Generating AI voices in n8n with ElevenLabs: text-to-speech, voice cloning and automated voice messages

A morning digest you listen to while making coffee instead of reading, a meeting reminder left as a WhatsApp voice message, a support reply that arrives as an audio note instead of cold text: AI voice synthesis has gone in a few years from a robotic tone you'd spot in two seconds to a voice nearly indistinguishable from a human recording. Much of that quality leap traces back to Shen et al., presented at ICASSP 2018 ("Natural TTS Synthesis by Conditioning WaveNet on Mel Spectrogram Predictions", see on Google Scholar) — the Tacotron 2 architecture, whose naturalness score (MOS 4.53) already sat right behind recorded human speech (MOS 4.58). Those same foundations now power APIs like ElevenLabs, easy to wire into n8n. Here's how to build a complete text-to-voice pipeline, and where to draw the line.

Three ways to wire ElevenLabs into n8n

The official ElevenLabs node. ElevenLabs publishes a native node covering voice synthesis (text-to-speech), transcription (speech-to-text) and conversational voice agents. On n8n Cloud it's available directly in the verified node list; self-hosted, it installs like any community node, with your ElevenLabs API key (xi-api-key) set in the credentials. It's the fastest path: pick a voice, provide the text, get the audio back as binary data.

HTTP Request, to keep everything in your own hands. ElevenLabs' REST API is also very straightforward to call with HTTP Request: a POST to the text-to-speech/{voice_id} endpoint with the text and voice settings (stability, similarity), the API key in a header, and the audio comes back directly as binary in the response. Useful if you'd rather avoid a dependency on a third-party node, or need a parameter the official node doesn't yet expose — the same reflex as for any API not yet natively covered, with the usual retries and timeouts.

Conversational voice agents. For a more advanced case — an AI phone desk that answers and speaks in real time — ElevenLabs offers a full conversational building block, which combines with the patterns already covered in our AI voice agent for phone calls guide: ElevenLabs synthesis then replaces the audio-generation piece in a Twilio → transcription → LLM → voice chain.

Building a text-to-voice pipeline

The robust pattern breaks into four stages, following the same model we detail for AI image generation in n8n:

1. An LLM writes the script, not a human upstream. A chat node turns raw data (today's email summary, a support ticket, a customer follow-up) into spoken text — short sentences, punctuation meant for the ear, no bullet lists or markdown. A Structured Output Parser guarantees a clean script field in the output.
2. Voice synthesis converts that script into audio via the ElevenLabs node or HTTP Request, using a voice chosen ahead of time and kept consistent across messages — your own, a brand character's, or one of ElevenLabs' public library voices.
3. Storage of the generated audio, as n8n binary data, to Google Drive or an S3 bucket depending on your retention needs.
4. Delivery, to the relevant channel: a voice message on WhatsApp Business or Telegram, an audio attachment in a Slack digest, or live playback in a Chat Trigger for an internal voice assistant.

Concrete use cases

Morning audio digest. FlowKit's Inbox AI Pack (€79) already summarizes your inbox into a daily Slack or Telegram digest; adding an ElevenLabs stage on the output turns that text into a thirty-second voice note you can listen to on the way to the coffee machine — a two-node addition to an existing workflow.

Automated voice follow-ups. Rather than yet another ignored follow-up email, a short, personalized voice message ("Hi Marc, your quote expires Friday…") sent on WhatsApp often catches attention better — as long as you stay sober on frequency and tone, so it doesn't turn into automated sales harassment.

Support replies as audio notes. For long-answer questions (a procedure, a technical explanation), a two-minute voice note often lands better than a wall of text — the LLM drafts the reply just like for a regular ticket, and ElevenLabs voices it before sending.

Multilingual content from a single script. ElevenLabs' multilingual model covers dozens of languages while keeping the chosen voice's timbre: a digest or product announcement can be generated in French and English in the same run, changing only a language parameter.

Voice cloning: what the technology allows, what the law constrains

ElevenLabs lets you clone a voice from a few minutes of recording — your own, a founder's for brand messages, or a customer's who has explicitly consented. It's also the product's most sensitive feature: the same technology that makes an audio digest pleasant to listen to also feeds the voice-impersonation risks documented for years by voice-biometrics research, notably the ASVspoof campaigns (Wang et al., "ASVspoof 2019: A large-scale public database of synthesized, converted and replayed speech", see on Google Scholar), which have evaluated since 2015 how well voice-verification systems detect a synthetic voice.

Three reflexes before automating voice cloning in production:

• Explicit, documented consent for any voice that isn't your own — a checked box or a contract clause, not a verbal go-ahead.
• Traceability: log who authorized which voice, when, and for what use — the same audit trail logic as for any sensitive personal-data processing, a building block covered by FlowKit's Compliance & Audit Pack (€149).
• Transparency toward the recipient where relevant (an AI-generated voice message sent to a customer is better off saying so, even in one sentence), to avoid any confusion about the message's nature.

Costs and guardrails

ElevenLabs bills per character of text converted, with a limited free tier and several paid plans by volume. For a daily digest of a few hundred words, the bill stays around a few cents per run — but as with any usage-billed AI call, an unbounded loop or an overly long source text can push the cost up without an obvious warning. The same guardrails as for any usage-billed AI call apply: cap the length of text sent to synthesis, and track consumption with our AI call cost tracking method.

Verdict

An AI voice pipeline in n8n fits into four nodes: an LLM that writes a script meant for the ear, ElevenLabs (official node or HTTP Request) that voices it, storage as binary data, and delivery to the relevant channel. It's one of the fastest additions to graft onto an existing workflow — a digest, a follow-up, a support reply — and often the one that changes the recipient's perception the most, as long as voice cloning stays on a documented-consent basis rather than an afterthought.