Voice Lab

Voice Lab is the dedicated workspace for testing and debugging voice experiences.

Voice Lab uses the shared native OpenAI browser client, with local settings, captions, task results and saved-call status. Native calling depends on the deployment's admission switch, start/end coordinator and provider-scoped configuration; see Voice.

What It Covers

Voice Lab is much broader than a simple microphone test. Depending on your setup, it can help you:

• edit voice-related configuration
• control retrieval behavior during voice tests
• choose microphone and speaker devices
• run live voice sessions
• inspect captions separately from backend task replies, media and citations
• inspect saved call, recording and settlement state without replaying work

Plain-Language Definitions

When To Use Voice Lab Instead Of Playground

Use Voice Lab when the question is about:

• audio input or output
• latency in voice interactions
• live tool calls during voice sessions
• native voice settings and connection recovery

Use Playground when the question is mainly about text quality or retrieval logic.

Typical Voice Validation Flow

1. confirm the agent's voice defaults in Agent Settings
2. open Voice Lab and choose the correct devices
3. run a short live session with common customer intents
4. inspect transcript quality and tool calls
5. inspect saved recordings, gaps and settlement status separately

Best Practices

• after connection loss, inspect the partial-history warning and use Resume voice deliberately

• distinguish a shown result from something the assistant actually said or the caller heard

• if saved-state reads fail, the last observed state remains visible while reads retry; the call is not automatically restarted

• use End to release media; do not treat end acceptance as proof of final accounting or complete recordings

• test in the same environment your users will use, not only on a perfect office setup

• validate both success paths and interruption-heavy paths

• inspect tool-call latency when voice feels slow

• keep voice prompts and retrieval expectations aligned with text behavior

Troubleshooting

Voice feels slow

• inspect tool calls and their duration
• compare retrieval latency with Playground
• confirm the issue is not device or network related

Transcript quality is poor

• test a different microphone and environment
• check whether the spoken language matches the configured expectation
• compare multiple sessions so you do not overreact to one noisy run

New calls are unavailable

VOICE_NOT_AVAILABLE means new-call admission is paused. It does not disable historical reads or end active calls. Check local settings and deployment readiness; there is no retired-provider fallback or remote sync action.

Related Docs

Agent Settings

Tune AI, retrieval, escalation, and voice defaults for one agent.

Playground

Test text answers, retrieval, prompts, and overrides before production.

Conversations

Review chats, reply as human, and inspect voice sessions.