The Macha × ElevenLabs Integration: Give Your AI Agents Ears

The Macha × ElevenLabs integration closes that gap. It gives your agents a single, focused capability: take an audio file and turn it into accurate text they can then summarize, triage, search, and act on — exactly like any typed message. It runs on ElevenLabs Scribe, one of the most accurate speech-to-text models available, and it plugs into the same agents you already run on top of Zendesk, Freshdesk, Gorgias, or Front.

A quick, honest note up front, because it changes how you should think about this: **this integration is speech-to-text, not text-to-speech.** ElevenLabs is famous for generating lifelike voices, but that is not what Macha wires up here. Macha doesn't make your agents talk — it lets them hear. We'll come back to that distinction in the watch-outs, because it matters for what you should and shouldn't expect.

What the integration actually does

Connect ElevenLabs to Macha with an API key and your agents gain one tool: Transcribe Audio.

Transcribe Audio — Upload an audio file and get an accurate text transcription using ElevenLabs speech recognition.

That's the whole surface area. It's deliberately small, and that's the point — one reliable primitive that the rest of your agent logic can build on. The integration sits in Macha's Productivity category, authenticates via API key, and adds no knowledge source of its own (it produces text for your agents, it isn't a place agents go to look things up).

In practice the flow looks like this:

1. An audio file enters the workflow — uploaded into a chat, attached to a ticket, or dropped in by an upstream step.
2. Your agent calls Transcribe Audio.
3. ElevenLabs returns a clean text transcript.
4. The agent does whatever you instructed it to do with that text — summarize it, pull out the order number, decide the ticket's priority, draft a reply, route it to the right team.

Macha already supports dropping audio straight into a conversation. The product's drag-and-drop uploader accepts "images, documents (PDF, DOCX, CSV, XLSX), and audio files," so getting a recording in front of an agent is as simple as dragging it onto the input. And once the agent fires the tool, you'll see a live "Using Transcribe Audio" status while it runs — the same real-time tool-status surface every Macha agent shows when it reaches for a connector.

ElevenLabs sits in that same Connectors directory next to your helpdesk and commerce tools. Connecting it is the standard Macha flow — pick the connector, paste a credential, and its tools become available to your agents. It exposes exactly one tool, which is why the directory shows "1 tool available."

Why this matters: voice becomes a first-class input

The reason "give your agents ears" is worth a whole integration is that voice data is usually trapped. It sits in a call-recording system, or a voicemail inbox, or an attachment nobody opens, precisely because turning it into something actionable used to require a human to sit down and listen.

When transcription is a tool your agent can call mid-workflow, voice stops being a dead end and becomes just another input. A few concrete shapes this takes:

• Voicemail triage. A customer's voicemail gets transcribed, summarized into a ticket, tagged with the likely topic, and prioritized — before an agent ever picks it up.
• Call-note automation. Drop a sales or support call recording in and the agent produces structured notes: who, what they wanted, next steps, objections. No more typing up notes after every call.
• Audio attachments on tickets. A customer films a defect and narrates the problem. The agent reads the narration, not just the pixels, and routes accordingly.
• Meeting recaps for the team. ElevenLabs' own suggested pattern: "When an audio file is shared, transcribe it and summarize the key points for the team," then drop that summary into Slack or a Notion doc via Macha's other connectors.

The leverage isn't the transcription on its own — plenty of tools transcribe. It's that the transcript lands inside an agent that can immediately take action on it using everything else it's connected to: your helpdesk, your order system, your knowledge base.

How accurate is it? The Scribe model under the hood

Transcription is only useful if you can trust it, so the model matters. Macha uses ElevenLabs Scribe, which in independent and vendor benchmarks sits at the top of the speech-to-text field.

ElevenLabs' Scribe v2 reports a word error rate of ≤ 5%, and the realtime variant scores 93.5% accuracy on the FLEURS multilingual benchmark spanning 30 languages — ahead of Google's Gemini Flash 2.5 (90%), OpenAI's GPT-4o Mini (85%), and Deepgram's Nova 3 (80%) on the same test. Scribe is built to hold up in the messy conditions support audio actually lives in: noisy backgrounds, overlapping speakers, and strong accents. (ElevenLabs, Scribe v2)

On languages, it's worth being precise, because the number moved between versions. ElevenLabs' Scribe v1 originally documented 99 languages; the current Scribe v2 and Scribe v2 Realtime pages list 90+ languages, including smart multi-language workflows that handle more than one language in a single file. (ElevenLabs, Speech to Text) We use the conservative 90+ figure throughout this post because that's what the live v2 documentation states today — if you need a specific language confirmed, check ElevenLabs' current list rather than trusting a blog.

For a support team, the practical read is: you're not feeding your agents a rough, error-strewn approximation that poisons every downstream decision. The text the agent reasons over is close to what the human actually said.

Setting it up

The setup is intentionally short:

1. In Macha, open Integrations and connect ElevenLabs.
2. Paste your ElevenLabs API key. That's the only credential — there's no OAuth dance.
3. Add the Transcribe Audio tool to whichever agents should handle audio. In the agent builder, the tool shows up in the same picker as the rest of your connectors.
4. Write the instruction. Tell the agent when to transcribe and what to do with the result — for example, "If an audio file is attached, transcribe it, summarize the customer's issue in two sentences, and set the ticket priority based on urgency."

Because Macha agents are model-agnostic, the reasoning on top of the transcript — the summarizing, the routing decision, the reply — runs on whatever model you've assigned that agent. The transcription step is its own tool call.

Where it fits next to the alternatives

It's worth being clear about what this replaces and what it doesn't.

Versus your helpdesk's native call transcription. Zendesk, for instance, can transcribe and summarize calls — but only calls that happen through Zendesk Talk. (Zendesk help) If your audio comes from anywhere else — a voicemail box, an uploaded file, a recording from another phone system — that native feature can't see it. Macha's ElevenLabs tool transcribes any audio file you can hand the agent, regardless of where it came from.

Versus standalone transcription tools. Sonix, Fireflies, Otter, Rev and friends transcribe accurately, but they hand you a transcript and stop. The text then sits in their product, waiting for a human to do something with it. The Macha integration's whole reason to exist is the next step: the transcript lands inside an agent that immediately acts on it inside your support stack.

That said — if all you need is a searchable archive of meeting transcripts and you never want an agent to act, a dedicated transcription app is the simpler tool. Use the right one for the job.

Watch-outs and when not to use it

A capabilities post that only lists upsides isn't honest. Here's where this integration's edges are.

• It does not give your agents a voice. This is the big one. ElevenLabs is best known for text-to-speech and voice agents, and the working assumption people bring to "Macha × ElevenLabs" is often spoken AI replies. Macha doesn't do that today. The integration is transcription only — speech in, text out. If you came looking for agents that talk back on the phone, this isn't that (yet).
• No real-time live-call streaming. The tool transcribes audio files. It's not a live agent listening on an active phone call and whispering answers in real time. Think recordings and attachments, not a streaming call assistant.
• It's a tool call, so it costs credits. Transcription runs as an AI action and draws from your Macha credit balance like any other tool use. Macha credits are billed per action — 0.5–9 credits depending on the model, with the default GPT-5.4 Mini at 1 credit — so a workflow that transcribes and summarizes and routes is a few actions, not one. High-volume audio pipelines are worth metering. See pricing for the full picture.
• Garbage in, garbage out. Scribe is excellent, but a recording that's near-inaudible to a human will still produce a shaky transcript. For mission-critical actions (refunds, account changes), keep a human in the loop on anything the agent flags as low-confidence.
• PII lives in voice too. Call recordings often contain card numbers, addresses, and names. If your compliance posture requires redaction, handle that in your workflow — don't assume the transcript is automatically scrubbed.

A worked example

Say you run support for an e-commerce brand and customers can leave voicemails for after-hours issues. You build one Macha agent on top of your helpdesk and give it the Transcribe Audio tool plus your usual ticket and order connectors. The instruction:

"When a ticket has an audio attachment, transcribe it. Summarize the customer's issue in two sentences. If they mention an order number, look it up. Set priority to high if they sound upset or mention a failed delivery. Post the summary as an internal note and tag the ticket with the topic."

Now an overnight voicemail arrives as an attachment. By the time a human opens the queue in the morning, the agent has already turned that 45-second recording into a tidy internal note, surfaced the order, set the priority, and tagged it. The human starts from a decision, not a "press play."

That's the pattern across every voice use case: the integration doesn't replace the human's judgment — it removes the part where the human had to listen first.

FAQ

Does Macha's ElevenLabs integration make my agents speak? No. It's speech-to-text only. It transcribes audio into text your agents can act on; it does not generate spoken voice replies.

What audio can it transcribe? Audio files you hand the agent — voicemails, call recordings, audio notes, and audio attachments uploaded into chat or onto a ticket. It runs on ElevenLabs Scribe, which supports 90+ languages.

How accurate is the transcription? It uses ElevenLabs Scribe, which reports ≤5% word error rate and tops major speech-to-text benchmarks, including noisy and multilingual audio.

Does it work in real time on live calls? No — it transcribes recorded audio files, not live streaming calls.

How is it priced? Transcription runs as a Macha tool call and draws from your credit balance (credits are per AI action). There's no separate per-minute fee from Macha; you'll also need your own ElevenLabs API key. Start with the 7-day free trial, no credit card required and see pricing for plan details.

What do I need to connect it? Just an ElevenLabs API key. Add it under Integrations, then attach the Transcribe Audio tool to your agents.

Voice is the input your agents have been deaf to. The ElevenLabs integration is a small, sharp fix: one reliable tool that turns recordings into text, dropped right into the agents that already run your support. It won't make your agents talk — but it will finally let them listen.

Ready to wire it up? Browse the full integrations catalog, read the docs, or start a 7-day free trial, no credit card required. For more patterns like this one, the Macha blog breaks down how teams wire agents and connectors together across real support workflows.