Skip to content
MIRA docs

Voice replies & talking to MIRA

MIRA can speak — its replies and check-ins can arrive as voice notes — and it can listen — voice notes you send are transcribed so you can talk instead of type. This guide sets both up.

Voice works per channel: you decide, for each channel, whether MIRA speaks. On a phone channel like Telegram or Signal a spoken reply arrives as a voice note; in the web UI it plays in the browser.

You’ll need a channel connected first — see Connect a channel.

Choose when MIRA speaks (per channel)

Section titled “Choose when MIRA speaks (per channel)”

Each channel has its own voice policy, set under Settings:

  • Off — text only on this channel.
  • On request — MIRA speaks only when you ask it to, or when you sent a voice note (it replies in kind).
  • Always — MIRA sends a voice note alongside its text reply, every time.

For example, set Telegram or Signal to always if you’d like every reply as a voice note on your phone, but leave email as text. The policy is per channel, so mixing styles is fine.

Pick a voice

Section titled “Pick a voice”

MIRA ships with a choice of voices. Pick one under your voice settings — for example af_heart or bf_emma (Kokoro voices). Your voice preference follows you across channels.

Enable Kokoro for natural local speech

Section titled “Enable Kokoro for natural local speech”

MIRA’s voice runs entirely on your own machine by default — no cloud calls needed. There are three local quality tiers:

  • Kokoro — natural, studio-quality speech that runs locally. Enable this if you want MIRA to actually sound good. It’s a larger download (a few hundred MB) but needs no cloud round-trip.
  • Piper — the lightweight local default. Small, fast, and good enough for most uses; works out of the box.
  • eSpeak — a tiny last-resort fallback used only if a model download fails. It sounds robotic.

Turn on Kokoro in your voice settings and choose a Kokoro voice (such as af_heart or bf_emma). MIRA downloads the model the first time it’s needed.

Kokoro speaks English only (American or British) and has no speech-rate control. For other languages, MIRA falls back to eSpeak or a configured cloud / OpenAI-compatible voice backend. On a plain CPU, expect Piper-quality speed; Kokoro shines on capable hardware.

If your operator has configured a cloud or self-hosted voice backend (OpenAI-compatible, ElevenLabs, Cartesia, and similar), those voices appear as options too. The local backends are the privacy-preserving default.

Hear it in the web UI

Section titled “Hear it in the web UI”

In the web chat, spoken replies play in the browser — no phone channel needed. With your web channel’s voice policy set to on request or always, replies are synthesised and played back inline as MIRA writes them.

Talk back: send MIRA voice notes

Section titled “Talk back: send MIRA voice notes”

Voice isn’t one-way. On a channel that carries audio — Telegram, Signal, and custom voice-capable channels — you can send MIRA a voice note and it will transcribe it (using the built-in whisper speech-to-text) and treat it as a normal message. Combined with an always voice policy, that’s a full spoken back-and-forth: you talk, MIRA talks back.

Troubleshooting

Section titled “Troubleshooting”
  • No voice note on my phone. Check that the channel’s voice policy is on request or always, and that the channel is connected and enabled.
  • The voice sounds robotic. You’re likely on eSpeak (the fallback). Enable Kokoro (or Piper) and pick a Kokoro voice.
  • My language isn’t English. Kokoro is English-only; use a configured cloud voice backend, or accept eSpeak for other languages.
Section titled “Related”