Voice Input Claude Russian Offline

Local Voice Input for Claude in Russian: Built in an Evening

A Flask server on localhost:5555 receives audio from a Chrome extension, transcribes it locally with Whisper (small model), and sends the text to the Claude.ai input field. A Menu Bar app built with rumps handles starting and stopping, with the menu icon switching from 🤖 to 🎤. Everything runs fully offline, no limits, completely free. Works on macOS, easily adaptable to Linux.

Prerequisites

Install dependencies:

• Whisper: brew install openai-whisper
• ffmpeg: brew install ffmpeg
• Python 3 and Chrome

Test Whisper:

whisper --version

Look for Russian in the output. Set up the project structure:

mkdir -p ~/projects/whisper-claude
cd ~/projects/whisper-claude
python3 -m venv venv
source venv/bin/activate
pip install flask flask-cors rumps
mkdir -p extension/icons

Virtual environment is essential on macOS Ventura+ to avoid conflicts with system Python.

Transcription Server (server.py)

The server handles POST /transcribe with base64 or file audio, calls Whisper with --language ru --fp16 False (critical for Apple Silicon), and returns JSON with the text. Includes /ping and /shutdown endpoints.

from flask import Flask, request, jsonify
from flask_cors import CORS
import subprocess
import tempfile
import os
import threading

app = Flask(__name__)
CORS(app)

WHISPER_PATH = "/opt/homebrew/bin/whisper"
WHISPER_MODEL = "small"
PORT = 5555

@app.route("/ping", methods=["GET"])
def ping():
    return jsonify({"status": "ok"})

@app.route("/transcribe", methods=["POST"])
def transcribe():
    if "audio" not in request.files:
        return jsonify({"error": "No audio file found"}), 400

    audio_file = request.files["audio"]

    with tempfile.NamedTemporaryFile(suffix=".webm", delete=False) as tmp:
        tmp_path = tmp.name
        audio_file.save(tmp_path)

    try:
        result = subprocess.run(
            [
                WHISPER_PATH, tmp_path,
                "--model", WHISPER_MODEL,
                "--language", "ru",
                "--output_format", "txt",
                "--output_dir", tempfile.gettempdir(),
                "--fp16", "False",
            ],
            capture_output=True, text=True, timeout=60
        )

        txt_path = tmp_path.replace(".webm", ".txt")
        if os.path.exists(txt_path):
            with open(txt_path, "r", encoding="utf-8") as f:
                text = f.read().strip()
            os.unlink(txt_path)
        else:
            text = result.stdout.strip()

        return jsonify({"text": text})

    except subprocess.TimeoutExpired:
        return jsonify({"error": "Timeout"}), 500
    except Exception as e:
        return jsonify({"error": str(e)}), 500
    finally:
        os.unlink(tmp_path)

@app.route("/shutdown", methods=["POST"])
def shutdown():
    threading.Timer(0.5, lambda: os._exit(0)).start()
    return jsonify({"status": "shutting down"})

if __name__ == "__main__":
    print(f"[Whisper] Starting on port {PORT}, model: {WHISPER_MODEL}")
    app.run(host="127.0.0.1", port=PORT, debug=False)

Key tip: --fp16 False prevents crashes on M1/M2/M3 chips.

Menu Bar Controller (menubar.py)

The app checks ports every 5 seconds with rumps.Timer, starts/stops servers in threads. TOOLS config lets you add services with one line. Status shows as emojis: ✅ running, ⏸ paused.

import rumps
import subprocess
import threading
import os
import time
import urllib.request

BASE_DIR = "~/projects/whisper-claude"
VENV_PYTHON = os.path.join(BASE_DIR, "venv/bin/python3")

TOOLS = [
    {
        "name": "Whisper → Claude",
        "script": os.path.join(BASE_DIR, "server.py"),
        "port": 5555,
        "description": "Russian voice input",
    },
]

# ... (rest of code as in original, abbreviated for example)

class AILauncher(rumps.App):
    # implementation as above
    pass

if __name__ == "__main__":
    AILauncher().run()

Full code includes ToolController with toggle, check_all, and graceful shutdown.

Chrome Extension

manifest.json (MV3):

{
  "manifest_version": 3,
  "name": "Whisper for Claude",
  "version": "1.0",
  "permissions": ["activeTab", "scripting"],
  "host_permissions": [
    "https://claude.ai/*",
    "http://127.0.0.1:5555/*"
  ],
  "content_scripts": [
    {
      "matches": ["https://claude.ai/*"],
      "js": ["content.js"],
      "run_at": "document_end"
    }
  ],
  "background": {
    "service_worker": "background.js"
  }
}

background.js handles mixed content: base64 audio from content.js → fetch in background → server.

const SERVER_URL = "http://127.0.0.1:5555";

chrome.runtime.onMessage.addListener((message, sender, sendResponse) => {
  if (message.action === "transcribe") {
    handleTranscribe(message.audio).then(sendResponse);
    return true;
  }
});

async function handleTranscribe(base64Audio) {
  // base64 to Blob to FormData to POST /transcribe
}

Content.js adds a 🎤 button next to the input field and captures MediaRecorder.

Key Highlights

• Local Whisper small model: ~1 GB, transcribes Russian offline in 10-30 seconds.
• Extensible architecture: add new AI tools to TOOLS without code changes.
• Security: localhost-only, no cloud calls, CORS for Chrome.
• Compatibility: macOS Apple Silicon (fp16=False), Linux with path tweaks.
• Performance: 60-second timeout, automatic temp file cleanup.

— Editorial Team