Voice To Text Config

Voice-to-Text Configuration

Set up local Whisper-based transcription for Telegram voice messages.

Steps

Run each step sequentially. Report status clearly after each one.

1. Check faster-whisper installation

python -c "import faster_whisper; print(f'faster-whisper {faster_whisper.__version__} installed')" 2>&1

• If installed: print the version, move to step 2.
• If ImportError: tell the user and install it:

  pip install faster-whisper
  

  Verify the install succeeded before continuing.

2. Check / download Whisper model

The default model is controlled by the WHISPER_MODEL env var (default: base). Available sizes: tiny (40MB, fastest), base (75MB, good balance), small (250MB), medium (750MB, most accurate for CPU).

Ask the user which model size they want if they haven't specified one. Then check if it's cached:

python -c "
import os, sys
from huggingface_hub import try_to_load_from_cache
model = os.environ.get('WHISPER_MODEL', 'base')
cached = try_to_load_from_cache(f'Systran/faster-whisper-{model}', 'model.bin')
if cached:
    print(f'Model \"{model}\" is cached at: {cached}')
else:
    print(f'Model \"{model}\" is NOT cached yet — needs download')
    sys.exit(1)
" 2>&1

If not cached, download it:

python -c "
import os, sys
model = os.environ.get('WHISPER_MODEL', 'base')
print(f'Downloading whisper model \"{model}\"... (this may take a minute)')
from faster_whisper import WhisperModel
m = WhisperModel(model, device='cpu', compute_type='int8')
print(f'Model \"{model}\" downloaded and ready.')
" 2>&1

Important: This download can take 1-3 minutes on first run. Let the user know progress is happening.

3. End-to-end test

If there's a voice file in the Telegram inbox, test transcription against it:

python -c "
import glob, os, sys
from faster_whisper import WhisperModel
inbox = os.path.expanduser(r'~\.claude\channels\telegram\inbox')
files = sorted(glob.glob(os.path.join(inbox, '*.oga')), key=os.path.getmtime, reverse=True)
if not files:
    print('No .oga voice files found in inbox to test against.')
    sys.exit(0)
f = files[0]
print(f'Testing transcription on: {os.path.basename(f)}')
model_size = os.environ.get('WHISPER_MODEL', 'base')
model = WhisperModel(model_size, device='cpu', compute_type='int8')
segments, info = model.transcribe(f, beam_size=5)
text = ' '.join(seg.text.strip() for seg in segments)
lang = getattr(info, 'language', 'unknown')
prob = getattr(info, 'language_probability', 0)
print(f'Language: {lang} ({prob:.0%})')
print(f'Transcription: {text}')
" 2>&1

4. Verify hook is registered

Check that the transcription hook exists in the plugin:

cat "${CLAUDE_PLUGIN_ROOT}/.claude-plugin/hooks/hooks.json" 2>/dev/null || cat "${CLAUDE_PLUGIN_ROOT}/hooks/hooks.json" 2>/dev/null

Confirm it targets mcp__plugin_telegram_telegram__download_attachment.

5. Report

Summarize:

• faster-whisper version
• Model size and cache status
• Test transcription result (if available)
• Hook registration status

If everything passed, tell the user: "Voice transcription is ready. Send a voice message on Telegram to test it live."

If the user wants a different model size, they can set WHISPER_MODEL in their environment (e.g., in .claude/settings.json env section or system env vars).