Telegram bot that sorts text and voice and describes photos

What it is

This n8n workflow lets a Telegram bot understand three kinds of message. A text, or the transcript of a voice note, is labelled as a task or another message and sent back with that label. A photo is described in words. Thirty nodes do the work, beside ten sticky notes.

Who it suits

People who want to capture quick thoughts by text or voice and have photos explained. Skip it if you expect the workflow to create tasks in another app, since it only labels and replies.

You need a Telegram bot, an OpenAI key and an n8n that Telegram can reach over the internet.

Where to use it

Telegram posts every message to a web address, so your n8n needs a public HTTPS address. After that you use the bot from any Telegram chat with it.

• n8n Cloud: import the file there and switch it on
• Your own n8n server: the same file, with a public HTTPS address
• Other labels: rewrite the two categories in the classifier nodes
• Other photo questions: set a prompt in Analyze Image

Setup

Download saves a zip with workflow.json and Mindket’s setup notes in your downloads folder. The last step tells Telegram where to send messages, and three helper nodes can do it for you.

1. In n8n, create a workflow, open its menu, choose Import from File and pick workflow.json.
2. In Telegram, ask BotFather for a new bot and copy its token.
3. Add a Telegram credential and select it in every Telegram node.
4. Add an OpenAI credential and select it in Audio Classifier Model, Text Classifier Model, Transcribe Recording and Analyze Image.
5. In Validation, replace the example first name, last name and numeric id with your own Telegram details.
6. In Telegram Token & Webhooks, replace <YOUR_TELEGRAM_BOT_TOKEN> with your token and keep the word bot in front of it. Replace <YOUR_N8N_ADDRESS> in the two web addresses with your public n8n address. In the three status nodes, replace <YOUR_TELEGRAM_CHAT_ID> with your chat id.
7. Switch the workflow on and run Set Webhook Production URL from the editor to register the address with Telegram.
8. Send your bot a text, a voice note and a photo.

How to use it

Message the bot. A text comes back as “Task message” or “Other message” with your words in italics. A voice note does the same with its transcript. A photo comes back as a written description. Messages from anyone else get a short error reply.

Example

You send a voice note that says to book the dentist for Tuesday. The workflow downloads it, turns it into text, and the classifier files it under task. The bot replies with “Task message” and your words.

You then send a photo of a cat on a sofa. The bot replies with a sentence that describes a cat on a sofa.

For newcomers and experts

If you are new. n8n draws a workflow as boxes joined by lines, and each box, a node, does one job. A webhook is an address n8n listens on, and Telegram posts each new message to it. A classifier is an AI step that puts a text into one of your groups.

For experts. Message Router has four outputs, one each for voice, text and photo plus a fallback that sends the error reply. The photo goes through Extract from File to Base64 and Convert to Image File so that Analyze Image can read it as base64. Validation must pass the incoming fields on, or the router sees no message.

Good to know

The original left its models unset, used a speech model that is ending, and had two flaws that stopped the flow.

• Models: the two classifier models had no model set, so they ran on the node’s default gpt-5-mini, although they were named gpt-4o-mini. OpenAI’s deprecations page, read on 12 November 2026, lists gpt-5-mini as shutting down on 11 December 2026. The image step named gpt-4o-mini, and OpenAI lists the dated snapshot of that model for shutdown on 14 January 2027 with gpt-6-luna as the replacement. Mindket’s copy uses gpt-6-luna for all three, which OpenAI’s model page lists for text and image input
• Speech to text: n8n’s OpenAI node always sends whisper-1, which OpenAI’s page lists for shutdown on 26 February 2027 with gpt-transcribe as a replacement. Mindket’s copy calls OpenAI’s transcription address with an HTTP Request node and gpt-transcribe. OpenAI’s speech to text guide lists ogg but not oga, the ending Telegram gives voice notes, so Name the voice file renames the file first
• A message path that stopped: Validation did not pass the incoming fields on, so the router never saw the message and every message got the error reply. Mindket’s copy turns on the include other fields option
• A file name that was missing: Convert to Image File reads result.file_path, which Extract from File to Base64 drops unless it keeps the source JSON. Mindket’s copy turns Keep Source to JSON on
• The photo limit: Analyze Image allows 300 tokens by default, and OpenAI’s reasoning guide says a reasoning model counts its thinking in the limit, so the copy raises it to 1500
• Labels only: nothing creates a task anywhere, so add your own step after the task messages
• Sizes: Telegram’s Bot API page says bots can download files up to 20 MB, and OpenAI’s guide allows 25 MB for audio

• A public address: anyone who learns the webhook address can post a made-up update. Telegram can send a secret token in a header when you register the webhook, and the Webhook node can check it with Header Auth
• Telegram’s HTML mode: the task and other messages repeat your text in HTML mode. Telegram’s documentation asks for <, > and & to be escaped, so a text with them can fail to send
• A caption: the photo’s caption is saved by Image Schema but no step uses it
• Your bot token: it sits in Telegram Token & Webhooks as plain text, so it is stored with the workflow and shows in execution data. Clear it after you register the webhook
• Not run against Telegram or OpenAI: Mindket ran the flow with stand-ins for both and saw all three kinds of message reach a reply. The helper requests to Telegram were not run
• Your data: texts, voice notes and photos go to OpenAI, and the chat runs through Telegram
• Credentials: the JSON holds none, so add the Telegram and OpenAI credentials yourself