Settings
The Settings tab is where you configure your agent's personality, voice, language, and basic behavior.
For technical settings like turn detection, DTMF keypress collection, and logging levels, see Advanced settings.
Agent Name
A descriptive name for internal use in Studio. Customers do not see this name.
Voice Selection
Stellar offers multiple voices, each with distinct characteristics. All voices support over 50 languages and can be tuned to speak in regional dialects or accents. Some voices are warm and expressive, others are more neutral and professional — choose the one that matches your brand's tone of voice.
Available voices: Alloy, Ash, Ballad, Coral, Echo, Sage, Shimmer, Verse, Marin, Cedar
Experiment with different voices in the Playground before committing to one. Voice characteristics can vary between languages.
Speaking Speed
Controls how fast the agent speaks, from 1.0x (normal) down to 0.75x (slowest). Slowing the agent down helps callers who need more time to follow what is being said.
This setting is post-processing: the generated audio is time-stretched afterwards, rather than the model being told to speak differently. It therefore changes only the delivery — never the words, the voice, or the agent's behavior — and it applies consistently on every turn. The trade-off is quality: because the audio is stretched, slower settings sound progressively more processed, and the effect is clearly audible by 0.75x. Faster-than-normal speeds are deliberately not offered, since they make agents harder to understand over a phone line.
Ask for a calmer, slower pace in the personality prompt before reaching for this setting. Prompting costs nothing in audio quality, and for many agents it is enough. Use the speed setting when prompting alone does not hold up reliably across calls — 0.9x is a good starting point, and worth combining with a prompt instruction rather than pushing the speed lower.
Always listen to your chosen speed in the Playground before publishing.
Speaking speed is available for OpenAI voices. It does not appear for agents running on other AI models, which have no equivalent setting.
Personality / System Prompt
The personality prompt is the most important setting for your agent. It defines how your agent behaves, what it should and shouldn't do, and how it responds to customers. It works together with state-level instructions in the Dialogue Builder — the personality applies globally, while each state can add context-specific behavior.
Example personality prompt
## Identity and role
- You are Emma, a customer service representative for Meridian Bank.
- Help callers with card and transaction questions using information returned by actions or provided in approved knowledge.
- Never guess balances, transaction details, eligibility, or processing times.
- Connect the caller to a representative when their request requires access or authority you do not have.
## Language
- British English is the default response language.
- Switch languages only when the caller explicitly asks.
- Do not infer language from an accent, name, address, filler word, or isolated foreign word.
- Keep preambles, action updates, and answers in the active conversation language.
## Accent
- Speak English with a standard British accent.
- Keep the accent stable throughout the call and use a moderate conversational pace.
- Accent adaptation must not change the response language.
## Tone and conversation style
- Sound warm, calm, and capable.
- Use short sentences and natural pauses.
- Ask one clear question at a time.
- Briefly acknowledge frustration before moving to the next useful step.
- Do not repeat the caller's topic unless confirmation would resolve a real ambiguity.
## Preambles
- Use one short, action-specific preamble before a noticeable lookup or change.
- Keep state transitions, language updates, transcript recovery, and background logging silent.
- After an action succeeds, go directly to the result.
## Unclear audio and silence
- If speech is clipped, noisy, or ambiguous, ask one short clarification question. Do not guess, call an action, or transition.
- If background audio produces a turn, briefly say you could not hear the caller and resume only when they address you.
## Reasoning
- Respond quickly to direct answers and short confirmations.
- Reason before multi-step tasks, tool decisions, troubleshooting, or escalation.
- Do not spend time reasoning over unclear audio; ask for clarification.
Use short labeled sections and keep only global behavior in the personality. Production personalities can be substantial when every rule applies throughout the conversation; move workflows and action-specific procedures into dialogue states. Remove conflicting instructions rather than shortening the prompt for its own sake.
Model-specific prompting
How you phrase your prompt matters as much as what it says, and the most effective techniques depend on the AI model your agent runs on. The Stellar prompting guide explains how to divide instructions between the personality, dialogue states, transitions, and actions, with production-derived examples in English and Dutch. It adapts OpenAI's realtime models prompting guide to Stellar's configuration model.
The Stellar guide targets GPT Realtime 2. Review and retest your prompts when switching the AI model, because different models can interpret the same wording differently.
Transcription Language
Select the primary language for transcriptions. This determines the speech-to-text accuracy for your agent's conversations.
Supported languages
Afrikaans, Arabic, Armenian, Azerbaijani, Belarusian, Bosnian, Bulgarian, Catalan, Chinese, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, Galician, German, Greek, Hebrew, Hindi, Hungarian, Icelandic, Indonesian, Italian, Japanese, Kannada, Kazakh, Korean, Latvian, Lithuanian, Macedonian, Malay, Marathi, Maori, Nepali, Norwegian, Persian, Polish, Portuguese, Romanian, Russian, Serbian, Slovak, Slovenian, Spanish, Swahili, Swedish, Tagalog, Tamil, Thai, Turkish, Ukrainian, Urdu, Vietnamese, and Welsh.
Allowed Languages for Language Switching
Control which languages your agent can switch to during a conversation. If left empty, all supported languages are available. If you specify languages, the agent can only switch to those.
Language switching is handled by the built-in set_language action. The agent calls it when a caller explicitly asks to switch languages — it does not switch automatically based on detected language. When the action runs, it validates the requested language against your allowed list, updates the transcription language for the rest of the conversation, and retranscribes the last user message in the new language to correct any misrecognition from the previous language setting.
If your allowed list contains five or fewer languages, the agent's prompt includes the specific language names. For larger lists, it uses ISO 639-1 language codes instead.
There are two separate layers at play: transcription (speech-to-text accuracy) depends on the configured language and is updated by set_language, while generation (the agent's spoken responses) can use any language regardless of the transcription setting. Until a caller explicitly requests a switch, transcription stays on the default language even if the agent is responding in another language.
Genesys language parameter
If the Genesys language parameter is active for a conversation, it takes precedence over the allowed languages setting. The agent can only use the language provided by Genesys, and set_language requests for other languages are declined.
Transcription Prompt
The transcription prompt helps the speech-to-text model accurately recognize terms specific to your business — product names, industry jargon, customer names, or technical acronyms that might otherwise be misheard.
Example
Common terms:
- TechCo (company name)
- CloudSync (product name)
- API (not "a P I")
- OAuth (not "O off")
Common customer locations:
- Schiphol (airport in Amsterdam)
- Den Haag (The Hague)
Saving changes
Click Save in the editor header after making changes. Changes take effect immediately for new conversations.
Next steps
- Add knowledge to help your agent answer specific questions
- Build dialogue flows to handle different conversation scenarios
- Set up actions to let your agent interact with external systems
- Configure guardrails to ensure safe, compliant conversations
- Set up evaluation rules to automatically assess conversation quality