Voice Messages
Without Opening Telegram

The Friction of Switching to Send a Message

You're in the middle of a debugging session. Your partner messages asking if you'll be late for dinner. You need to reply — quick, human, no big deal.

So you switch to Telegram. The app takes a second to focus. You find the chat. You start typing. You mistype because you're still half-thinking about the bug. You fix it. You send. You switch back to the terminal.

By the time you're done, you've lost the mental model you were holding. The context switch cost you more than the 15 seconds it took.

What if you could just speak?

Three seconds. No app switch. No typing. You're still looking at the same screen.

How It Works

Your voice ──▶ SpeechButton STT ──▶ Local AI Transform ──▶ Telegram Bot API
  (7ms)         (Apple Neural         (Gemma 4, local —       (delivers message
                 Engine, offline)      cleans up speech,        to chat)
                                       formats for chat)

Three components:

1. SpeechButton captures and transcribes your voice locally on Apple Neural Engine
2. Local AI (Gemma 4) reads your prompt file and cleans up raw transcription into a natural, well-punctuated message — entirely on your Mac
3. A Python script in your integrations/ folder sends the text to the Telegram Bot API

The result: a clean Telegram message from spoken words, in under 4 seconds total. Everything is local except the final Telegram API call — no Anthropic API key needed.

Setup

Five steps. Under five minutes.

curl https://api.telegram.org/bot<TOKEN>/getUpdates | jq '.result[0].message.chat.id'

You are a message formatter for Telegram chat.

Clean up raw speech transcription into a natural, well-punctuated
chat message. Fix filler words, false starts, and run-on sentences.
Keep the tone conversational. Preserve the original meaning exactly.

Output ONLY the cleaned message text — no quotes, no explanation.

# ~/.config/speechbutton/config.toml

[global]
model = "parakeet-tdt-0.6b-v3-int8"
language = "en"
auto_punctuation = true

[audio]
vad_enabled = true
vad_silence_threshold = 1.0

# Telegram — send message from voice
[[hotkey]]
key = "RightCommand"
channel = "6"
name = "telegram"
transform = "prompts/telegram_message.md"
exec = "TELEGRAM_BOT_TOKEN=123:ABC TELEGRAM_CHAT_ID=456 integrations/send_telegram.py"

#!/usr/bin/env python3
"""Send a message to Telegram via Bot API."""
import json, os, sys, urllib.request

def main():
    text = sys.stdin.read().strip()
    if not text: sys.exit(0)

    token = os.environ.get("TELEGRAM_BOT_TOKEN")
    chat_id = os.environ.get("TELEGRAM_CHAT_ID")
    if not token or not chat_id:
        print("TELEGRAM_BOT_TOKEN and TELEGRAM_CHAT_ID required", file=sys.stderr)
        sys.exit(1)

    req = urllib.request.Request(
        f"https://api.telegram.org/bot{token}/sendMessage",
        json.dumps({"chat_id": chat_id, "text": text, "parse_mode": "Markdown"}).encode(),
        {"Content-Type": "application/json"})
    result = json.loads(urllib.request.urlopen(req, timeout=10).read())
    if result.get("ok"):
        print(f"Telegram: {text[:60]}...")
    else:
        print(f"Error: {result}", file=sys.stderr)
        sys.exit(1)

if __name__ == "__main__":
    main()

chmod +x ~/.config/speechbutton/integrations/send_telegram.py

Local AI or Claude API?

SpeechButton supports two ways to clean up your speech before sending to Telegram. Option A is recommended — it's free, offline, and private.

The integration script is identical for both options. Only the transform line in config.toml changes. If you choose Option B, set ANTHROPIC_API_KEY in your environment and swap the transform line.

Real Workflows

Privacy

Here's exactly what stays on your Mac and what goes to the cloud:

With the default local AI transform, everything stays on your Mac except the final Telegram Bot API call: voice → local STT (Apple Neural Engine) → local AI transform (Gemma 4, on your Mac) → Telegram API. No audio leaves your machine. No transcription leaves your machine. Only the cleaned message text reaches Telegram's servers — which is where it needs to go anyway. No Anthropic API key required. If you prefer higher-quality message polish at the cost of privacy, you can optionally switch to the Claude API transform (Option B) — in that case, transcribed text is sent to Anthropic's servers for processing.

Prerequisites

• ✓ macOS 15+ (Sequoia) — required for local STT
• ✓ Apple Silicon (M1 or later) — required for Apple Neural Engine
• ✓ Telegram account — to receive messages
• ✓ Bot token from @BotFather — free, takes 60 seconds to create
• ✓ Python 3 — pre-installed on macOS, no packages needed

Get Started