Skip to main content

Configure voice delivery

Voice delivery controls how an agent speaks after the model has decided what to say. Start with the default Warm professional, Natural, and Subtle settings. Tune delivery only after the system prompt, language, and action flow behave correctly.

Support varies by architecture and provider. Wirevox can apply a control natively, translate it into provider settings, manage it in the Modular pipeline, or add it as model guidance. The editor's capability label explains which path is active.

Core controls​

Personality​

Choose Professional, Warm professional, Calm and reassuring, Friendly and conversational, or Energetic sales. This setting guides delivery; it does not replace the agent's role and behavioral rules in the system prompt.

Expressiveness​

  • Restrained minimizes emotional variation and suits sensitive or regulated conversations.
  • Natural is the general-purpose default.
  • Expressive permits stronger variation and should be tested for consistency.

Adaptive delivery​

When enabled, delivery can adjust to the conversation. Realtime models can use the caller's words and vocal delivery. Modular models receive transcript text and conversation context but do not reliably hear vocal emotion.

Do not instruct a Modular agent to infer stress, age, health, or intent from tone of voice. It does not receive the same audio awareness as a Realtime model.

Reactions​

Reactions can be Off, Subtle, or Natural. Provider support differs, so Wirevox may render a reaction natively or approximate it with speech and punctuation. Keep reactions restrained for complaints, emergencies, financial conversations, and other sensitive contexts.

Advanced controls​

Custom delivery instructions​

Add short instructions for spoken delivery, up to 1,000 characters. Keep them about sound, not business policy.

Sound calm and confident. Slow slightly when repeating phone numbers, dates,
addresses, and appointment times. Do not use theatrical voices.

Business rules belong in the system prompt. Conflicting instructions make provider behavior less predictable.

Pitch​

With adaptive delivery, choose a Subtle, Natural, or Wide pitch range. Without it, choose a Lower, Natural, or Higher baseline. Pitch can be native, hybrid, or guided depending on the provider; it is not an exact musical value.

Pronunciation dictionary​

Use one replacement per line with = or =>:

Wirevox = Wire-vox
HVAC => H V A C
Etobicoke = eh-TOE-bih-coh

Wirevox accepts up to 50 parsed entries. Each written and spoken side can contain up to 100 characters, and the editor limits the raw field to 3,000 characters. In Modular calls, Wirevox applies replacements before text-to-speech without changing the saved transcript. Realtime support is provider-dependent and may be native or guided.

Use replacements narrowly. Test them in full sentences because a phonetic spelling that works alone can sound wrong next to punctuation or numbers.

Realtime turn taking​

Responsive, Balanced, and Patient turn-taking profiles are available for Realtime agents and map to provider-owned voice activity behavior. Modular agents use their speech recognition and patience configuration instead.

Grok Realtime reasoning​

Grok Realtime exposes Fastest response and Higher accuracy response reasoning. Higher accuracy can add latency. Use it only when the improvement is visible in representative task tests.

Cartesia Sonic controls​

Cartesia Sonic voices support native emotion selection and a volume control from 0.5× to 2×. Start with automatic emotion and 1× volume. Avoid compensating for poor source audio with excessive volume.

Provider differences​

  • xAI TTS supports context and native speech cues; some pitch behavior is hybrid.
  • OpenAI TTS uses provider speed plus prompted emotional delivery.
  • ElevenLabs maps delivery into provider stability and style behavior.
  • Cartesia supports native emotion, speed, volume, pauses, and laughter.
  • Rime uses semantic phrasing and punctuation, with native pause handling in supported voices.
  • Deepgram native voices have more limited explicit emotional and speed controls.

These differences mean that identical editor settings can sound different after a voice-provider change. Re-run the delivery test set whenever you change the voice.