Voice stack

The voice stack gives you two things: voice login — sign in with your voice instead of (or in addition to) a password — and the Live Agent, a talking-head assistant inside your research page that you can talk to conversationally. Several specialist services work together behind the scenes to make this possible, and voice usage is metered against your plan.

What each part does

Service	Role in the voice flow
Deepgram (speech-to-text)	Transcribes what you say into text the agent can understand.
ElevenLabs (text-to-speech)	Speaks the agent’s replies aloud in a natural voice.
ASI One (AI brain)	Runs the conversation — understands your questions, uses your idea’s data, and takes actions with your confirmation.
Self-hosted biometrics (ECAPA-TDNN)	Confirms that the voice signing in is your voice — self-hosted, so your voiceprint stays private and is never sent to a third party.
LiveAvatar (talking-head embed)	Gives the agent a face: an embeddable avatar that lip-syncs the agent’s speech on the research page.

This appears in two places: voice login under Settings → Security, and the Live Agent panel inside an idea’s research page.

Prerequisites

• Voice is enabled for your plan, and usage is metered — transcription, synthesis, AI conversation, and avatar time each count against your plan’s voice allowances, shown in your usage view.
• A working microphone for enrollment, voice login, and voice conversations.
• The talking-head avatar is optional: if no avatar is available, the agent still works in voice-only mode.

Connect

There are no external accounts to connect — the voice stack is built into the product. The only setup is enrolling your voice for voice login:

1. Go to Settings → Security → Voice sign-in.
2. Click Enroll and record three short phrases when prompted.
3. You are enrolled. Your voiceprint is stored securely and used only to verify that it is really you.

Note

Enrollment requires your consent, and you can remove your voiceprint at any time from the same page.

Use it

• Sign in with your voice. On the login screen, choose the voice sign-in option and speak your phrase. If multi-factor authentication is enabled, you will complete the usual MFA step afterwards.
• Talk to the Live Agent. Open an idea → research page → Live Agent, allow the microphone, and start talking. The agent listens, thinks, and replies aloud — use it to interrogate your research, test assumptions, or draft next steps.
• Ask it to do things. The agent can work with your idea’s data and will confirm with you before making changes.
• Watch your usage. Transcription, synthesis, AI conversation, and avatar time are metered against your plan, so you can keep an eye on consumption from the usage view.

Disconnect / troubleshoot

• Remove voice login. Go to Settings → Security → Voice sign-in and delete your voiceprint. You can always re-enroll later.
• The agent does not hear you. Check your microphone permissions in the browser and make sure you are not on mute. Background noise can also reduce accuracy.
• The avatar does not appear. The talking head is optional and can be unavailable or paused on your plan — the agent continues in voice-only mode.
• A service is temporarily down. If a voice component is unreachable, the agent degrades gracefully (for example, plain-text replies) rather than failing.
• Voice usage looks high. Voice features are metered; review the usage view to see which part (transcription, synthesis, AI conversation, or avatar time) is driving consumption.

Related

• Voice agent — the full toolkit guide
• AI chat & agents — text-based AI conversations