Build a Gemini Voice Agent for Telegram in n8n

This n8n workflow turns a Telegram bot into an assistant you can type or talk to. Gemini transcribes voice notes, a Gemini agent answers and can send email through Gmail or read a Google Sheet, and Google Cloud's text to speech API speaks the reply back. In my video it ran at $0 on free tiers; two parts need updating for today's Gemini models.

The demo shows both routes. Typed, I asked the bot to email that I would be late, to look up December in a sample gym sheet and to name the highest mountain; it returned 300 members and 35,000 in revenue, sent the email and answered. Spoken, the same request for September came back as a voice note with 250 members and 30,000 in revenue from the sheet.

The video was recorded in July 2025, and the Gemini docs have moved since. This is what changed, checked on 23 October 2026:

• gemini-2.0-flash, the model the video uses for transcription, was shut down on 1 June 2026, according to Google's deprecations page.
• n8n's Google Gemini node now has a Transcribe a Recording operation, which replaces the three HTTP requests the video uses for speech to text.
• Gemini 2.5 Flash, the agent's model in the video, still runs but is limited to accounts that already used it. Google points new projects to Gemini 3.8 Flash or 3.5 Flash Lite.
• The Gemini 1.5 models the video suggests as alternatives are no longer listed.
• Cloud Text to Speech now gives WaveNet voices 4 million free characters a month, not the 1 million mentioned in the video, per its pricing page.

The Telegram trigger only works on an n8n instance with a public HTTPS address. If you do not have one yet, set up n8n on the Google Cloud Free Tier first; the other n8n guides on this blog use the same server.

Every message enters through one trigger, gets sorted by type, and both types reach the same AI agent as plain text. The reply then goes out the way the question came in. In node order:

1. Telegram Trigger, on new messages, with file downloads switched on.
2. A Set node that copies the typed text into a field called response message, to keep later expressions short.
3. A Switch node: messages with a voice file ID go to the voice route, messages with text go straight on.
4. Voice route: Telegram Get File downloads the recording, then Gemini turns it into text.
5. The AI Agent, with a Gemini chat model, simple memory, a Think tool, a Gmail tool and a Google Sheets tool.
6. An IF node that checks whether the original message was a voice note.
7. Text reply: Telegram Send Message. Voice reply: text to speech, read the audio file from disk, Telegram Send Audio.

That split matters for cost. Everything up to the agent runs on the Gemini API's free tier with no billing account, while the spoken reply uses Google Cloud, which in the video needed a billing account even though the usage stayed inside the free allowance.

In Telegram, open BotFather (the account with the blue check mark), send /start, then /newbot. Give the bot a display name and then a username. Telegram's BotFather docs say usernames are 5 to 32 characters long, may use only Latin letters, numbers and the underscore sign, and must end in bot, which is the error the video runs into first. BotFather then sends a token; anyone holding it controls the bot, so keep it private.

In n8n, add a Telegram Trigger, create a Telegram credential with that token and choose the Message event. Then switch on Download Images/Files: n8n's trigger docs describe it as including attached files in the output, which is how the voice note arrives as binary data the later nodes can send to Gemini.

Two things trip people up here, both covered on n8n's Telegram trigger issues page. Telegram refuses a webhook that is not HTTPS, so an n8n on localhost or plain HTTP fails with the error above. And a bot has a single webhook, so while the workflow is published the test button stops receiving messages, and the other way round. Unpublish while you test, or make a second bot for testing.

The Switch node needs two rules. The voice rule checks that the trigger's message.voice.file_id field exists, because only voice notes carry one; the text rule checks that message.text exists. On the voice route, a Telegram node with resource File and operation Get takes that same file ID and downloads the recording as binary data.

For the transcription, use n8n's Google Gemini node with the Audio resource and the Transcribe a Recording operation, which n8n's Gemini node docs describe as turning audio into text. It takes the binary file from the Get File node and uses a Gemini API key, which you create in Google AI Studio and save as a Google Gemini(PaLM) credential, as the credential docs explain.

Telegram voice notes are OGG files encoded with Opus, and Gemini's audio docs list both audio/ogg and audio/opus as supported input types, so no conversion is needed.

If you keep the video's HTTP requests

The video built the same step from three HTTP Request nodes against the Gemini Files API, and that route still works with one change. The upload uses the headers Google's Files API docs show today: X-Goog-Upload-Protocol: resumable, X-Goog-Upload-Command: start and the file's size and type, then a second request with X-Goog-Upload-Offset: 0 and X-Goog-Upload-Command: upload, finalize. Uploaded files are deleted automatically after 48 hours.

The change is in the third request. Its URL named gemini-2.0-flash, which no longer answers, so put a current model in the generateContent URL and keep the body:

POST https://generativelanguage.googleapis.com/v1beta/models/gemini-3.8-flash:generateContent

{
  "contents": [{
    "parts": [
      {"text": "Transcribe the following audio message"},
      {
        "file_data": {
          "mime_type": "audio/ogg",
          "file_uri": "FILE_URI_FROM_THE_UPLOAD"
        }
      }
    ]
  }]
}

The AI Agent node gets its prompt from the previous node instead of the chat trigger. The system prompt in the video is short; in substance it says:

You are a helpful assistant. You can answer any question, or send emails
to the address set in the Gmail tool. You don't have to ask me each time
who the email is for, and you can write the subject and body yourself.
Use the Google Sheets tool for any question about the data in it, like the
total revenue or the number of members in a certain month.

Without the line about writing the subject and body, the agent keeps asking for them before it sends anything, as the video shows. For the chat model, the video picks Gemini 2.5 Flash. Google's deprecations page now says the 2.5 models are served only to accounts that already used them and recommend Gemini 3.8 Flash or 3.5 Flash Lite for new projects. Both are free of charge on the free tier of the Gemini API pricing page.

Free has two conditions. Free tier requests count against per model rate limits that you can see in AI Studio, per the rate limit docs, and the pricing page marks free tier content as used to improve Google's products, which matters if your sheet holds customer data.

Add a Simple Memory so the bot remembers the conversation and a Think tool so the agent can reason before answering. Then the two working tools:

• Gmail tool: resource Message, operation Send, your own address in To. For Subject and Message press the sparkle button that lets the model define the value, and give it a short description such as "make up a subject that fits the request". Under options, turn off Append n8n Attribution.
• Google Sheets tool: resource Sheet Within Document, operation Get Row(s). Pick the document from the list, by URL or by ID (the part of the sheet URL between /d/ and /edit), choose the sheet, and describe the tool as reading rows to answer questions about revenue or members by month.

An IF node checks whether the trigger's voice file ID exists. If not, a Telegram Send Message node replies with the agent's output to the chat ID from the trigger, again with Append n8n Attribution turned off. If it does, the voice route starts.

The video speaks the reply with the community node n8n-nodes-google-speech, installed under Settings, Community Nodes, Install. Its npm page shows one version, 0.3.0, published in March 2025, and n8n's community node docs warn that a community node runs unverified code with full access to the machine and to your workflow data. Read its source before you install it on a server that holds real credentials.

The node signs in with a Google Cloud service account. In the Cloud console, create a project, link a billing account, enable the Cloud Text to Speech API and the Cloud Speech to Text API, then create a service account and download a JSON key, and paste the whole file into the n8n credential. The video grants the account Owner and Editor as well as the speech roles. Google's IAM docs warn that basic roles like those carry thousands of permissions, and the node's README asks only for Cloud Speech Client, so start with that role and add another only if a request is refused.

Set the node to Text to Speech, Synthesize, with the agent's output as the text, and choose OGG Opus as the audio encoding; Google's encoding reference describes it as Opus audio in an Ogg container, the format of Telegram voice notes. Pick a language and a Standard voice, then under options turn on saving to a temporary directory (/tmp) and give the file a name.

Standard and WaveNet voices are free for the first 4 million characters a month and then cost $4 per million, while Neural2 and Studio voices get 1 million free characters at $16 and $160 per million, according to the Cloud Text to Speech pricing page. The video uses a Standard voice for the largest free allowance; WaveNet, which the video describes as better at pronunciation, now lists the same 4 million characters.

A Read/Write Files from Disk node with the Read operation picks up the file from the path the speech node returns and puts it in a binary field called data. The last node is Telegram Send Audio with Binary File on, input field data, the chat ID from the trigger and the original message ID as the reply target.

One detail to know about that last node. Telegram's Bot API docs say sendAudio shows a file in the music player and expects MP3 or M4A, while sendVoice is the method for a playable voice message and accepts OGG with Opus. n8n's Telegram node lists Send Audio but no voice operation in its message operations, so if the reply arrives as a music file, call sendVoice with an HTTP Request node instead.

Since the video, Google has added text to speech models to the Gemini API itself. gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts are free of charge on the free tier of the Gemini API pricing page, so in principle the whole bot could run on one Gemini API key with no Google Cloud billing account.

The catch is the file format. Google's speech generation docs say a normal request returns WAV audio at 24 kHz, and Telegram's voice method takes OGG Opus, MP3 or M4A. n8n's Gemini node also lists no speech generation operation, so you would call the API with an HTTP Request node and add a conversion step before Telegram. For a voice note reply today, the Cloud Text to Speech route with OGG Opus output stays the shorter path.

Google also lists a dedicated gemini-3.5-transcribe model with a free tier, built for speaker labels, word timestamps and custom vocabulary. For short voice notes to a personal bot, the Gemini node's transcription is enough.

Most failures come from one of five places, roughly in the order a message travels:

• "An HTTPS URL must be provided for webhook": your n8n has no public HTTPS address. The Google Cloud hosting guide gives it one.
• The bot answers in testing but not once published, or the reverse: Telegram keeps one webhook per bot, so only one of the two URLs receives messages.
• Voice notes fail while typed messages work: the transcription step still names gemini-2.0-flash. Switch to the Gemini node or a current model.
• Requests stop working after a burst of messages: the free tier hit a rate limit. Check your limits in AI Studio.
• The spoken reply arrives as a music track: send it with sendVoice as described above.

The video ends by announcing a YouTube cross posting automation, which now has its own write up in YouTube cross posting with n8n, and AI automation tools for business operations puts agents like this one next to the other options. If you want a voice or chat agent like this answering your customers instead of you, GenovaFlow builds them as custom AI agents.