Chapter 03The E:Voice Manual

The voice on the other end

What you'll get. Which parts of speech run on your phone, which run in the cloud, and how to change either one.

Settings → Voice/Speech, recorded on a phone. Input and Output are set separately.

Speech has two halves, and E:Voice lets you set them separately. Hearing you is speech-to-text. Speaking back is the voice. Either one can run on your phone or in the cloud, and they do not have to match.

Everything lives in one place: Settings → Voice/Speech, described in the menu as “STT, TTS, voice training”. Inside are six sheets — Input (“Choose how E:Voice hears you.”), Output (“Choose how E:Voice speaks back.”), Voice, Speech behavior, Voice provider and Diagnostics.

What already runs on your phone

Two speech models ship inside the app. There is nothing to download and no account needed to use them.

Local speech recognition — the offline transcriber. Pick it in Input, under “Speech recognition”. The Input sheet shows a “Local speech model” row that reads Ready or Missing, with the note “Vosk local recognition uses the bundled offline model.”
Kokoro local voice — the on-device voice, and the app's default for reading replies. Pick it in Output, under “Default readback provider”.

The app is plain about the boundary. Local providers are labelled “speech is transcribed on this device when the bundled model is available” and “Kokoro voice is generated on this device when the bundled model is available.”

The cloud options

Under Input you can choose Local speech recognition, OpenAI cloud speech, E:Voice powered by ElevenLabs, or ElevenLabs (your API key). Under Output: Kokoro local voice, E:Voice powered by ElevenLabs, ElevenLabs (your API key), OpenAI cloud voice, Paired runtime voice, or Muted.

Choosing a “your API key” option without a key stored sends you to the API Key Vault with a nudge: “Add an ElevenLabs key in API Key Vault to enable direct speech recognition.” A key of your own bills to your provider account. Chapter 04 covers the vault.

Premium voices, said straight

Premium voices: details coming soon.

Local voice runs on the phone.
Premium voices: details coming soon.
You can bring your own ElevenLabs key.

Giving an agent its own voice

Voice is set per agent, so Frank and a paired runtime can sound different. Each agent offers three routes:

The premium voice: speaks with a premium voice. Premium voices: details coming soon.
Local voice: always speaks with the local voice, which runs on the phone.
Follow app default — whatever you chose in Output. The screen names it for you rather than making you remember.

Each agent also gets a local fallback voice you can pick and preview, so the drop from premium to local is a change in polish rather than a change of person. There is a Preview button next to Save voice, and preview works on every plan.

If it mishears you

Two things help. Voice training (Settings → Voice/Speech, action “Tune”) teaches local recognition how you say product names and phrases — “Corrections stay on this device and are applied before Frank sees the transcript.” And the noise filter slider in Speech behavior, at 50 by default, is worth raising in a loud truck.