Telegram voice translator that replies in text and speech

Telegram voice translator that replies in text and speech

An n8n workflow that translates voice messages in Telegram. You send your bot a voice message in one of two languages. The
workflow turns the speech into text and finds out which language it was in. It translates the text into the other language
and sends the result back twice, as a text message and as a spoken audio file.

What is inside

• workflow.json: the workflow as n8n exports it, 12 nodes and three sticky notes. It holds no credential.
• README.md: these notes.

What we changed from the original copy

• The speech to text step was n8n's OpenAI node, which always asks for whisper-1. OpenAI's deprecations page, read on
  12 November 2026, gives 26 February 2027 as the removal date of whisper-1 and names gpt-transcribe as a replacement. The
  new Transcribe voice node calls /v1/audio/transcriptions with gpt-transcribe.
• A new node, Name the voice file, renames Telegram's .oga voice file to voice.ogg. OpenAI's speech to text guide lists ogg
  among the formats and does not list oga.
• Chat Model left the model unset. In n8n's source, read on 12 November 2026, an unset model on node version 1 means
  gpt-5-mini, which OpenAI shuts down on 11 December 2026. This copy is on version 1.2 and names gpt-6-luna. The chain
  has no tools, so the note about tool calls on OpenAI's model page does not apply.
• Input Error Handling set a field that nothing read, so a text message crashed the download step. The new node
  Is it a voice message checks for a voice message, and Ask for a voice message answers anything else with a short request.
• Telegram Trigger listened to every kind of update. It now listens to messages only, and it is on node version 1.2, which
  can limit who may use the bot.
• Three nodes have clearer names. Telegram1 is Download voice file, OpenAI Chat Model is Chat Model and OpenAI is
  Speak translation, which now names its operation, Generate Audio.
• The three sticky notes are rewritten, and the outside link in one of them is gone.
• Every credential reference and the webhook id of the author's trigger are removed, so you pick your own.

Before you start

• An n8n instance with a public HTTPS address. Telegram delivers updates by an HTTPS POST to the webhook address.
• A Telegram bot and its token. You make one by chatting with BotFather inside Telegram.
• An OpenAI API key with some credit. It pays for the transcription, the translation and the speech.

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. In Telegram Trigger, Download voice file, Text reply, Audio reply and Ask for a voice message, add a Telegram
   credential with your bot token.
3. In Chat Model and Speak translation, add an OpenAI credential with your API key. In Transcribe voice, which is an HTTP
   Request node, choose the credential type OpenAI and select the same credential.
4. Open Settings and type your two languages as plain words in language_native and language_translate. The file starts
   with english and french.
5. Press Test workflow and send a voice message to your bot. A text reply and an audio reply arrive in the same chat.
6. Turn the workflow on, so that the bot keeps answering when you are away.

How a run goes

1. Telegram Trigger starts the run when someone sends a message to your bot.
2. Settings holds the two languages for the later steps.
3. Is it a voice message looks for a voice note. Anything else goes to Ask for a voice message, which asks for one.
4. Download voice file gets the audio from Telegram, and Name the voice file renames it to voice.ogg.
5. Transcribe voice sends the file to OpenAI and gets the text back.
6. Auto-detect and translate asks Chat Model to find the language of the text. If it is the first language, it is translated
   into the second, and the other way round. The model must answer with the translation only.
7. Text reply sends the translation as a text message.
8. Speak translation turns the same text into speech, and Audio reply sends the audio file.

Change the look

• The two languages: language_native and language_translate in Settings.
• The model of the translation: the model name in Chat Model.
• The voice: the Voice option of Speak translation. The file keeps its default, alloy.
• The wording of the translation request: the text in Auto-detect and translate.
• Who may use the bot: add Restrict to User IDs under Additional Fields in Telegram Trigger.

Good to know

• The spoken reply stops working in January 2027. Speak translation uses tts-1, and OpenAI's deprecations page gives
  6 January 2027 as its shutdown date, with the same date for tts-1-hd. The replacement that the page names,
  gpt-realtime-2.1-mini, is not a supported model for the speech endpoint, according to OpenAI's text to speech guide. The text
  reply goes out first, so it still arrives, but the run then ends with an error at Speak translation.
• OpenAI's text to speech guide asks you to tell people that the voice is AI generated. The audio reply carries no such
  note, so add a line to Text reply if listeners could take it for a human voice.
• Two languages only. The prompt translates between the two languages in Settings, and a third language is left to the model.
• Voice messages must be small. A bot can download a file of up to 20 MB from Telegram, and OpenAI accepts up to 25 MB.
  The speech step takes up to 4096 characters, so a long translation is refused there.
• gpt-transcribe costs 0.0045 US dollars per minute of audio on OpenAI's model page. The translation and the speech are billed
  separately.
• Anyone who finds the bot can use it and spend your OpenAI credit. Restrict to User IDs in Telegram Trigger limits it, and it
  is enforced on node version 1.2 and later.
• Replies go out in Telegram's Markdown mode on Text reply. A lone _ or * in a translation can make Telegram refuse the
  message.
• Your data: the audio goes to OpenAI for the transcription, and the text goes to OpenAI for the translation and the speech.
• Not run: we did not call Telegram or OpenAI to test this file. We ran it in a test runner with stand-in answers for a voice
  message and for a text message.
• Credentials: the JSON holds none.