diff --git a/apps/api/src/speech/speech.gateway.spec.ts b/apps/api/src/speech/speech.gateway.spec.ts
new file mode 100644
index 0000000..dac50e1
--- /dev/null
+++ b/apps/api/src/speech/speech.gateway.spec.ts
@@ -0,0 +1,683 @@
+/**
+ * SpeechGateway Tests
+ *
+ * Issue #397: WebSocket streaming transcription endpoint tests.
+ * Written FIRST following TDD (Red-Green-Refactor).
+ *
+ * Tests cover:
+ * - Authentication via handshake token
+ * - Session lifecycle: start -> audio chunks -> stop
+ * - Transcription result emission
+ * - Session cleanup on disconnect
+ * - Error handling
+ * - Buffer size limit enforcement
+ */
+
+import { describe, it, expect, beforeEach, vi } from "vitest";
+import { Socket } from "socket.io";
+import { SpeechGateway } from "./speech.gateway";
+import { SpeechService } from "./speech.service";
+import { AuthService } from "../auth/auth.service";
+import { PrismaService } from "../prisma/prisma.service";
+import type { SpeechConfig } from "./speech.config";
+import type { TranscriptionResult } from "./interfaces/speech-types";
+
+// ==========================================
+// Test helpers
+// ==========================================
+
+interface AuthenticatedSocket extends Socket {
+  data: {
+    userId?: string;
+    workspaceId?: string;
+  };
+}
+
+function createMockConfig(): SpeechConfig {
+  return {
+    stt: {
+      enabled: true,
+      baseUrl: "http://localhost:8000/v1",
+      model: "test-model",
+      language: "en",
+    },
+    tts: {
+      default: { enabled: true, url: "http://localhost:8880/v1", voice: "test", format: "mp3" },
+      premium: { enabled: false, url: "" },
+      fallback: { enabled: false, url: "" },
+    },
+    limits: {
+      maxUploadSize: 25_000_000,
+      maxDurationSeconds: 600,
+      maxTextLength: 4096,
+    },
+  };
+}
+
+function createMockSocket(overrides?: Partial<AuthenticatedSocket>): AuthenticatedSocket {
+  return {
+    id: "test-socket-id",
+    join: vi.fn(),
+    leave: vi.fn(),
+    emit: vi.fn(),
+    disconnect: vi.fn(),
+    data: {},
+    handshake: {
+      auth: { token: "valid-token" },
+      query: {},
+      headers: {},
+    },
+    ...overrides,
+  } as unknown as AuthenticatedSocket;
+}
+
+function createMockAuthService(): {
+  verifySession: ReturnType<typeof vi.fn>;
+} {
+  return {
+    verifySession: vi.fn().mockResolvedValue({
+      user: { id: "user-123" },
+      session: { id: "session-123" },
+    }),
+  };
+}
+
+function createMockPrismaService(): {
+  workspaceMember: { findFirst: ReturnType<typeof vi.fn> };
+} {
+  return {
+    workspaceMember: {
+      findFirst: vi.fn().mockResolvedValue({
+        userId: "user-123",
+        workspaceId: "workspace-456",
+        role: "MEMBER",
+      }),
+    },
+  };
+}
+
+function createMockSpeechService(): {
+  transcribe: ReturnType<typeof vi.fn>;
+  isSTTAvailable: ReturnType<typeof vi.fn>;
+} {
+  return {
+    transcribe: vi.fn().mockResolvedValue({
+      text: "Hello world",
+      language: "en",
+      durationSeconds: 2.5,
+    } satisfies TranscriptionResult),
+    isSTTAvailable: vi.fn().mockReturnValue(true),
+  };
+}
+
+// ==========================================
+// Tests
+// ==========================================
+
+describe("SpeechGateway", () => {
+  let gateway: SpeechGateway;
+  let mockAuthService: ReturnType<typeof createMockAuthService>;
+  let mockPrismaService: ReturnType<typeof createMockPrismaService>;
+  let mockSpeechService: ReturnType<typeof createMockSpeechService>;
+  let mockConfig: SpeechConfig;
+  let mockClient: AuthenticatedSocket;
+
+  beforeEach(() => {
+    mockAuthService = createMockAuthService();
+    mockPrismaService = createMockPrismaService();
+    mockSpeechService = createMockSpeechService();
+    mockConfig = createMockConfig();
+    mockClient = createMockSocket();
+
+    gateway = new SpeechGateway(
+      mockAuthService as unknown as AuthService,
+      mockPrismaService as unknown as PrismaService,
+      mockSpeechService as unknown as SpeechService,
+      mockConfig
+    );
+
+    vi.clearAllMocks();
+  });
+
+  // ==========================================
+  // Authentication
+  // ==========================================
+  describe("handleConnection", () => {
+    it("should authenticate client and populate socket data on valid token", async () => {
+      mockAuthService.verifySession.mockResolvedValue({
+        user: { id: "user-123" },
+        session: { id: "session-123" },
+      });
+      mockPrismaService.workspaceMember.findFirst.mockResolvedValue({
+        userId: "user-123",
+        workspaceId: "workspace-456",
+        role: "MEMBER",
+      });
+
+      await gateway.handleConnection(mockClient);
+
+      expect(mockAuthService.verifySession).toHaveBeenCalledWith("valid-token");
+      expect(mockClient.data.userId).toBe("user-123");
+      expect(mockClient.data.workspaceId).toBe("workspace-456");
+    });
+
+    it("should disconnect client without token", async () => {
+      const clientNoToken = createMockSocket({
+        handshake: { auth: {}, query: {}, headers: {} },
+      } as Partial<AuthenticatedSocket>);
+
+      await gateway.handleConnection(clientNoToken);
+
+      expect(clientNoToken.disconnect).toHaveBeenCalled();
+    });
+
+    it("should disconnect client with invalid token", async () => {
+      mockAuthService.verifySession.mockResolvedValue(null);
+
+      await gateway.handleConnection(mockClient);
+
+      expect(mockClient.disconnect).toHaveBeenCalled();
+    });
+
+    it("should disconnect client without workspace access", async () => {
+      mockAuthService.verifySession.mockResolvedValue({
+        user: { id: "user-123" },
+        session: { id: "session-123" },
+      });
+      mockPrismaService.workspaceMember.findFirst.mockResolvedValue(null);
+
+      await gateway.handleConnection(mockClient);
+
+      expect(mockClient.disconnect).toHaveBeenCalled();
+    });
+
+    it("should disconnect client when auth throws", async () => {
+      mockAuthService.verifySession.mockRejectedValue(new Error("Auth failure"));
+
+      await gateway.handleConnection(mockClient);
+
+      expect(mockClient.disconnect).toHaveBeenCalled();
+    });
+
+    it("should extract token from handshake.query as fallback", async () => {
+      const clientQueryToken = createMockSocket({
+        handshake: {
+          auth: {},
+          query: { token: "query-token" },
+          headers: {},
+        },
+      } as Partial<AuthenticatedSocket>);
+
+      mockAuthService.verifySession.mockResolvedValue({
+        user: { id: "user-123" },
+        session: { id: "session-123" },
+      });
+      mockPrismaService.workspaceMember.findFirst.mockResolvedValue({
+        userId: "user-123",
+        workspaceId: "workspace-456",
+        role: "MEMBER",
+      });
+
+      await gateway.handleConnection(clientQueryToken);
+
+      expect(mockAuthService.verifySession).toHaveBeenCalledWith("query-token");
+    });
+  });
+
+  // ==========================================
+  // start-transcription
+  // ==========================================
+  describe("handleStartTranscription", () => {
+    beforeEach(async () => {
+      mockAuthService.verifySession.mockResolvedValue({
+        user: { id: "user-123" },
+        session: { id: "session-123" },
+      });
+      mockPrismaService.workspaceMember.findFirst.mockResolvedValue({
+        userId: "user-123",
+        workspaceId: "workspace-456",
+        role: "MEMBER",
+      });
+      await gateway.handleConnection(mockClient);
+      vi.clearAllMocks();
+    });
+
+    it("should create a transcription session", () => {
+      gateway.handleStartTranscription(mockClient, { language: "en" });
+
+      expect(mockClient.emit).toHaveBeenCalledWith(
+        "transcription-started",
+        expect.objectContaining({ sessionId: expect.any(String) })
+      );
+    });
+
+    it("should create a session with optional language parameter", () => {
+      gateway.handleStartTranscription(mockClient, { language: "fr" });
+
+      expect(mockClient.emit).toHaveBeenCalledWith(
+        "transcription-started",
+        expect.objectContaining({ sessionId: expect.any(String) })
+      );
+    });
+
+    it("should create a session with no options", () => {
+      gateway.handleStartTranscription(mockClient, {});
+
+      expect(mockClient.emit).toHaveBeenCalledWith(
+        "transcription-started",
+        expect.objectContaining({ sessionId: expect.any(String) })
+      );
+    });
+
+    it("should emit error if client is not authenticated", () => {
+      const unauthClient = createMockSocket();
+      // Not connected through handleConnection, so no userId set
+
+      gateway.handleStartTranscription(unauthClient, {});
+
+      expect(unauthClient.emit).toHaveBeenCalledWith(
+        "transcription-error",
+        expect.objectContaining({ message: expect.any(String) })
+      );
+    });
+
+    it("should replace existing session if one already exists", () => {
+      gateway.handleStartTranscription(mockClient, {});
+      gateway.handleStartTranscription(mockClient, { language: "de" });
+
+      // Should have emitted transcription-started twice (no error)
+      const startedCalls = (mockClient.emit as ReturnType<typeof vi.fn>).mock.calls.filter(
+        (call: unknown[]) => call[0] === "transcription-started"
+      );
+      expect(startedCalls).toHaveLength(2);
+    });
+  });
+
+  // ==========================================
+  // audio-chunk
+  // ==========================================
+  describe("handleAudioChunk", () => {
+    beforeEach(async () => {
+      mockAuthService.verifySession.mockResolvedValue({
+        user: { id: "user-123" },
+        session: { id: "session-123" },
+      });
+      mockPrismaService.workspaceMember.findFirst.mockResolvedValue({
+        userId: "user-123",
+        workspaceId: "workspace-456",
+        role: "MEMBER",
+      });
+      await gateway.handleConnection(mockClient);
+      vi.clearAllMocks();
+      gateway.handleStartTranscription(mockClient, {});
+      vi.clearAllMocks();
+    });
+
+    it("should accumulate audio data in the session", () => {
+      const chunk = Buffer.from("audio-data-1");
+      gateway.handleAudioChunk(mockClient, chunk);
+
+      // No error emitted
+      const errorCalls = (mockClient.emit as ReturnType<typeof vi.fn>).mock.calls.filter(
+        (call: unknown[]) => call[0] === "transcription-error"
+      );
+      expect(errorCalls).toHaveLength(0);
+    });
+
+    it("should accept Uint8Array data and convert to Buffer", () => {
+      const chunk = new Uint8Array([1, 2, 3, 4]);
+      gateway.handleAudioChunk(mockClient, chunk);
+
+      const errorCalls = (mockClient.emit as ReturnType<typeof vi.fn>).mock.calls.filter(
+        (call: unknown[]) => call[0] === "transcription-error"
+      );
+      expect(errorCalls).toHaveLength(0);
+    });
+
+    it("should emit error if no active session exists", () => {
+      const noSessionClient = createMockSocket({ id: "no-session" });
+      noSessionClient.data = { userId: "user-123", workspaceId: "workspace-456" };
+
+      const chunk = Buffer.from("audio-data");
+      gateway.handleAudioChunk(noSessionClient, chunk);
+
+      expect(noSessionClient.emit).toHaveBeenCalledWith(
+        "transcription-error",
+        expect.objectContaining({ message: expect.any(String) })
+      );
+    });
+
+    it("should emit error if client is not authenticated", () => {
+      const unauthClient = createMockSocket({ id: "unauth" });
+      // Not authenticated
+
+      const chunk = Buffer.from("audio-data");
+      gateway.handleAudioChunk(unauthClient, chunk);
+
+      expect(unauthClient.emit).toHaveBeenCalledWith(
+        "transcription-error",
+        expect.objectContaining({ message: expect.any(String) })
+      );
+    });
+
+    it("should emit error when buffer size exceeds max upload size", () => {
+      // Set a very small max upload size
+      const smallConfig = createMockConfig();
+      smallConfig.limits.maxUploadSize = 10;
+
+      const limitedGateway = new SpeechGateway(
+        mockAuthService as unknown as AuthService,
+        mockPrismaService as unknown as PrismaService,
+        mockSpeechService as unknown as SpeechService,
+        smallConfig
+      );
+
+      // We need to manually set up the authenticated client in the new gateway
+      const limitedClient = createMockSocket({ id: "limited-client" });
+      limitedClient.data = { userId: "user-123", workspaceId: "workspace-456" };
+
+      // Start session directly (since handleConnection populates data)
+      limitedGateway.handleStartTranscription(limitedClient, {});
+      vi.clearAllMocks();
+
+      // Send a chunk that exceeds the limit
+      const largeChunk = Buffer.alloc(20, "a");
+      limitedGateway.handleAudioChunk(limitedClient, largeChunk);
+
+      expect(limitedClient.emit).toHaveBeenCalledWith(
+        "transcription-error",
+        expect.objectContaining({ message: expect.stringContaining("exceeds") })
+      );
+    });
+
+    it("should emit error when accumulated buffer size exceeds max upload size", () => {
+      const smallConfig = createMockConfig();
+      smallConfig.limits.maxUploadSize = 15;
+
+      const limitedGateway = new SpeechGateway(
+        mockAuthService as unknown as AuthService,
+        mockPrismaService as unknown as PrismaService,
+        mockSpeechService as unknown as SpeechService,
+        smallConfig
+      );
+
+      const limitedClient = createMockSocket({ id: "limited-client-2" });
+      limitedClient.data = { userId: "user-123", workspaceId: "workspace-456" };
+
+      limitedGateway.handleStartTranscription(limitedClient, {});
+      vi.clearAllMocks();
+
+      // Send two chunks that together exceed the limit
+      const chunk1 = Buffer.alloc(10, "a");
+      const chunk2 = Buffer.alloc(10, "b");
+      limitedGateway.handleAudioChunk(limitedClient, chunk1);
+      limitedGateway.handleAudioChunk(limitedClient, chunk2);
+
+      expect(limitedClient.emit).toHaveBeenCalledWith(
+        "transcription-error",
+        expect.objectContaining({ message: expect.stringContaining("exceeds") })
+      );
+    });
+  });
+
+  // ==========================================
+  // stop-transcription
+  // ==========================================
+  describe("handleStopTranscription", () => {
+    beforeEach(async () => {
+      mockAuthService.verifySession.mockResolvedValue({
+        user: { id: "user-123" },
+        session: { id: "session-123" },
+      });
+      mockPrismaService.workspaceMember.findFirst.mockResolvedValue({
+        userId: "user-123",
+        workspaceId: "workspace-456",
+        role: "MEMBER",
+      });
+      await gateway.handleConnection(mockClient);
+      vi.clearAllMocks();
+    });
+
+    it("should transcribe accumulated audio and emit final result", async () => {
+      gateway.handleStartTranscription(mockClient, { language: "en" });
+
+      const chunk1 = Buffer.from("audio-part-1");
+      const chunk2 = Buffer.from("audio-part-2");
+      gateway.handleAudioChunk(mockClient, chunk1);
+      gateway.handleAudioChunk(mockClient, chunk2);
+
+      vi.clearAllMocks();
+
+      const expectedResult: TranscriptionResult = {
+        text: "Hello world",
+        language: "en",
+        durationSeconds: 2.5,
+      };
+      mockSpeechService.transcribe.mockResolvedValue(expectedResult);
+
+      await gateway.handleStopTranscription(mockClient);
+
+      // Should have called transcribe with concatenated buffer
+      expect(mockSpeechService.transcribe).toHaveBeenCalledWith(
+        expect.any(Buffer),
+        expect.objectContaining({})
+      );
+
+      // Should emit transcription-final
+      expect(mockClient.emit).toHaveBeenCalledWith(
+        "transcription-final",
+        expect.objectContaining({ text: "Hello world" })
+      );
+    });
+
+    it("should pass language option to SpeechService.transcribe", async () => {
+      gateway.handleStartTranscription(mockClient, { language: "fr" });
+      gateway.handleAudioChunk(mockClient, Buffer.from("audio"));
+
+      vi.clearAllMocks();
+
+      await gateway.handleStopTranscription(mockClient);
+
+      expect(mockSpeechService.transcribe).toHaveBeenCalledWith(
+        expect.any(Buffer),
+        expect.objectContaining({ language: "fr" })
+      );
+    });
+
+    it("should clean up session after stop", async () => {
+      gateway.handleStartTranscription(mockClient, {});
+      gateway.handleAudioChunk(mockClient, Buffer.from("audio"));
+
+      await gateway.handleStopTranscription(mockClient);
+
+      vi.clearAllMocks();
+
+      // Sending more audio after stop should error (no session)
+      gateway.handleAudioChunk(mockClient, Buffer.from("more-audio"));
+
+      expect(mockClient.emit).toHaveBeenCalledWith(
+        "transcription-error",
+        expect.objectContaining({ message: expect.any(String) })
+      );
+    });
+
+    it("should emit transcription-error when transcription fails", async () => {
+      gateway.handleStartTranscription(mockClient, {});
+      gateway.handleAudioChunk(mockClient, Buffer.from("audio"));
+
+      vi.clearAllMocks();
+
+      mockSpeechService.transcribe.mockRejectedValue(new Error("STT service down"));
+
+      await gateway.handleStopTranscription(mockClient);
+
+      expect(mockClient.emit).toHaveBeenCalledWith(
+        "transcription-error",
+        expect.objectContaining({ message: expect.stringContaining("STT service down") })
+      );
+    });
+
+    it("should emit error if no active session exists", async () => {
+      await gateway.handleStopTranscription(mockClient);
+
+      expect(mockClient.emit).toHaveBeenCalledWith(
+        "transcription-error",
+        expect.objectContaining({ message: expect.any(String) })
+      );
+    });
+
+    it("should emit error if client is not authenticated", async () => {
+      const unauthClient = createMockSocket({ id: "unauth-stop" });
+
+      await gateway.handleStopTranscription(unauthClient);
+
+      expect(unauthClient.emit).toHaveBeenCalledWith(
+        "transcription-error",
+        expect.objectContaining({ message: expect.any(String) })
+      );
+    });
+
+    it("should emit error when stopping with no audio chunks received", async () => {
+      gateway.handleStartTranscription(mockClient, {});
+
+      vi.clearAllMocks();
+
+      await gateway.handleStopTranscription(mockClient);
+
+      expect(mockClient.emit).toHaveBeenCalledWith(
+        "transcription-error",
+        expect.objectContaining({ message: expect.stringContaining("No audio") })
+      );
+    });
+  });
+
+  // ==========================================
+  // handleDisconnect
+  // ==========================================
+  describe("handleDisconnect", () => {
+    beforeEach(async () => {
+      mockAuthService.verifySession.mockResolvedValue({
+        user: { id: "user-123" },
+        session: { id: "session-123" },
+      });
+      mockPrismaService.workspaceMember.findFirst.mockResolvedValue({
+        userId: "user-123",
+        workspaceId: "workspace-456",
+        role: "MEMBER",
+      });
+      await gateway.handleConnection(mockClient);
+      vi.clearAllMocks();
+    });
+
+    it("should clean up active session on disconnect", () => {
+      gateway.handleStartTranscription(mockClient, {});
+      gateway.handleAudioChunk(mockClient, Buffer.from("audio"));
+
+      gateway.handleDisconnect(mockClient);
+
+      // Session should be gone. Verify by trying to add a chunk to a new
+      // socket with the same ID (should error since session was cleaned up).
+      const newClient = createMockSocket({ id: mockClient.id });
+      newClient.data = { userId: "user-123", workspaceId: "workspace-456" };
+
+      gateway.handleAudioChunk(newClient, Buffer.from("more"));
+
+      expect(newClient.emit).toHaveBeenCalledWith(
+        "transcription-error",
+        expect.objectContaining({ message: expect.any(String) })
+      );
+    });
+
+    it("should not throw when disconnecting client without active session", () => {
+      expect(() => gateway.handleDisconnect(mockClient)).not.toThrow();
+    });
+
+    it("should not throw when disconnecting unauthenticated client", () => {
+      const unauthClient = createMockSocket({ id: "unauth-disconnect" });
+      expect(() => gateway.handleDisconnect(unauthClient)).not.toThrow();
+    });
+  });
+
+  // ==========================================
+  // Edge cases
+  // ==========================================
+  describe("edge cases", () => {
+    beforeEach(async () => {
+      mockAuthService.verifySession.mockResolvedValue({
+        user: { id: "user-123" },
+        session: { id: "session-123" },
+      });
+      mockPrismaService.workspaceMember.findFirst.mockResolvedValue({
+        userId: "user-123",
+        workspaceId: "workspace-456",
+        role: "MEMBER",
+      });
+      await gateway.handleConnection(mockClient);
+      vi.clearAllMocks();
+    });
+
+    it("should handle multiple start-stop cycles for the same client", async () => {
+      // First cycle
+      gateway.handleStartTranscription(mockClient, {});
+      gateway.handleAudioChunk(mockClient, Buffer.from("cycle-1"));
+      await gateway.handleStopTranscription(mockClient);
+
+      vi.clearAllMocks();
+
+      // Second cycle
+      gateway.handleStartTranscription(mockClient, { language: "de" });
+      gateway.handleAudioChunk(mockClient, Buffer.from("cycle-2"));
+      await gateway.handleStopTranscription(mockClient);
+
+      expect(mockSpeechService.transcribe).toHaveBeenCalledTimes(1);
+      expect(mockClient.emit).toHaveBeenCalledWith(
+        "transcription-final",
+        expect.objectContaining({ text: "Hello world" })
+      );
+    });
+
+    it("should isolate sessions between different clients", async () => {
+      const client2 = createMockSocket({ id: "client-2" });
+      client2.data = { userId: "user-456", workspaceId: "workspace-789" };
+
+      // Client 2 also needs to be "connected"
+      mockAuthService.verifySession.mockResolvedValue({
+        user: { id: "user-456" },
+        session: { id: "session-456" },
+      });
+      mockPrismaService.workspaceMember.findFirst.mockResolvedValue({
+        userId: "user-456",
+        workspaceId: "workspace-789",
+        role: "MEMBER",
+      });
+      await gateway.handleConnection(client2);
+      vi.clearAllMocks();
+
+      // Start sessions for both clients
+      gateway.handleStartTranscription(mockClient, {});
+      gateway.handleStartTranscription(client2, {});
+
+      // Send audio to client 1 only
+      gateway.handleAudioChunk(mockClient, Buffer.from("audio-for-client-1"));
+
+      // Stop client 2 (no audio)
+      await gateway.handleStopTranscription(client2);
+
+      // Client 2 should get an error (no audio received)
+      expect(client2.emit).toHaveBeenCalledWith(
+        "transcription-error",
+        expect.objectContaining({ message: expect.stringContaining("No audio") })
+      );
+
+      vi.clearAllMocks();
+
+      // Stop client 1 (has audio) -- should succeed
+      await gateway.handleStopTranscription(mockClient);
+      expect(mockSpeechService.transcribe).toHaveBeenCalled();
+      expect(mockClient.emit).toHaveBeenCalledWith(
+        "transcription-final",
+        expect.objectContaining({ text: "Hello world" })
+      );
+    });
+  });
+});
diff --git a/apps/api/src/speech/speech.gateway.ts b/apps/api/src/speech/speech.gateway.ts
new file mode 100644
index 0000000..907ec57
--- /dev/null
+++ b/apps/api/src/speech/speech.gateway.ts
@@ -0,0 +1,366 @@
+/**
+ * SpeechGateway
+ *
+ * WebSocket gateway for real-time streaming transcription.
+ * Uses a separate `/speech` namespace from the main WebSocket gateway.
+ *
+ * Protocol:
+ * 1. Client connects with auth token in handshake
+ * 2. Client emits `start-transcription` with optional { language }
+ * 3. Client streams audio via `audio-chunk` events (Buffer/Uint8Array)
+ * 4. Client emits `stop-transcription` to finalize
+ * 5. Server responds with `transcription-final` containing the result
+ *
+ * Session management:
+ * - One active transcription session per client
+ * - Audio chunks accumulated in memory (Buffer array)
+ * - On stop: chunks concatenated and sent to SpeechService.transcribe()
+ * - Sessions cleaned up on disconnect
+ *
+ * Rate limiting:
+ * - Total accumulated audio size is capped by config limits.maxUploadSize
+ *
+ * Issue #397
+ */
+
+import {
+  WebSocketGateway as WSGateway,
+  WebSocketServer,
+  SubscribeMessage,
+  OnGatewayConnection,
+  OnGatewayDisconnect,
+} from "@nestjs/websockets";
+import { Logger, Inject } from "@nestjs/common";
+import { Server, Socket } from "socket.io";
+import { AuthService } from "../auth/auth.service";
+import { PrismaService } from "../prisma/prisma.service";
+import { SpeechService } from "./speech.service";
+import { speechConfig, type SpeechConfig } from "./speech.config";
+
+// ==========================================
+// Types
+// ==========================================
+
+interface AuthenticatedSocket extends Socket {
+  data: {
+    userId?: string;
+    workspaceId?: string;
+  };
+}
+
+interface TranscriptionSession {
+  chunks: Buffer[];
+  totalSize: number;
+  language: string | undefined;
+  startedAt: Date;
+}
+
+interface StartTranscriptionPayload {
+  language?: string;
+}
+
+// ==========================================
+// Gateway
+// ==========================================
+
+@WSGateway({
+  namespace: "/speech",
+  cors: {
+    origin: process.env.WEB_URL ?? "http://localhost:3000",
+    credentials: true,
+  },
+})
+export class SpeechGateway implements OnGatewayConnection, OnGatewayDisconnect {
+  @WebSocketServer()
+  server!: Server;
+
+  private readonly logger = new Logger(SpeechGateway.name);
+  private readonly sessions = new Map<string, TranscriptionSession>();
+  private readonly CONNECTION_TIMEOUT_MS = 5000;
+
+  constructor(
+    private readonly authService: AuthService,
+    private readonly prisma: PrismaService,
+    private readonly speechService: SpeechService,
+    @Inject(speechConfig.KEY)
+    private readonly config: SpeechConfig
+  ) {}
+
+  // ==========================================
+  // Connection lifecycle
+  // ==========================================
+
+  /**
+   * Authenticate client on connection using the same pattern as the main WebSocket gateway.
+   * Extracts token from handshake, verifies session, and checks workspace membership.
+   */
+  async handleConnection(client: Socket): Promise<void> {
+    const authenticatedClient = client as AuthenticatedSocket;
+
+    const timeoutId = setTimeout(() => {
+      if (!authenticatedClient.data.userId) {
+        this.logger.warn(`Client ${authenticatedClient.id} timed out during authentication`);
+        authenticatedClient.disconnect();
+      }
+    }, this.CONNECTION_TIMEOUT_MS);
+
+    try {
+      const token = this.extractTokenFromHandshake(authenticatedClient);
+
+      if (!token) {
+        this.logger.warn(`Client ${authenticatedClient.id} connected without token`);
+        authenticatedClient.disconnect();
+        clearTimeout(timeoutId);
+        return;
+      }
+
+      const sessionData = await this.authService.verifySession(token);
+
+      if (!sessionData) {
+        this.logger.warn(`Client ${authenticatedClient.id} has invalid token`);
+        authenticatedClient.disconnect();
+        clearTimeout(timeoutId);
+        return;
+      }
+
+      const user = sessionData.user as { id: string };
+      const userId = user.id;
+
+      const workspaceMembership = await this.prisma.workspaceMember.findFirst({
+        where: { userId },
+        select: { workspaceId: true, userId: true, role: true },
+      });
+
+      if (!workspaceMembership) {
+        this.logger.warn(`User ${userId} has no workspace access`);
+        authenticatedClient.disconnect();
+        clearTimeout(timeoutId);
+        return;
+      }
+
+      authenticatedClient.data.userId = userId;
+      authenticatedClient.data.workspaceId = workspaceMembership.workspaceId;
+
+      clearTimeout(timeoutId);
+      this.logger.log(
+        `Speech client ${authenticatedClient.id} connected (user: ${userId}, workspace: ${workspaceMembership.workspaceId})`
+      );
+    } catch (error) {
+      clearTimeout(timeoutId);
+      this.logger.error(
+        `Authentication failed for speech client ${authenticatedClient.id}:`,
+        error instanceof Error ? error.message : "Unknown error"
+      );
+      authenticatedClient.disconnect();
+    }
+  }
+
+  /**
+   * Clean up transcription session on client disconnect.
+   */
+  handleDisconnect(client: Socket): void {
+    const authenticatedClient = client as AuthenticatedSocket;
+    const sessionId = authenticatedClient.id;
+
+    if (this.sessions.has(sessionId)) {
+      this.sessions.delete(sessionId);
+      this.logger.log(`Cleaned up transcription session for client ${sessionId}`);
+    }
+
+    this.logger.debug(`Speech client ${sessionId} disconnected`);
+  }
+
+  // ==========================================
+  // Transcription events
+  // ==========================================
+
+  /**
+   * Start a new transcription session for the client.
+   * Replaces any existing session for this client.
+   *
+   * @param client - The connected socket client
+   * @param payload - Optional parameters: { language?: string }
+   */
+  @SubscribeMessage("start-transcription")
+  handleStartTranscription(client: Socket, payload: StartTranscriptionPayload): void {
+    const authenticatedClient = client as AuthenticatedSocket;
+
+    if (!authenticatedClient.data.userId) {
+      authenticatedClient.emit("transcription-error", {
+        message: "Not authenticated. Connect with a valid token.",
+      });
+      return;
+    }
+
+    const sessionId = authenticatedClient.id;
+
+    // Clean up any existing session for this client
+    if (this.sessions.has(sessionId)) {
+      this.sessions.delete(sessionId);
+      this.logger.debug(`Replaced existing session for client ${sessionId}`);
+    }
+
+    const language = payload.language;
+
+    const session: TranscriptionSession = {
+      chunks: [],
+      totalSize: 0,
+      language,
+      startedAt: new Date(),
+    };
+
+    this.sessions.set(sessionId, session);
+
+    authenticatedClient.emit("transcription-started", {
+      sessionId,
+      language,
+    });
+
+    this.logger.debug(
+      `Transcription session started for client ${sessionId} (language: ${language ?? "auto"})`
+    );
+  }
+
+  /**
+   * Receive an audio chunk and accumulate it in the active session.
+   * Enforces maximum buffer size from configuration.
+   *
+   * @param client - The connected socket client
+   * @param data - Audio data as Buffer or Uint8Array
+   */
+  @SubscribeMessage("audio-chunk")
+  handleAudioChunk(client: Socket, data: Buffer | Uint8Array): void {
+    const authenticatedClient = client as AuthenticatedSocket;
+
+    if (!authenticatedClient.data.userId) {
+      authenticatedClient.emit("transcription-error", {
+        message: "Not authenticated. Connect with a valid token.",
+      });
+      return;
+    }
+
+    const sessionId = authenticatedClient.id;
+    const session = this.sessions.get(sessionId);
+
+    if (!session) {
+      authenticatedClient.emit("transcription-error", {
+        message: "No active transcription session. Send start-transcription first.",
+      });
+      return;
+    }
+
+    const chunk = Buffer.isBuffer(data) ? data : Buffer.from(data);
+    const newTotalSize = session.totalSize + chunk.length;
+
+    if (newTotalSize > this.config.limits.maxUploadSize) {
+      authenticatedClient.emit("transcription-error", {
+        message: `Audio buffer size (${String(newTotalSize)} bytes) exceeds maximum allowed size (${String(this.config.limits.maxUploadSize)} bytes).`,
+      });
+      // Clean up the session on overflow
+      this.sessions.delete(sessionId);
+      return;
+    }
+
+    session.chunks.push(chunk);
+    session.totalSize = newTotalSize;
+  }
+
+  /**
+   * Stop the transcription session, concatenate audio chunks, and transcribe.
+   * Emits `transcription-final` on success or `transcription-error` on failure.
+   *
+   * @param client - The connected socket client
+   */
+  @SubscribeMessage("stop-transcription")
+  async handleStopTranscription(client: Socket): Promise<void> {
+    const authenticatedClient = client as AuthenticatedSocket;
+
+    if (!authenticatedClient.data.userId) {
+      authenticatedClient.emit("transcription-error", {
+        message: "Not authenticated. Connect with a valid token.",
+      });
+      return;
+    }
+
+    const sessionId = authenticatedClient.id;
+    const session = this.sessions.get(sessionId);
+
+    if (!session) {
+      authenticatedClient.emit("transcription-error", {
+        message: "No active transcription session. Send start-transcription first.",
+      });
+      return;
+    }
+
+    // Always remove session before processing (prevents double-stop)
+    this.sessions.delete(sessionId);
+
+    if (session.chunks.length === 0) {
+      authenticatedClient.emit("transcription-error", {
+        message: "No audio data received. Send audio-chunk events before stopping.",
+      });
+      return;
+    }
+
+    try {
+      const audioBuffer = Buffer.concat(session.chunks);
+      const options: { language?: string } = {};
+      if (session.language) {
+        options.language = session.language;
+      }
+
+      this.logger.debug(
+        `Transcribing ${String(audioBuffer.length)} bytes for client ${sessionId} (language: ${session.language ?? "auto"})`
+      );
+
+      const result = await this.speechService.transcribe(audioBuffer, options);
+
+      authenticatedClient.emit("transcription-final", {
+        text: result.text,
+        language: result.language,
+        durationSeconds: result.durationSeconds,
+        confidence: result.confidence,
+        segments: result.segments,
+      });
+
+      this.logger.debug(`Transcription complete for client ${sessionId}: "${result.text}"`);
+    } catch (error: unknown) {
+      const message = error instanceof Error ? error.message : String(error);
+      this.logger.error(`Transcription failed for client ${sessionId}: ${message}`);
+      authenticatedClient.emit("transcription-error", {
+        message: `Transcription failed: ${message}`,
+      });
+    }
+  }
+
+  // ==========================================
+  // Private helpers
+  // ==========================================
+
+  /**
+   * Extract authentication token from Socket.IO handshake.
+   * Checks auth.token, query.token, and Authorization header (in that order).
+   */
+  private extractTokenFromHandshake(client: Socket): string | undefined {
+    const authToken = client.handshake.auth.token as unknown;
+    if (typeof authToken === "string" && authToken.length > 0) {
+      return authToken;
+    }
+
+    const queryToken = client.handshake.query.token as unknown;
+    if (typeof queryToken === "string" && queryToken.length > 0) {
+      return queryToken;
+    }
+
+    const authHeader = client.handshake.headers.authorization as unknown;
+    if (typeof authHeader === "string") {
+      const parts = authHeader.split(" ");
+      const [type, token] = parts;
+      if (type === "Bearer" && token) {
+        return token;
+      }
+    }
+
+    return undefined;
+  }
+}
diff --git a/apps/api/src/speech/speech.module.ts b/apps/api/src/speech/speech.module.ts
index d2151ef..42978f9 100644
--- a/apps/api/src/speech/speech.module.ts
+++ b/apps/api/src/speech/speech.module.ts
@@ -11,15 +11,18 @@
  *
  * Imports:
  * - ConfigModule.forFeature(speechConfig) for speech configuration
+ * - AuthModule for WebSocket authentication
+ * - PrismaModule for workspace membership queries
  *
  * Providers:
  * - SpeechService: High-level speech operations with provider selection
+ * - SpeechGateway: WebSocket gateway for streaming transcription (Issue #397)
  * - TTS_PROVIDERS: Map<SpeechTier, ITTSProvider> populated by factory based on config
  *
  * Exports:
  * - SpeechService for use by other modules (e.g., controllers, brain)
  *
- * Issue #389, #390, #391
+ * Issue #389, #390, #391, #397
  */
 
 import { Module, type OnModuleInit, Logger } from "@nestjs/common";
@@ -32,15 +35,19 @@ import {
 } from "./speech.config";
 import { SpeechService } from "./speech.service";
 import { SpeechController } from "./speech.controller";
+import { SpeechGateway } from "./speech.gateway";
 import { STT_PROVIDER, TTS_PROVIDERS } from "./speech.constants";
 import { SpeachesSttProvider } from "./providers/speaches-stt.provider";
 import { createTTSProviders } from "./providers/tts-provider.factory";
+import { AuthModule } from "../auth/auth.module";
+import { PrismaModule } from "../prisma/prisma.module";
 
 @Module({
-  imports: [ConfigModule.forFeature(speechConfig)],
+  imports: [ConfigModule.forFeature(speechConfig), AuthModule, PrismaModule],
   controllers: [SpeechController],
   providers: [
     SpeechService,
+    SpeechGateway,
     // STT provider: conditionally register SpeachesSttProvider when STT is enabled
     ...(isSttEnabled()
       ? [