diff --git a/server/src/ws/voice.ts b/server/src/ws/voice.ts
new file mode 100644
index 0000000..f441f9d
--- /dev/null
+++ b/server/src/ws/voice.ts
@@ -0,0 +1,233 @@
+import type { ServerWebSocket } from "bun";
+import type { WebSocketData } from "./sessions.js";
+
+// ── Types ────────────────────────────────────────────────
+
+interface VoiceSession {
+  chunks: Buffer[];
+  totalBytes: number;
+  partialTimer: ReturnType<typeof setInterval> | null;
+  lastPartialOffset: number;
+}
+
+// ── State ────────────────────────────────────────────────
+
+const activeSessions = new Map<string, VoiceSession>();
+
+// ── Audio constants ──────────────────────────────────────
+
+const SAMPLE_RATE = 16_000;
+const CHANNELS = 1;
+const BITS_PER_SAMPLE = 16;
+const PARTIAL_INTERVAL_MS = 2_000;
+/** Minimum bytes before attempting first transcription (100ms of 16kHz mono 16-bit) */
+const MIN_AUDIO_BYTES = 3_200;
+
+// ── Exported handlers ────────────────────────────────────
+
+/**
+ * Start a voice session for a device. Creates a buffer and starts a
+ * periodic timer that sends accumulated audio to Groq Whisper for
+ * partial transcripts every ~2s.
+ */
+export function handleVoiceStart(
+  ws: ServerWebSocket<WebSocketData>,
+  deviceId: string,
+  groqApiKey: string
+): void {
+  // Clean up any existing session for this device
+  cleanupSession(deviceId);
+
+  const session: VoiceSession = {
+    chunks: [],
+    totalBytes: 0,
+    partialTimer: null,
+    lastPartialOffset: 0,
+  };
+
+  activeSessions.set(deviceId, session);
+
+  // Start periodic partial transcription
+  session.partialTimer = setInterval(async () => {
+    // Only transcribe if there's new audio since last partial
+    if (session.totalBytes <= session.lastPartialOffset) return;
+    if (session.totalBytes < MIN_AUDIO_BYTES) return;
+
+    try {
+      const pcm = concatChunks(session.chunks);
+      const text = await transcribeAudio(pcm, groqApiKey);
+      session.lastPartialOffset = session.totalBytes;
+
+      if (text) {
+        sendToDevice(ws, { type: "transcript_partial", text });
+      }
+    } catch (err) {
+      console.error(`[Voice] Partial transcription failed for ${deviceId}:`, err);
+    }
+  }, PARTIAL_INTERVAL_MS);
+}
+
+/**
+ * Append a base64-encoded PCM audio chunk to the session buffer.
+ */
+export function handleVoiceChunk(deviceId: string, base64Data: string): void {
+  const session = activeSessions.get(deviceId);
+  if (!session) {
+    console.warn(`[Voice] Chunk received for unknown session: ${deviceId}`);
+    return;
+  }
+
+  const decoded = Buffer.from(base64Data, "base64");
+  session.chunks.push(decoded);
+  session.totalBytes += decoded.length;
+}
+
+/**
+ * Stop the partial timer, send the complete audio to Groq for a final
+ * transcript, relay it to the device, clean up, and return the text.
+ */
+export async function handleVoiceSend(
+  ws: ServerWebSocket<WebSocketData>,
+  deviceId: string,
+  groqApiKey: string
+): Promise<string> {
+  const session = activeSessions.get(deviceId);
+  if (!session) {
+    console.warn(`[Voice] Send requested for unknown session: ${deviceId}`);
+    return "";
+  }
+
+  // Stop partial timer
+  if (session.partialTimer !== null) {
+    clearInterval(session.partialTimer);
+    session.partialTimer = null;
+  }
+
+  let transcript = "";
+
+  if (session.totalBytes >= MIN_AUDIO_BYTES) {
+    try {
+      const pcm = concatChunks(session.chunks);
+      transcript = await transcribeAudio(pcm, groqApiKey);
+    } catch (err) {
+      console.error(`[Voice] Final transcription failed for ${deviceId}:`, err);
+    }
+  }
+
+  sendToDevice(ws, { type: "transcript_final", text: transcript });
+
+  // Clean up session
+  activeSessions.delete(deviceId);
+
+  return transcript;
+}
+
+/**
+ * Cancel a voice session: stop the timer and discard all audio.
+ */
+export function handleVoiceCancel(deviceId: string): void {
+  cleanupSession(deviceId);
+}
+
+// ── Internal helpers ─────────────────────────────────────
+
+/**
+ * Concatenate all buffered chunks into a single Buffer.
+ */
+function concatChunks(chunks: Buffer[]): Buffer {
+  return Buffer.concat(chunks);
+}
+
+/**
+ * Clean up and remove a voice session.
+ */
+function cleanupSession(deviceId: string): void {
+  const session = activeSessions.get(deviceId);
+  if (!session) return;
+
+  if (session.partialTimer !== null) {
+    clearInterval(session.partialTimer);
+  }
+  activeSessions.delete(deviceId);
+}
+
+/**
+ * Wrap raw PCM data in a WAV container and send it to Groq's
+ * Whisper API for transcription. Returns the transcribed text.
+ */
+async function transcribeAudio(pcmBuffer: Buffer, apiKey: string): Promise<string> {
+  const wav = pcmToWav(pcmBuffer, SAMPLE_RATE, CHANNELS, BITS_PER_SAMPLE);
+
+  const formData = new FormData();
+  const wavBytes = new Uint8Array(wav.buffer, wav.byteOffset, wav.byteLength) as BlobPart;
+  formData.append("file", new Blob([wavBytes], { type: "audio/wav" }), "audio.wav");
+  formData.append("model", "whisper-large-v3");
+
+  const response = await fetch("https://api.groq.com/openai/v1/audio/transcriptions", {
+    method: "POST",
+    headers: {
+      Authorization: `Bearer ${apiKey}`,
+    },
+    body: formData,
+  });
+
+  if (!response.ok) {
+    const body = await response.text();
+    throw new Error(`Groq Whisper API error ${response.status}: ${body}`);
+  }
+
+  const result = (await response.json()) as { text: string };
+  return result.text ?? "";
+}
+
+/**
+ * Create a 44-byte WAV header + PCM data buffer.
+ */
+function pcmToWav(
+  pcm: Buffer,
+  sampleRate: number,
+  channels: number,
+  bitsPerSample: number
+): Buffer {
+  const byteRate = (sampleRate * channels * bitsPerSample) / 8;
+  const blockAlign = (channels * bitsPerSample) / 8;
+  const dataSize = pcm.length;
+  const headerSize = 44;
+
+  const buffer = Buffer.alloc(headerSize + dataSize);
+
+  // RIFF header
+  buffer.write("RIFF", 0);
+  buffer.writeUInt32LE(36 + dataSize, 4); // file size - 8
+  buffer.write("WAVE", 8);
+
+  // fmt subchunk
+  buffer.write("fmt ", 12);
+  buffer.writeUInt32LE(16, 16); // subchunk1 size (PCM = 16)
+  buffer.writeUInt16LE(1, 20); // audio format (PCM = 1)
+  buffer.writeUInt16LE(channels, 22);
+  buffer.writeUInt32LE(sampleRate, 24);
+  buffer.writeUInt32LE(byteRate, 28);
+  buffer.writeUInt16LE(blockAlign, 32);
+  buffer.writeUInt16LE(bitsPerSample, 34);
+
+  // data subchunk
+  buffer.write("data", 36);
+  buffer.writeUInt32LE(dataSize, 40);
+
+  // PCM data
+  pcm.copy(buffer, headerSize);
+
+  return buffer;
+}
+
+/**
+ * Send a JSON message to a device WebSocket (safe — catches send errors).
+ */
+function sendToDevice(ws: ServerWebSocket<WebSocketData>, msg: Record<string, unknown>): void {
+  try {
+    ws.send(JSON.stringify(msg));
+  } catch {
+    // device disconnected
+  }
+}