Free and open-source local AI dictation tool

GPU‑accelerated or CPU‑only microphone transcription using NVIDIA NeMo Parakeet TDT 0.6B v3, with a clean CLI and a debug mode for device/timing diagnostics.

Features

• GPU/CPU switch via --cpu, with explicit device placement and verification using PyTorch device checks.
• --debug prints device, timings, and CUDA memory stats and truncates transcriber.debug.log on each run to keep logs fresh.
• --list-devices enumerates input devices and --input-device selects one for recording.
• Copies transcript to clipboard (disable with --no-clipboard).

Requirements

• Python 3.10+
• PortAudio for microphone access (PyAudio wheels bind to it)
• For GPU: a CUDA-enabled PyTorch build and compatible NVIDIA drivers
• NVIDIA NeMo Toolkit (ASR) and its runtime dependencies

OS Support

• Linux (Ubuntu/Debian) and WSL2 Ubuntu: supported; this is the primary/validated target.
• Windows (native): not supported currently. The CLI uses select on stdin and NeMo does not provide Windows wheels; use WSL2 Ubuntu instead.
• macOS: not supported currently. NeMo does not publish macOS wheels; consider running inside a Linux VM/container, or use an alternative engine (e.g., Faster‑Whisper) if native macOS support is required.

Linux / WSL2 prerequisites

On Debian/Ubuntu (including WSL2), install system audio dependencies first:

sudo apt update
sudo apt install -y portaudio19-dev pulseaudio libasound2-plugins

Clipboard helpers (optional, for pyperclip):

sudo apt install -y xclip   # X11
# or
sudo apt install -y wl-clipboard  # Wayland

Notes for WSL2/WSLg: audio is bridged via the PulseAudio server at unix:/mnt/wslg/PulseServer, so starting a separate PulseAudio daemon is not needed and is typically refused. RDP sink/source devices may show as suspended until audio flows; that's expected under the WSLg bridge.

Install (uv or pip)

Create and activate a virtual environment:

uv venv
source .venv/bin/activate

Install PyTorch first (choose one):

# GPU (CUDA 12.1)
uv pip install torch torchaudio torchvision --index-url https://download.pytorch.org/whl/cu121

# CPU-only
uv pip install torch torchaudio torchvision --index-url https://download.pytorch.org/whl/cpu

Then install the rest from PyPI:

uv pip install 'nemo_toolkit[asr]' pyaudio 'numpy<2.0' pyperclip

Why two steps? The PyTorch wheels are hosted on a separate index; installing them first avoids accidentally directing all packages to the PyTorch index (which would fail for NeMo and others).

Optional ALSA → Pulse config (Linux)

Direct ALSA to PulseAudio by placing this in ~/.asoundrc if needed:

pcm.!default {
    type pulse
}
ctl.!default {
    type pulse
}

Usage

Make the script executable and run:

chmod +x transcriber.py
python transcriber.py

Helpful options:

# Verbose diagnostics to file (device, timings, GPU memory), spinner preserved
python transcriber.py --debug

# Force CPU even if CUDA is available
python transcriber.py --cpu

# List and select audio input devices
python transcriber.py --list-devices
python transcriber.py --input-device 2

# Disable clipboard copy (avoid needing xclip/wl-clipboard)
python transcriber.py --no-clipboard

Help-first CLI prints instructions and exits without loading the model:

python transcriber.py -h

Verify GPU usage

In --debug, the script prints Model device: cuda:N when on GPU and shows CUDA memory growth during inference. If --cpu is used or only CPU wheels are installed, the device prints as cpu and CUDA memory remains zero.

Model cache

Parakeet is downloaded once and cached under the Hugging Face Hub cache (default ~/.cache/huggingface/hub), and subsequent runs load it locally. This location can be customized via HF_HOME / HF_HUB_CACHE if desired.

Notes

• The script truncates transcriber.debug.log on each --debug run to keep logs fresh.
• The spinner writes directly to the original terminal stream to avoid interference from redirected stdout/stderr during model/model‑load.
• Argparse parsing occurs before heavy imports so -h/--help returns immediately.

Troubleshooting

• PyAudio error about sample rate: some devices only support 48kHz. Try selecting a device via --input-device (see --list-devices) or switch your default input device/sample rate in system settings.
• Clipboard copy fails on Linux: install xclip (X11) or wl-clipboard (Wayland), or run with --no-clipboard.
• GPU memory stays at 0 in --debug: you are likely running the CPU PyTorch build or --cpu is set. Reinstall a CUDA build of PyTorch and omit --cpu.

Windows/macOS status

• Windows native: not supported (NeMo wheels and the select-based key handling are Linux‑only). Use WSL2 Ubuntu and follow the Linux instructions.
• macOS: not supported (no official NeMo/macOS wheels). Use a Linux VM/container or an alternative engine.

License

MIT for this glue code; refer to upstream projects (NeMo, PyAudio, PyTorch) for their licenses.