yaptype

Push-to-talk voice dictation powered by NVIDIA NeMo Parakeet-unified-en-0.6b.

Hold a key → speak → release → transcription is pasted into the active window.

Project structure

yaptype/
├── .python-version       # Python 3.10
├── pyproject.toml        # Dependencies & uv config (torch CUDA, NeMo)
├── uv.lock               # Lockfile
├── yapctl.py             # Server manager CLI (start/stop/status/logs)
├── asr_server.py         # ASR server (loads model once, stays hot in VRAM)
├── dictate_client.py     # Push-to-talk client (connects to server, starts instantly)
├── dictate.py            # Standalone push-to-talk (no server needed, slower startup)
├── transcribe.py         # Batch file transcription script
└── README.md

Setup

# 1. Install dependencies (torch CUDA 12.8 + NeMo)
uv sync

# 2. Upgrade NeMo to main branch (required for parakeet-unified-en-0.6b)
#    NeMo 2.7.3+ needed but not yet on PyPI as of 2026-04-18.
uv pip install "nemo_toolkit[asr] @ git+https://github.com/NVIDIA-NeMo/NeMo.git@main" --no-deps

Note: Re-run step 2 after any uv sync. Once NeMo 2.7.3 is on PyPI, step 2 won't be needed.

Usage

Quick start

# Start everything (server + dictation) in one command
uv run --no-sync python yapctl.py start

# Hold Right Ctrl to record, release to transcribe & paste
# Press Ctrl+C to stop dictating (server stays running in background)

# Stop the background server when done
uv run --no-sync python yapctl.py stop

Server management

uv run --no-sync python yapctl.py status       # Check if server is running
uv run --no-sync python yapctl.py server       # Start ONLY the server (background)
uv run --no-sync python yapctl.py logs         # View server logs
uv run --no-sync python yapctl.py logs -f      # Follow server logs
uv run --no-sync python yapctl.py restart      # Restart server + dictation

After the server is running, you can start/stop the dictation client independently:

# Terminal 2: start client (connects to server, no model loading)
uv run --no-sync python dictate_client.py

Standalone mode

If you don't want to run a separate server:

uv run --no-sync python dictate.py

Options

# Custom hotkey (default: Right Ctrl)
--key scroll_lock
--key f13

# Specific microphone (list with: python -m sounddevice)
--device 5

# Custom server address (client only)
--host 127.0.0.1 --port 9876

Batch file transcription

# Edit transcribe.py to point at your audio files, then:
uv run --no-sync python transcribe.py

Performance (RTX 3060 Ti, 8GB VRAM)

Key config notes

• exclude-newer-package = { torch = false } — Exempts torch from global exclude-newer date filter (PyTorch CUDA wheels lack upload dates)
• pytorch-cu128 index — Routes torch to NVIDIA's CUDA 12.8 wheel index
• uv run --no-sync — Prevents uv from reverting manually-installed NeMo

Development

# Lint
uv run ruff check .

# Format
uv run ruff format .

# Type check
uv run ty check

Dependencies

• Python 3.10
• PyTorch 2.11.0+cu128
• NeMo 2.8.0rc0 (from main branch)
• Model (nemo): nvidia/parakeet-unified-en-0.6b (~1.2 GB, cached in ~/.cache/huggingface/)

llama.cpp backend (alternative)

You can also use Qwen3-ASR via llama-server instead of NeMo:

# Install llama.cpp and start the server
llama-server -hf ggml-org/Qwen3-ASR-0.6B-GGUF

# Use the llama backend
uv run --no-sync python yapctl.py --backend llama start

The llama backend requires no NeMo/PyTorch — just llama-server running separately. Uses ~800MB VRAM (vs ~2GB for NeMo) and has comparable speed.