Where is the audio transcribed?

On a shared speech server, not inside your bot. Your instance installs only a lightweight connector and the audio is uploaded for processing, which avoids hundreds of megabytes of model storage and the transcription CPU and memory cost in every container.

Does it work on both OpenClaw and Hermes?

Yes, from the same Tools page on either framework. On Hermes, restart the bot after installing if it does not notice the tool.

It runs a lightweight base model with CPU int8, which is quick and inexpensive rather than best-in-class. Clear speech transcribes well; heavy accents, crosstalk and background noise produce errors. Ask for meaning rather than a verbatim record.

← Home

Guide

faster-whisper

A large share of real messaging is spoken, not typed — especially on WhatsApp and WeChat. Without transcription your bot simply cannot hear those users. This adds it, in one click, on OpenClaw and Hermes alike.

Why It Matters More Than It Sounds

If your bot serves customers on a phone-first platform, voice notes are not an edge case. A bot that silently drops them looks broken to the person who sent one, and you never see the message they thought they sent.

With transcription installed, a voice note becomes text the agent handles like any other message — it answers, and the person never thinks about it again.

Install It

Open the dashboard and go to Tools.
Pick a running bot — tools install per instance.
Install faster-whisper.

On Hermes, restart the bot afterwards if it does not seem to notice the tool.

How It Runs, and Where the Audio Goes

This one is not a purely local install, and the reason is practical. A speech model is hundreds of megabytes and transcription is CPU-hungry — putting that inside every bot container would cost you disk and memory you would rather spend on the agent.

So your instance installs a lightweight connector, and the model runs on a shared speech server. Audio is uploaded there to be transcribed and the text comes back. The trade is that your bot stays small and fast; the thing to know is that the audio leaves the container.

For most support and assistant use that is unremarkable — it is the same audio the messaging platform already carried. If you handle recordings that cannot leave your own boundary, that is a reason to run your own speech stack instead, and a self-hosted bot can do exactly that: faster-whisper is MIT-licensed and runs anywhere.

What You Get

Voice messages on your connected platforms transcribed automatically, then handled like text.
Audio files you send the bot directly transcribed on request — “what does this recording say?”
A CPU-efficient setup using int8 and a lightweight base model, which is what keeps it quick and cheap rather than best-in-class on hard audio.

Getting Good Results

Expect a base-model transcript: excellent on clear speech, imperfect on heavy accents, crosstalk or background noise.
Ask for meaning rather than a verbatim record. “What is this person asking for?” survives a few wrong words; a legal transcript does not.
For a long recording, ask for a summary and the action items rather than the full text — that is usually what you wanted anyway.

Troubleshooting

Voice notes still ignored. Confirm the tool is installed on that bot, and on Hermes restart it once.
The transcript is wrong in places. That is the small model doing its job cheaply. Ask the sender for text when accuracy is critical.
Nothing happens on one platform only. Platforms differ in how they deliver audio. Test the same clip in web chat to separate a transcription problem from a delivery one.

Try It

Install it, then send your bot a voice note from your phone asking it something ordinary. The reply should arrive as if you had typed it.