fix: #494 Voice input transcription failing in realtime-demo (#530)

Author: seratch

.changeset/wet-snakes-stare.md

@@ -0,0 +1,5 @@
+---
+'@openai/agents-realtime': patch
+---
+
+fix: #494 Voice input transcription failing in realtime-demo

packages/agents-realtime/src/realtimeSession.ts

@@ -29,7 +29,10 @@ import {
   RealtimeOutputGuardrailSettings,
 } from './guardrail';
 import { RealtimeItem } from './items';
-import { OpenAIRealtimeModels } from './openaiRealtimeBase';
+import {
+  DEFAULT_OPENAI_REALTIME_SESSION_CONFIG,
+  OpenAIRealtimeModels,
+} from './openaiRealtimeBase';
 import { OpenAIRealtimeWebRTC } from './openaiRealtimeWebRtc';
 import { OpenAIRealtimeWebSocket } from './openaiRealtimeWebsocket';
 import { RealtimeAgent } from './realtimeAgent';
@@ -147,6 +150,12 @@ export type RealtimeSessionConnectOptions = {
   url?: string;
 };
 
+function cloneDefaultSessionConfig(): Partial<RealtimeSessionConfig> {
+  return JSON.parse(
+    JSON.stringify(DEFAULT_OPENAI_REALTIME_SESSION_CONFIG),
+  ) as Partial<RealtimeSessionConfig>;
+}
+
 /**
  * A `RealtimeSession` is the cornerstone of building Voice Agents. It's the equivalent of a
  * Runner in text-based agents except that it automatically handles multiple turns by maintaining a
@@ -206,7 +215,8 @@ export class RealtimeSession<
   // modalities, speed, toolChoice, turnDetection, etc.). Without this, updating
   // the agent would drop audio format overrides (e.g. g711_ulaw) and revert to
   // transport defaults causing issues for integrations like Twilio.
-  #lastSessionConfig: Partial<RealtimeSessionConfig> | null = null;
+  #lastSessionConfig: Partial<RealtimeSessionConfig> | null =
+    cloneDefaultSessionConfig();
   #automaticallyTriggerResponseForMcpToolCalls: boolean = true;
 
   constructor(

packages/agents-realtime/test/realtimeSession.test.ts

@@ -11,7 +11,11 @@ import {
 } from '@openai/agents-core';
 import * as utils from '../src/utils';
 import type { TransportToolCallEvent } from '../src/transportLayerEvents';
-import { OpenAIRealtimeBase } from '../src/openaiRealtimeBase';
+import {
+  DEFAULT_OPENAI_REALTIME_SESSION_CONFIG,
+  OpenAIRealtimeBase,
+} from '../src/openaiRealtimeBase';
+import { toNewSessionConfig } from '../src/clientMessages';
 
 function createMessage(id: string, text: string): RealtimeItem {
   return {
@@ -122,6 +126,21 @@ describe('RealtimeSession', () => {
     expect(t.connectCalls[0]?.url).toBe('ws://example');
   });
 
+  it('includes default transcription config when connecting', async () => {
+    const t = new FakeTransport();
+    const agent = new RealtimeAgent({ name: 'A', handoffs: [] });
+    const s = new RealtimeSession(agent, { transport: t });
+    await s.connect({ apiKey: 'test' });
+
+    const normalizedConfig = toNewSessionConfig(
+      t.connectCalls[0]?.initialSessionConfig ?? {},
+    );
+
+    expect(normalizedConfig.audio?.input?.transcription).toEqual(
+      DEFAULT_OPENAI_REALTIME_SESSION_CONFIG.audio?.input?.transcription,
+    );
+  });
+
   it('updateHistory accepts callback', () => {
     const item = createMessage('1', 'hi');
     session.updateHistory([item]);