Voice chat over a webhook with memory and ElevenLabs

Voice chat over a webhook with memory and ElevenLabs

An n8n workflow with a webhook for spoken questions. You post a recording. OpenAI turns it into text, Gemini answers it
with the earlier conversation in its prompt, and ElevenLabs speaks the answer. The reply is an audio file.

What is inside

• workflow.json: the workflow as n8n exports it, 11 nodes and four sticky notes. It holds no credential.
• README.md: these notes.

What we changed from the original copy

• ElevenLabs Generate Audio held the text {{voice id}} at the end of its address, after an equals sign. n8n reads
  that as an expression and rejects it until you replace it. The address is now plain text that ends in
  <YOUR_VOICE_ID>.
• Google Gemini Chat Model called models/gemini-1.5-flash, a model Google no longer lists. It now calls
  models/gemini-3.6-flash.
• Window Buffer Memory used a long test string as its session key. The key is now voice-chat.
• The sticky note about ElevenLabs is rewritten and shows <YOUR_ELEVENLABS_KEY>.
• The original copy was exported switched on. This one is off until you publish it.
• Two node names lost their spaced hyphen. OpenAI - Speech to Text is now OpenAI Speech to Text, and
  ElevenLabs - Generate Audio is now ElevenLabs Generate Audio.

Before you start

• An n8n instance with an address the caller can reach. n8n Cloud works, and so does your own server on the internet.
• An OpenAI API key, for the speech to text step.
• A Google AI Studio key for the Gemini API.
• An ElevenLabs account, its API key and the ID of a voice from its voice library.
• Something that can post a recording as a form upload, such as a web page, an app or a script.

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. Add an OpenAI credential and select it in OpenAI Speech to Text.
3. Add a Google Gemini (PaLM) API credential with your key and select it in Google Gemini Chat Model.
4. Add a Custom Auth credential for ElevenLabs. Its JSON holds your key in a header:

{
  "headers": {
    "xi-api-key": "<YOUR_ELEVENLABS_KEY>"
  }
}

1. Select that credential in ElevenLabs Generate Audio, under Generic Credential Type and Custom Auth. In the same
   node, replace <YOUR_VOICE_ID> at the end of the address with your voice's ID.
2. Save, open Webhook and press Listen for test event. Post a short recording to its test address. This command
   does it and saves the answer:

curl -X POST -F "voice_message=@question.mp3" "<TEST ADDRESS>" --output answer.mp3

1. Press Publish and use the production address from then on.

How a run goes

1. Webhook takes a POST on the path voice_message. It waits for a node to answer, so the caller gets the audio in
   the same call.
2. OpenAI Speech to Text transcribes the file in the form field voice_message. OpenAI accepts flac, mp3, mp4,
   mpeg, mpga, m4a, ogg, wav or webm, and the node sends the model whisper-1.
3. Get Chat loads the stored messages from Window Buffer Memory. Aggregate gathers them into one list, context.
4. Basic LLM Chain sends the transcript to Google Gemini Chat Model. A message ahead of it tells the model to read
   the earlier conversation first, and lists it as Human: and AI Assistant: lines.
5. Insert Chat adds two messages to the memory, the transcript as the user's and the answer as the AI's.
6. Limit keeps the first item. Insert Chat in this version sends back every stored message as its own item, so
   without Limit the next node would run once for each of them.
7. ElevenLabs Generate Audio posts the answer text to ElevenLabs' text to speech endpoint for your voice.
8. Respond to Webhook sends the first binary file it receives back to the caller.

Change the look

• The voice: <YOUR_VOICE_ID> in ElevenLabs Generate Audio.
• The speaking model: add model_id as a second body parameter there. ElevenLabs uses eleven_multilingual_v2 when none
  is given.
• The assistant's personality: the message in Basic LLM Chain. It holds only the instruction to read the earlier
  conversation, so add the role you want to it.
• The answering model: models/gemini-3.6-flash in Google Gemini Chat Model.
• The conversation: voice-chat in Window Buffer Memory. A different key starts a separate history.

Good to know

• Google's deprecations page, read on 12 November 2026, does not list gemini-1.5-flash, and its models page offers no
  Gemini 1.5 model. The deprecations page names gemini-3.6-flash as the replacement for the 2.0 flash family, and the
  models page marks it stable.
• OpenAI Speech to Text has no model setting and sends whisper-1. OpenAI's deprecations page, read on 12 November
  2026, lists whisper-1 for shutdown on 26 February 2027, with gpt-live-transcribe or gpt-transcribe as the
  replacement. After that date, swap the node for an HTTP Request to OpenAI's transcription endpoint.
• Every caller shares one conversation, because the session key is fixed. To keep callers apart, build the key from
  something the request carries.
• The memory lives in the n8n process. n8n drops a history after an hour without use and a restart clears it all.
• Basic LLM Chain writes undefined wherever a history entry has no human or AI text, so a new session can show it.
• The webhook has no authentication. Anyone with its address can start a call that spends your three accounts, so turn on
  the webhook's own authentication or keep the address private.
• OpenAI bills the recording, Google bills the answer, and ElevenLabs bills the characters it speaks.
• The recording goes to OpenAI, its text and the history go to Google, and the answer goes to ElevenLabs.