Testing Voice In The Editor

The workflow editor's Playground panel includes a mic button when the open workflow has a voice trigger. Click it to speak directly to your workflow without going through the embedded chat widget or AI Hub — useful for iterating on prompts, models, and turn-taking behavior while building the workflow.

When the mic button appears

The mic icon only renders when the open workflow has at least one trigger carrying a websocketTasks extension — i.e. the embedded sub-workflow that runs during the voice session. Practically:

• Open a workflow with a browser/v1/voiceSession trigger → mic appears
• Open a workflow with no voice trigger → no mic
• Open a brand-new workflow → no mic until you add the trigger and save

Adding the trigger does NOT require deploying the workflow. The editor's test path resolves the draft workflow definition directly via the workflow service, so you can iterate on the trigger configuration and click mic without publishing.

How testing differs from production voice

The editor's test path uses a separate WebSocket endpoint:

production:   wss://host/api/automation/webhooks/{webhookId}/wss
editor test:  wss://host/api/platform/internal/workflow-tests/{workflowId}/wss

Both terminate in the same handler class but key off different identifiers. The editor path:

• Is gated by your ByteChef session cookie (no cross-origin embedding)
• Reads the workflow's draft definition (no publish required)
• Mints session tokens via POST /api/platform/internal/workflow-tests/{workflowId}/voice-session-token
• Runs the same audio bridge / inter-task chain wiring as production

What you test in the editor is what your customers will get from the embedded widget or AI Hub.

Click flow

1. Click the mic icon in the Playground header. The icon flips to a red mic-off when active.
2. The browser prompts for microphone permission. Grant it. (If you've denied it before, you'll need to re-grant via browser site settings — see Troubleshooting below.)
3. A status banner appears: "Connecting microphone…"
4. Once the WS is open and Deepgram has accepted the audio: the banner flips to "🎙 Listening… click mic to end." (Or "🔊 Assistant is speaking…" once it responds.)
5. Speak. Transcripts and assistant responses appear in the chat thread as they arrive.
6. Click the mic again (now showing the mic-off icon) to end the session.

Iteration loop

Voice testing is the fastest way to evaluate prompt + model changes:

1. Edit the system prompt or LLM model in the Deepgram voiceAgent task
2. Save the workflow
3. Click mic, speak, listen
4. Repeat

Each click of the mic starts a fresh session — the previous Deepgram WS closes, a new sub-workflow job spawns, and chat memory resets (unless you've configured a fixed threadId on a streaming agent task). This is the right semantics for prompt iteration; you don't want the previous test's context bleeding into the next.

For multi-turn conversation memory across sessions, set a stable threadId on the agent task (e.g. ${callSid} interpolates per-session; 'editor-test' reuses across all editor sessions).

Cost during testing

Every editor voice session bills you for real Deepgram usage. A 1-minute test = ~$0.13. The 30-minute max-session-duration cap applies in the editor too. For prompt iteration, keep sessions short and exit explicitly via the mic-off button.

Browser requirements

Same as production voice: Chrome 66+, Firefox 76+, or Safari 14.1+. The mic button is disabled with a tooltip explaining why on unsupported browsers — no silent failures.

Troubleshooting

Mic button doesn't appear

The workflow doesn't have a voice trigger. Add a browser/v1/voiceSession trigger with an embedded sub-workflow under websocketTasks. See the Voice Quickstart.

Mic button greyed out

Either the browser is unsupported (hover for the tooltip explaining why) or the workflow hasn't been saved yet (the test path requires a persisted workflow id).

"Workflow has no websocketTasks defined on its trigger"

The trigger exists but the websocketTasks extension is missing or empty. Re-open the trigger config and add the sub-workflow under "Real-Time Workflow".

Audio works but no response

Almost always a sample-rate mismatch. The trigger's sampleRate (16000 by default) must match the Deepgram task's audioInputSampleRate. The output sample rate (audioOutputSampleRate, typically 24000) is independent.

Permission denied / mic button does nothing

The browser remembers permission decisions per origin. Reset via:

• Chrome: chrome://settings/content/microphone
• Firefox: about:preferences#privacy → Permissions → Microphone
• Safari: Safari → Settings → Websites → Microphone

Reload after resetting.

Session closes after 30 minutes

The server enforces a 30-minute max-session-duration to bound Deepgram cost. If you're testing long conversations, close and re-open the mic between iterations.

What's next

• Voice Quickstart — initial workflow setup