Voice configuration¶

Voice setup

Voice settings decide whether callers can understand and trust the assistant. Focus first on a clear greeting, one recognition language, sensible silence timing, and a calm professional voice.

Polish the voice only after the basic callback flow works.

Voice configuration is where you make the callback assistant easy to hear and easy to answer. The best first setup is usually the boring one: clear greeting, one recognition language, moderate voice settings, and no aggressive tuning.

Where these settings live¶

You will usually touch several product areas:

General Identity for greetings and timezone
Speech for recognition languages and silence behavior
Voice for the spoken voice and any extra voice tuning
Technical or Agent Options for the voice response mode
Tenant Pronunciation settings for shared pronunciation rules

Start with caller comprehension, not personality¶

For the Nuvoca Callback Assistant, the voice should sound:

clear
calm
professional
easy to understand over a real phone connection

Avoid voices that sound theatrical, highly emotional, or overly branded. A callback assistant does not need to impress the caller. It needs to be understood on the first try.

Greetings¶

Your greeting should do one job: orient the caller quickly.

Good greetings for callback flows:

identify the company
identify the assistant
explain the purpose in one sentence

Example:

Thanks for calling Acme. This is Ava, the Nuvoca callback assistant. I can collect your details and arrange a callback from the team.

Keep separate morning, afternoon, and evening greetings only if your team wants that tone. The wording should stay almost identical across time windows.

If the UI exposes Do not allow interruption for the greeting, enable it for the first pass so the introduction is delivered cleanly.

Speech recognition¶

In Speech, start simple:

choose one primary recognition language
keep silence settings near the defaults
add company-specific names or terms only if recognition is clearly failing

This matters because callback flows depend on capturing names, company names, and phone numbers accurately. If recognition is noisy, the rest of the assistant will feel unreliable even when the prompt is good.

When to adjust silence behavior¶

Change silence settings only if you see a specific problem:

if the agent interrupts the caller too quickly, increase the segmentation silence slightly
if the agent waits too long after the caller finishes, reduce it slightly
if the caller is often slow to begin speaking, increase the initial silence timeout

Do not tune these settings blindly. Make one change, place one call, and compare the result.

Voice selection¶

In Voice, choose one voice that matches the recognition language and sounds trustworthy over the phone.

For first launch:

prefer a neutral or professional style
keep speaking behavior moderate
avoid unusual expressive settings until the callback flow is stable

If several voices seem acceptable, pick the one that sounds clearest on a phone speaker, not the one that sounds most interesting in isolation.

Voice response mode¶

Voice response mode controls whether the voice agent prioritizes voice refinements or first-audio speed.

Mode	What it does	Use when
Quality first	Applies configured voice refinements before playback	Pronunciation, language behavior, and consistent voice quality matter most
Fast first audio	Starts playback as soon as text streams; some refinements may not apply	Response speed matters more than exact voice refinements

Keep Quality first for most production agents, especially when names, product terms, or non-English pronunciation must be reliable.

Use Fast first audio only after the conversation works and you are trying to reduce perceived response delay. Test a real call after enabling it.

Not available for all voices

Fast first audio is not available for Dragon HD Omni voices. Those voices use Quality first.

Pronunciation¶

Use tenant Pronunciation settings when the voice repeatedly says important terms incorrectly. Good candidates are company names, product names, abbreviations, currencies, units, and time expressions.

Start with the simplest rule that solves the problem:

Rule type	What it changes	Example use
Term rule: Alias	Replaces a term with a more speakable form	Brand or product names
Term rule: Language hint	Asks a compatible voice to pronounce a term with another language or accent	Foreign names or words in mixed-language calls
Term rule: Phoneme	Uses a precise pronunciation string	Critical terms where alias is not accurate enough
Formatting pattern	Adds speech hints for currency, units, or time	`24€`, `3kg`, `14 Uhr`

Practical workflow:

Test the voice without a rule.
Add one pronunciation rule.
Place the same test call again.
Keep the rule only if it improves the phone result.

Do not add a large dictionary before testing the baseline. Too many rules can create unexpected side effects.

Language accent override¶

Some HD voices support Language Accent Override in advanced voice settings. It lets you keep the selected voice while asking it to speak with a different supported accent.

Use it when:

the voice is otherwise right but the accent does not fit your callers,
a multilingual workflow needs a more specific locale,
test calls show that names or common phrases sound wrong with the base accent.

Leave it empty when the voice's base accent already fits the audience.

Advanced voice tuning¶

If the UI exposes extra controls such as Telephone EQ, enhanced pronunciation, expressiveness, speech tempo, word focus, speaking style, or guidance values, treat them as second-pass tuning. They are rarely the reason a first callback flow succeeds or fails.

Use them only when:

the agent sounds unnaturally flat
pronunciation is acceptable but delivery feels too stiff
you have already confirmed the callback flow itself works

If behavior and wording are wrong, fix the prompt first. If the words are right but the delivery feels off, then tune the voice.

Common mistake¶

The most common voice-configuration mistake is spending time polishing speech output before speech recognition and prompt logic are stable. If callers are not being understood or the callback number is not being confirmed, voice polish will not fix that.

Read this next¶

Go to Agent Options to tune prompt, model, memory settings, and voice response mode
Go to Common Issues if the agent interrupts too much, sounds inconsistent, or misses caller details
Go to Launch Your First Voice Agent if you want the full callback-assistant setup path again