async def _interrupt(self, last_chunk_id: int | None = None):
         log.info("Barge-in: cancelling response.")
         self.cancel_event.set()
         self.is_responding = False
+        if last_chunk_id is not None:
+            self._last_played_chunk_id = last_chunk_id
         # Tell client to stop audio immediately
         try:
             await self.send_json({"type": "interrupt"})
@@ -91,8 +147,8 @@ class ConversationSession:
         # LLM
         log.info(f"Conversation history ({len(self.conversation_history)} messages): "
                  + str([m['content'][:50] for m in self.conversation_history]))
-        response = await asyncio.to_thread(
-            self.models.llm_engine.generate, self.conversation_history
+        response, self.kv_cache_state = await asyncio.to_thread(
+            self.models.llm_engine.generate, self.conversation_history, 256, self.kv_cache_state
         )
 
         if self.cancel_event.is_set():
@@ -102,16 +158,23 @@ class ConversationSession:
         # TTS - stream chunks with per-sentence text
         await self.send_json({"type": "status", "state": "speaking"})
         chunk_queue = queue.Queue()
+        self._last_played_chunk_id = None
+
+        segments = _split_into_segments(response)
+        log.info(f"TTS: split response into {len(segments)} segments")
 
         def _tts_worker():
             try:
-                for graphemes, _ps, audio in self.models.tts_engine.pipeline(
-                    response, voice=self.models.tts_engine.voice
-                ):
+                for segment in segments:
                     if self.cancel_event.is_set():
                         break
-                    if audio is not None and len(audio) > 0:
-                        chunk_queue.put((graphemes, audio))
+                    for graphemes, _ps, audio in self.models.tts_engine.pipeline(
+                        segment, voice=self.models.tts_engine.voice
+                    ):
+                        if self.cancel_event.is_set():
+                            break
+                        if audio is not None and len(audio) > 0:
+                            chunk_queue.put((graphemes, audio))
             except Exception:
                 log.exception("TTS generation error")
             finally:
@@ -121,6 +184,9 @@ class ConversationSession:
         tts_thread.start()
 
         spoken_text = ""
+        chunk_id = 0
+        # Maps chunk_id -> cumulative text up to and including that chunk
+        chunk_text_map: dict[int, str] = {}
         while True:
             try:
                 item = await asyncio.to_thread(chunk_queue.get, timeout=10.0)
@@ -134,8 +200,14 @@ class ConversationSession:
 
             sentence_text, audio = item
             spoken_text += sentence_text
+            chunk_text_map[chunk_id] = spoken_text
 
-            await self.send_json({"type": "response_text", "text": sentence_text, "final": False})
+            await self.send_json({
+                "type": "response_text",
+                "text": sentence_text,
+                "chunk_id": chunk_id,
+                "final": False,
+            })
             pcm_bytes = float32_to_pcm_bytes(audio)
             try:
                 await self.send_bytes(pcm_bytes)
@@ -143,19 +215,42 @@ class ConversationSession:
                 log.warning("Failed to send audio, client disconnected.")
                 self.cancel_event.set()
                 break
+            chunk_id += 1
 
         tts_thread.join(timeout=2.0)
 
-        # Save only what was actually spoken
-        if spoken_text.strip():
+        # Determine what was actually heard by the client
+        was_interrupted = spoken_text.strip() != response.strip()
+        if was_interrupted and self._last_played_chunk_id is not None:
+            # Client told us the last chunk whose audio actually played
+            heard_text = chunk_text_map.get(self._last_played_chunk_id, "")
+            log.info(f"Interrupted: client heard up to chunk {self._last_played_chunk_id}")
+        else:
+            heard_text = spoken_text
+
+        # Save only what was actually spoken/heard
+        if heard_text.strip():
+            # Use original LLM response when fully spoken (keeps KV-cache valid);
+            # use heard_text only when interrupted.
+            final_content = heard_text.strip() if was_interrupted else response
             self.conversation_history.append(
-                {"role": "assistant", "content": spoken_text.strip()}
+                {"role": "assistant", "content": final_content}
             )
+            if was_interrupted and self.kv_cache_state is not None:
+                self.kv_cache_state = self.models.llm_engine.trim_cache(
+                    self.kv_cache_state, self.conversation_history
+                )
         elif self.conversation_history and self.conversation_history[-1]["role"] == "user":
             self.conversation_history.pop()
+            self.kv_cache_state = None
 
         if not self.cancel_event.is_set():
-            await self.send_json({"type": "response_text", "text": "", "final": True})
+            await self.send_json({
+                "type": "response_text",
+                "text": "",
+                "final": True,
+                "total_chunks": chunk_id,
+            })
 
         self.is_responding = False
         try:

static/app.js

@@ -13,6 +13,11 @@ const BARGE_IN_THRESHOLD = 0.03; // RMS energy threshold for barge-in
 const BARGE_IN_FRAMES = 2; // Consecutive frames above threshold to trigger
 let bargeInCount = 0;
 
+// --- Text-audio sync state ---
+let pendingTextChunks = []; // [{chunkId, text}] - text waiting for its audio to arrive
+let scheduledTextTimers = []; // timer IDs for text display scheduled to match audio playback
+let lastDisplayedChunkId = -1; // last chunk whose text was actually shown to the user
+
 const chatArea = document.getElementById("chat-area");
 const statusBadge = document.getElementById("status-badge");
 const micBtn = document.getElementById("mic-btn");
@@ -54,8 +59,8 @@ function handleJSON(msg) {
 
     case "interrupt":
       stopPlayback();
-      // Trim the assistant message to what was spoken, then finalize
-      finalizeAssistantMessage();
+      // Finalize with interrupted marker — text already reflects only what was heard
+      finalizeAssistantMessage(true);
       break;
 
     case "transcript":
@@ -64,9 +69,15 @@ function handleJSON(msg) {
 
     case "response_text":
       if (msg.final) {
-        finalizeAssistantMessage();
+        // All chunks sent; finalize will happen when last audio chunk plays
+        // (or immediately if nothing was queued)
+        if (pendingTextChunks.length === 0 && scheduledTextTimers.length === 0) {
+          finalizeAssistantMessage(false);
+        }
+        // Otherwise, playAudioChunk will finalize after the last scheduled text
       } else {
-        appendAssistantText(msg.text);
+        // Queue text — it will be displayed when corresponding audio starts playing
+        pendingTextChunks.push({ chunkId: msg.chunk_id, text: msg.text });
       }
       break;
   }
@@ -113,9 +124,20 @@ function appendAssistantText(text) {
   chatArea.scrollTop = chatArea.scrollHeight;
 }
 
-function finalizeAssistantMessage() {
+function finalizeAssistantMessage(interrupted = false) {
+  if (interrupted && currentAssistantEl && currentAssistantText) {
+    const marker = document.createElement("span");
+    marker.className = "interrupted-marker";
+    marker.textContent = " [interrupted]";
+    currentAssistantEl.appendChild(marker);
+  }
   currentAssistantEl = null;
   currentAssistantText = "";
+  // Reset sync state
+  pendingTextChunks = [];
+  for (const tid of scheduledTextTimers) clearTimeout(tid);
+  scheduledTextTimers = [];
+  lastDisplayedChunkId = -1;
 }
 
 // --- Audio Playback ---
@@ -146,18 +168,38 @@ function playAudioChunk(arrayBuffer) {
 
   activeSources.push(source);
   isPlaying = true;
-  source.onended = () => {
-    activeSources = activeSources.filter((s) => s !== source);
-    if (activeSources.length === 0) {
-      isPlaying = false;
-      bargeInCount = 0;
-    }
-  };
+
+  // Pair this audio chunk with the next queued text chunk
+  const textEntry = pendingTextChunks.shift();
 
   const now = ctx.currentTime;
   if (nextPlayTime < now) {
     nextPlayTime = now + 0.01;
   }
+
+  // Schedule text display to coincide with audio playback start
+  if (textEntry) {
+    const delayMs = Math.max(0, (nextPlayTime - now) * 1000);
+    const tid = setTimeout(() => {
+      appendAssistantText(textEntry.text);
+      lastDisplayedChunkId = textEntry.chunkId;
+      scheduledTextTimers = scheduledTextTimers.filter((t) => t !== tid);
+    }, delayMs);
+    scheduledTextTimers.push(tid);
+  }
+
+  source.onended = () => {
+    activeSources = activeSources.filter((s) => s !== source);
+    if (activeSources.length === 0) {
+      isPlaying = false;
+      bargeInCount = 0;
+      // If all audio has finished and no more text pending, finalize
+      if (pendingTextChunks.length === 0 && scheduledTextTimers.length === 0) {
+        finalizeAssistantMessage(false);
+      }
+    }
+  };
+
   source.start(nextPlayTime);
   nextPlayTime += buffer.duration;
 }
@@ -172,6 +214,10 @@ function stopPlayback() {
   nextPlayTime = 0;
   isPlaying = false;
   bargeInCount = 0;
+  // Cancel any pending text displays
+  for (const tid of scheduledTextTimers) clearTimeout(tid);
+  scheduledTextTimers = [];
+  pendingTextChunks = [];
 }
 
 // --- Microphone ---
@@ -229,8 +275,12 @@ async function startMic() {
             if (bargeInCount >= BARGE_IN_FRAMES) {
               // User is speaking over the assistant - interrupt
               stopPlayback();
-              finalizeAssistantMessage();
-              ws.send(JSON.stringify({ type: "interrupt" }));
+              const msg = { type: "interrupt" };
+              if (lastDisplayedChunkId >= 0) {
+                msg.last_chunk_id = lastDisplayedChunkId;
+              }
+              ws.send(JSON.stringify(msg));
+              finalizeAssistantMessage(true);
               isPlaying = false;
               bargeInCount = 0;
             }

static/style.css

@@ -84,6 +84,12 @@ header h1 {
   border-bottom-left-radius: 4px;
 }
 
+.interrupted-marker {
+  color: #888;
+  font-style: italic;
+  font-size: 13px;
+}
+
 #controls {
   padding: 16px 24px;
   border-top: 1px solid #222;