Skip to main content

Choose a voice

Choose a voice for intelligibility and task fit before choosing it for personality. A support agent must remain clear when reading names and reference numbers; a scheduling agent must make dates and times easy to confirm; a sales agent still needs to sound trustworthy when interrupted.

The voice picker is the authority for the voices currently available to the selected architecture, provider, and language. The catalog can change, so this guide describes provider behavior instead of maintaining a fragile count.

Understand voice compatibility​

Modular agents​

Modular agents can use configured voices from Deepgram, OpenAI, ElevenLabs, Fish Audio, xAI, Cartesia, and Rime. Wirevox infers the text-to-speech provider from the selected voice and routes generated text to that provider.

Some catalog voices are explicitly language-specific. Other provider voices are treated as multilingual and can remain selected when the language changes. A multilingual label does not prove that every language sounds equally natural; test the exact language and content you will deploy.

Realtime agents​

A Realtime voice must belong to the active Realtime provider:

  • OpenAI Realtime uses an OpenAI Realtime voice.
  • Grok Realtime uses an xAI voice.
  • Gemini Live uses a Gemini Live voice.

Switching between Modular and Realtime activates the compatible voice saved for that architecture. Wirevox preserves the inactive architecture's saved selection so you can switch back without rebuilding it.

Preview and compare voices​

Use the play control beside a voice to hear a sample. The Playground > Voices area is useful when you want to compare voices before editing an agent.

Compare candidates with the same script. Include:

  • the opening greeting;
  • business and product names;
  • names, addresses, dates, times, prices, and confirmation numbers;
  • an apology or sensitive response;
  • a short action confirmation; and
  • content in every enabled language.

Do not judge a production voice from a single polished sentence. Browser audio is also not a substitute for a phone test, where telephony compression and the caller's connection affect clarity.

Set speaking speed​

The editor shows the valid range for the selected voice:

  • Cartesia voices: 0.6–1.5×
  • Rime voices: 0.4–2.5×
  • Other currently selectable voices: 0.7–1.2×

Cartesia, Rime, Grok TTS, OpenAI TTS, ElevenLabs, and Fish Audio receive a provider speed control. Realtime models receive pacing guidance. Deepgram native voices use their natural provider pace, so do not assume the slider produces the same effect across every voice.

Start near 1×. Slow down only enough to improve comprehension, then retest interruption timing and total call length.

Create a custom voice​

Wirevox can create a private workspace voice through Fish Audio. In the voice picker, choose the custom voice option and either:

  1. record the three guided samples in the browser, with each complete clip lasting at least 10 seconds; or
  2. upload one to four WAV, MP3, M4A, or OPUS files, up to 20 MB each.

Use one speaker, a quiet room, and clean audio without music or background conversation. Confirm that the voice is yours or that you have the owner's permission. The resulting voice is private to the workspace. Wirevox sends the samples to Fish Audio to create the model and does not retain the uploaded recordings after that request.

Creating a clone and selecting it for an agent are separate steps. Wait for the voice to be ready, preview it, select it in the draft, test it, and publish the agent change.

Selection checklist​

  • The voice appears for the active architecture and provider.
  • It remains clear over a real phone call.
  • It pronounces the business vocabulary correctly.
  • Its language performance was tested, not assumed.
  • Speed and delivery suit the task.
  • The estimated credit rate is acceptable.
  • A custom voice has documented owner permission.