Front-desk reception, first-line enquiries, first-round job interviews, whistleblowing intake: too much volume for a person, too cold for a text chatbot. Talking AI avatars are appearing in exactly these places. The first thing to settle is the division of labour. The avatar is the part that speaks; it does not decide what to say.
How it works | Your chatbot answers, the avatar speaks
A talking avatar service takes text, generates speech and lip movement, and streams video to the user’s browser. The content of the conversation comes from your chatbot, an FAQ search, or a fixed script.
This split has advantages. Your existing chatbot and FAQ stay as they are and you keep control of answer quality. Swap the avatar and the answers do not change; improve the answers and the avatar needs no reconfiguration. Conversely, putting an avatar in front of a desk does not make the desk smarter: the answers still have to be prepared.
Scripted for reception and interviews, conversational for enquiries
Scripted use cases speak in a fixed order: reading interview questions, walking a whistleblower through intake items, confirming a visitor’s purpose. A script with branches is enough; no answer generation is involved. With speech start and end events you can advance to the next question only after the current one finishes.
Conversational use cases answer free-form questions. A chatbot or retrieval over internal documents (RAG) produces the answer and the avatar speaks it. Answer generation time and time-to-first-speech add up, so perceived speed depends on both.
Most deployments should start scripted: no risk of wrong answers, and results are easy to measure.
Response latency and how to avoid making people wait
Time-to-first-speech decides whether users feel ignored. Set phrases can be pre-synthesised and start fast; new sentences take the synthesis time, and conversational use adds generation time.
- Acknowledge first: speak a short set phrase such as “One moment” while the answer is prepared
- Chunk the answer: send sentence by sentence rather than one long block, so the first sound comes sooner
- Allow interruption: let the user cut in and stop the current utterance
Choosing an avatar and supporting languages
Choose a preset person or build one from a photo of a staff member. For reception, a fictional preset is often better received than a likeness of a real employee. Pick by clothing that fits the setting rather than by face: a white coat for medical, a uniform for reception.
If the service speaks whatever language the text is in, you do not need extra avatars for extra languages: let the visitor pick a language and pass that script. Translation quality is a script problem, so have native speakers review in advance.
Personal data and spoken text (what to check for whistleblowing and interviews)
- Is spoken text stored? Does the provider keep the text, or only character and second counts for billing?
- Is it used for training? Do the terms state that your text and video are never used to train the provider’s models?
- Error monitoring: is the text body excluded from what the provider sends to monitoring tools?
- Transport: is video encrypted end to end, and does it work on corporate networks that block UDP?
“Not stored” and “not used for training” in writing makes intake and interview use cases defensible.
Pricing and keeping costs down
Most services bill by connected time, so cost accrues only while the screen is open. For always-on reception, closing the session when nobody is present and ending it after inactivity changes the bill substantially. A monthly cap prevents surprises.
Summary | The answers are 80% of the work
Deploying a talking avatar is mostly about preparing the answers. Start scripted, secure latency with acknowledgements and chunking, and verify data handling in the terms for sensitive uses.
SHANNON provides avacast, an API that streams a talking avatar to the browser from text you send. Speech, lips and blinks are generated on avacast’s side; your chatbot stays in charge of the answers. Spoken text is never stored or used for training. Pricing is on the avacast website. Contact us about reception, help-desk or interview deployments.