typeout

CLI utility to transcribe audio or video to text using ASR (Automatic Speech Recognition).

A single self-contained script that auto-detects GPU availability and uses the appropriate backend.

Backend	Hardware	Models
OpenAI Whisper	CPU/GPU	tiny, base, small, medium, large
Distil-Whisper	CPU/GPU	distil-large-v3, distil-medium.en
Cohere Transcribe 2B	CPU/GPU	14 languages
NVIDIA NeMo ASR	GPU	Canary-1B-v2, Canary-Qwen-2.5B, Parakeet-0.6B

Input can be a local file (any format ffmpeg supports — mp3, wav, flac, mp4, mkv, webm, ...), a URL, or a YouTube video ID.

Features

• Single command — auto-detects GPU and uses the right backend
• Accepts audio or video in any format ffmpeg can read
• Transcribe YouTube videos by URL or video ID
• Multiple models: Whisper, Cohere Transcribe, NeMo ASR
• Multilingual transcription (14 languages with Cohere, 25 with NeMo Canary)
• Speech translation (NeMo Canary models only)
• Caches downloaded audio and transcripts for instant repeat lookups
• Transcript on stdout, diagnostics on stderr (pipe-friendly)

Installation

The only prerequisite is uv. The script is self-contained and manages all Python dependencies automatically on first run. A shortcut to the script can be found at https://tinyurl.com/typeout.

# Install uv
cargo install uv  # or: pip install uv

# Download typeout
curl -O https://raw.githubusercontent.com/miku/typeout/refs/heads/main/typeout

# Make executable and put somewhete into PATH
chmod +x typeout
mv typeout ~/.local/bin/

That's it. No virtualenvs, no pip install, no setup.py.

Usage

# Transcribe a local file
typeout recording.mp3
typeout lecture.mp4

# Write to file
typeout podcast.flac -o transcript.txt

# Use different Whisper model (CPU)
typeout recording.mp3 --model small

# Use Cohere Transcribe (CPU/GPU, requires Hugging Face login)
typeout recording.mp3 --model cohere-transcribe --lang en
typeout lecture.mp4 --model cohere-transcribe --lang ja

# GPU models (auto-detected on systems with NVIDIA GPU)
typeout recording.mp3 --model canary-qwen-2.5b
typeout lecture.mp4 --model parakeet-0.6b

# Use Whisper on GPU (fp16 acceleration)
typeout recording.mp3 --model large

# Multilingual: set source language
typeout interview.wav --lang de

# Translation: German audio to English text (NeMo Canary only)
typeout interview.wav --lang de --target-lang en

# From a URL or YouTube ID
typeout "https://www.youtube.com/watch?v=dQw4w9WgXcQ"
typeout dQw4w9WgXcQ

# Check external tools
typeout --check

# Clear cache
typeout --clear-cache

# List models
typeout --list-models

Models

Model	Size	Languages	Notes
base · CPU default	~140MB	multilingual	Whisper, good balance
tiny	~40MB	multilingual	Whisper, fastest
small	~460MB	multilingual	Whisper
medium	~1.5GB	multilingual	Whisper
large	~2.9GB	multilingual	Whisper, highest accuracy
distil-large-v3	~750MB	multilingual	Distil-Whisper, 6x faster than large
distil-medium.en	~400MB	English only	Distil-Whisper, fast
cohere-transcribe	~4.1GB	14 languages	Cohere, high accuracy, requires HF login
canary-1b-v2 · GPU default	~6.4GB	25 languages	NVIDIA only, NeMo, multilingual, translation
canary-qwen-2.5b	~5.1GB	multilingual	NVIDIA only, NeMo, highest quality, SLM
parakeet-0.6b	~2.5GB	English only	NVIDIA only, NeMo, fast and lightweight

Cohere Transcribe setup (gated model):

# 1. Accept terms at: https://huggingface.co/CohereLabs/cohere-transcribe-03-2026
# 2. Login to Hugging Face
huggingface-cli login

How it works

The typeout script is an amalgamation — it contains both CPU and GPU Python scripts embedded within it. On first run:

1. Detects if NVIDIA GPU is available (nvidia-smi)
2. Extracts the appropriate Python script to ~/.cache/typeout/
3. Runs it with uv run, which installs dependencies automatically

Subsequent runs reuse the extracted script (unless you upgrade typeout).

Caching

Downloaded audio (for URLs) and transcripts are cached in ~/.cache/typeout/ (respects $XDG_CACHE_HOME).

• URLs: keyed by URL — same URL hits cache instantly
• Local files: keyed by path + modification time + size — cache invalidates on edit
• Transcripts: keyed by source + model + language — different models/languages get separate entries

Use --no-cache to bypass, --clear-cache to remove all cached data.

Dependencies

• uv — runs the script and manages Python dependencies
• ffmpeg — audio extraction and normalization (check with --check)
• nvidia-smi — GPU detection (auto-detected)
• huggingface-cli login — required for Cohere Transcribe (gated model)

Models are cached in ~/.local/share/typeout/ (respects $XDG_DATA_HOME).

Examples

$ typeout --lang de \
    https://swr-pd.ard-mcdn.de/swr/swrkultur/hoerspiel/ard-hoerspiel-speicher/2303264.mp3

Transcribing a 7h+ audio book [UBIK] takes about 14 minutes on a 70W RTX 4000 SFF with canary-1b-v2.

$ typeout https://www.youtube.com/watch?v=P1qMKFMrpro # UBIK audiobook

Image

From: https://etc.usf.edu/clipart/