MINARAVoice
Talk to Minara and hear it answer — in the web terminal and across messaging platforms, with replies spoken sentence-by-sentence as they stream.
Minara supports voice on every chat surface. In the web terminal you can dictate a message with the microphone button and have replies read aloud as they stream in (the first sentence is audible in under a second). On messaging platforms, send a voice note and the agent transcribes it, answers in text, and (when enabled) replies with a voice note of its own.
Setup
Voice uses a pluggable speech service. Set one of:
ELEVENLABS_API_KEY— preferred when present (lowest latency synthesis, Scribe transcription).OPENAI_API_KEY— fallback, and the only provider able to mint OGG/Opus voice notes (Telegram, Lark, WhatsApp render those as native voice bubbles).
Turn on voice replies for messaging platforms with
MESSAGING_VOICE_REPLY=1 and inbound transcription with
MESSAGING_INBOUND_TRANSCRIBE=1. Voice turns persist into chat history:
the original recording on your message, the exact spoken audio on the
reply.
Platform support
| Platform | Receive voice | Send voice replies | Notes |
|---|---|---|---|
| Web terminal | Yes | Yes | Mic button + streamed read-aloud on every chat surface |
| Telegram | Yes | Yes | Native voice notes (OGG/Opus) |
| Discord | Yes | Yes | Audio attachment |
| Slack | Yes | Yes | Audio file upload |
| Yes | Yes | Cloud API audio type; Opus renders as a voice note | |
| Lark / Feishu | Yes | Yes | Opus uploads render as voice bubbles |
| Matrix | Yes | Yes | MSC3245 voice messages |
| Mattermost | Yes | Yes (audio file) | Inline player client-side |
| BlueBubbles (iMessage) | Yes | Yes (audio file) | |
| Google Chat | No | Yes (audio file) | Inbound download needs service-account media scope (planned) |
| WeChat OA | Yes | Yes | Uses WeChat's built-in transcript when enabled; AMR fallback needs ffmpeg |
| WeCom | Yes | Yes | AMR both ways — requires ffmpeg (VOICE_FFMPEG_PATH) |
| LINE | Yes | Text only | Sending audio needs a public media URL (planned) |
| Signal | No | Yes (audio file) | Inbound voice via signal-cli attachment store (planned) |
| Email / Gmail | No | Yes (mp3 attachment) | No inbound message channel |
| Microsoft Teams | No | Text only | File sends require the consent-card flow (planned) |
| DingTalk | No | Text only | Robot webhook API has no media upload (planned) |
| No | Text only | Voice uses the Silk codec, which standard ffmpeg can't encode |
Platforms marked "planned" still answer voice notes with text replies wherever inbound voice reaches the agent.