parakeet

name	parakeet
description	Local speech-to-text using NVIDIA Parakeet TDT (NeMo). 600M-param multilingual ASR with automatic punctuation/capitalization, word-level timestamps, and ~3380x realtime speed on GPU. Supports 25 European languages with auto-detection, long-form audio up to 3 hours, streaming output, SRT/VTT subtitles, batch processing, and URL/YouTube input.
version	1.0.0
author	ThePlasmak
tags	["audio","transcription","speech-to-text","parakeet","nemo","nvidia","ml","cuda","gpu","subtitles","multilingual","asr","timestamps"]
platforms	["linux","wsl2"]
metadata	{"openclaw":{"emoji":"🦜","requires":{"bins":["python3"],"optionalBins":["ffmpeg","yt-dlp"]}}}

🦜 Parakeet — NVIDIA Speech-to-Text

Local speech-to-text using NVIDIA's Parakeet TDT models via the NeMo toolkit. The default parakeet-tdt-0.6b-v3 is a 600-million-parameter multilingual ASR model that delivers state-of-the-art accuracy with automatic punctuation and capitalization, word-level timestamps, and an insane ~3380× realtime inference speed on GPU.

Only needs ~2GB VRAM to run — plenty of room on an RTX 3070 (8GB).

When to Use

Use this skill when you need:

• Best-accuracy transcription — Parakeet v3 achieves 6.34% average WER on the Open ASR Leaderboard
• Automatic punctuation & capitalization — no post-processing needed, output is clean and readable
• European language transcription — 25 languages with automatic detection (no prompting required)
• Word/segment/char timestamps — precise timing at every granularity level
• Long audio files — up to 24 min with full attention, up to 3 hours with --long-form local attention
• Streaming output — see segments as they're transcribed with --streaming
• Subtitle generation — SRT and VTT formats with word-level line splitting
• Batch processing — glob patterns, directories, skip-existing support with ETA
• URL/YouTube transcription — auto-downloads via yt-dlp
• Blazing speed — ~3380× realtime means a 10-minute file transcribes in under 0.2 seconds on GPU

Trigger phrases: "transcribe with parakeet", "nemo transcribe", "nvidia speech to text", "parakeet transcribe", "transcribe this audio", "convert speech to text", "what did they say", "make a transcript", "audio to text", "subtitle this video", "transcribe in [language]", "European language transcription", "best accuracy transcription", "high accuracy speech to text", "transcribe long audio", "transcribe lecture", "transcribe meeting", "word timestamps"

🦜 Parakeet vs 🗣️ faster-whisper — When to Use Which

Feature	Parakeet	faster-whisper
Accuracy	✅ Best (6.34% avg WER)	Good (distil: 7.08% WER)
Speed	✅ ~3380× realtime	~20× realtime
Auto punctuation	✅ Built-in	❌ Requires post-processing
Languages	25 European	✅ 99+ worldwide
Diarization	❌ Not supported	✅ pyannote speaker ID
Chapters/search	❌ Not supported	✅ Chapter detection, search
Output formats	text, JSON, SRT, VTT	✅ text, JSON, SRT, VTT, ASS, LRC, TTML, CSV, TSV, HTML
Translation	❌ Not supported	✅ Any language → English
VRAM usage	~2GB	~1.5GB (distil)
Long audio	✅ Up to 3 hours	Limited by VRAM
Streaming	✅ Chunked inference	✅ Segment streaming
Noise handling	Good (built-in robustness)	✅ --denoise, --normalize
Filler removal	❌	✅ --clean-filler

Rule of thumb:

• Parakeet when you want: best accuracy, fastest speed, auto-punctuation, European languages
• faster-whisper when you need: diarization, 99+ languages, translation, chapters/search, more output formats, audio preprocessing

⚠️ Agent guidance — keep invocations minimal:

CORE RULE: default command (./scripts/transcribe audio.wav) is the fastest path — add flags only when the user explicitly asks for that capability.

Transcription:

• Only add --timestamps if the user asks for word-level or segment-level timestamps (auto-enabled for srt, vtt, json formats)
• Only add --format srt/vtt if the user asks for subtitles/captions in that format
• Only add --format json if the user wants structured/programmatic output
• Only add --long-form if the audio is longer than 24 minutes
• Only add --streaming if the user wants live/progressive output for long files
• Only add -l/--language CODE if the user specifies a language (auto-detection is usually fine)
• Only add --model if the user wants a specific model variant (v2, 1.1b)
• Only add --device cpu if GPU is not available or user requests CPU
• Only add --batch-size N if the user reports OOM errors
• Only add --max-words-per-line or --max-chars-per-line for subtitle readability on long segments

Batch processing:

• Only add --skip-existing when resuming interrupted batch jobs
• For batch output, always use -o <dir> (directory, not file)
• ETA is shown automatically for batch jobs (no flag needed)

Output format for agent relay:

• Text (default) → safe to show directly to user for short files; summarise for long ones
• JSON → useful for programmatic post-processing; not ideal to paste in full to user
• SRT/VTT → always write to -o file; tell the user the output path, never paste raw subtitle content

When NOT to use:

• Need speaker diarization → use faster-whisper with --diarize
• Non-European languages (Asian, African, etc.) → use faster-whisper
• Need translation to English → use faster-whisper with --translate
• Cloud-only environments without local compute

Quick Reference

Task	Command	Notes
Basic transcription	./scripts/transcribe audio.wav	Auto-punctuated, GPU-accelerated
Transcribe MP3	./scripts/transcribe audio.mp3	Auto-converts via ffmpeg
SRT subtitles	./scripts/transcribe audio.wav --format srt -o subs.srt	Timestamps auto-enabled
VTT subtitles	./scripts/transcribe audio.wav --format vtt -o subs.vtt	WebVTT format
Word timestamps	./scripts/transcribe audio.wav --timestamps --format json -o out.json	Word/segment/char level
JSON output	./scripts/transcribe audio.wav --format json -o result.json	Full metadata + timestamps
Long audio (>24 min)	./scripts/transcribe lecture.wav --long-form	Local attention, up to ~3 hours
Streaming output	./scripts/transcribe audio.wav --streaming	Print segments as transcribed
Specify language	./scripts/transcribe audio.wav -l fr	Validate language (v3 auto-detects)
YouTube/URL	./scripts/transcribe https://youtube.com/watch?v=...	Auto-downloads via yt-dlp
Batch process	./scripts/transcribe *.wav -o ./transcripts/	Output to directory
Batch with skip	./scripts/transcribe *.wav --skip-existing -o ./out/	Resume interrupted batches
English-only model	./scripts/transcribe audio.wav -m nvidia/parakeet-tdt-0.6b-v2	v2 English-only
Larger model	./scripts/transcribe audio.wav -m nvidia/parakeet-tdt-1.1b	1.1B param English model
CPU mode	./scripts/transcribe audio.wav --device cpu	If no GPU available
Quiet mode	./scripts/transcribe audio.wav -q	Suppress progress messages
Subtitle word wrapping	./scripts/transcribe audio.wav --format srt --max-words-per-line 8 -o subs.srt	Split long subtitle cues
Char-based wrapping	./scripts/transcribe audio.wav --format srt --max-chars-per-line 42 -o subs.srt	Character limit per line
Show version	./scripts/transcribe --version	Print NeMo version
System check	./setup.sh --check	Verify GPU, Python, NeMo, ffmpeg
Upgrade NeMo	./setup.sh --update	Upgrade without full reinstall

Model Selection

Available Models

Model	Params	Languages	Speed	Use Case
nvidia/parakeet-tdt-0.6b-v3	600M	25 EU languages	~3380× RT	Default, best multilingual
nvidia/parakeet-tdt-0.6b-v2	600M	English only	~3380× RT	English-only, slightly better English WER
nvidia/parakeet-tdt-1.1b	1.1B	English only	Slower	Maximum English accuracy

Supported Languages (v3)

Bulgarian (bg), Croatian (hr), Czech (cs), Danish (da), Dutch (nl), English (en), Estonian (et), Finnish (fi), French (fr), German (de), Greek (el), Hungarian (hu), Italian (it), Latvian (lv), Lithuanian (lt), Maltese (mt), Polish (pl), Portuguese (pt), Romanian (ro), Slovak (sk), Slovenian (sl), Spanish (es), Swedish (sv), Russian (ru), Ukrainian (uk)

Language is auto-detected — no prompting or configuration needed. The model identifies the language from the audio content automatically.

Accuracy Benchmarks (v3)

Benchmark	WER
Open ASR Leaderboard (avg)	6.34%
LibriSpeech test-clean	1.93%
LibriSpeech test-other	3.59%
AMI	11.31%
GigaSpeech	9.59%
SPGI Speech	3.97%
TED-LIUM v3	2.75%
VoxPopuli	6.14%

Setup

Linux / WSL2

# Full install (creates venv, installs NeMo + CUDA PyTorch)
./setup.sh

# Verify installation
./setup.sh --check

# Upgrade NeMo toolkit
./setup.sh --update

Requirements:

• Python 3.10+
• NVIDIA GPU with CUDA (strongly recommended; CPU works but is much slower)
• ffmpeg (recommended for mp3/m4a/mp4 format conversion; not needed for .wav/.flac)
• Optional: yt-dlp (for URL/YouTube input)

Platform Support

Platform	Acceleration	Notes
Linux + NVIDIA GPU	CUDA	~3380× realtime 🚀
WSL2 + NVIDIA GPU	CUDA	~3380× realtime 🚀
Linux (no GPU)	CPU	Functional but slow
macOS	❌	NeMo has limited macOS support

VRAM: The 600M model only needs ~2GB VRAM to load. An RTX 3070 (8GB) has plenty of headroom.

GPU Support

The setup script auto-detects your GPU and installs PyTorch with CUDA. Always use GPU if available.

If setup didn't detect your GPU, manually install PyTorch with CUDA:

# For CUDA 12.x
uv pip install --python ./venv/bin/python torch torchaudio --index-url https://download.pytorch.org/whl/cu121

• WSL2 users: Ensure you have the NVIDIA CUDA drivers for WSL installed on Windows

Usage

# Basic transcription (auto-punctuated, auto-capitalized)
./scripts/transcribe audio.wav

# Transcribe an MP3 (auto-converts to WAV via ffmpeg)
./scripts/transcribe recording.mp3

# SRT subtitles
./scripts/transcribe audio.wav --format srt -o subtitles.srt

# WebVTT subtitles
./scripts/transcribe audio.wav --format vtt -o subtitles.vtt

# Full JSON with word/segment/char timestamps
./scripts/transcribe audio.wav --timestamps --format json -o result.json

# Transcribe from YouTube URL
./scripts/transcribe https://youtube.com/watch?v=dQw4w9WgXcQ

# Long lecture (>24 min, up to 3 hours)
./scripts/transcribe lecture.wav --long-form

# Streaming mode (print segments as they're transcribed)
./scripts/transcribe audio.wav --streaming

# Batch process a directory
./scripts/transcribe ./recordings/ -o ./transcripts/

# Batch with glob, skip already-done files
./scripts/transcribe *.wav --skip-existing -o ./transcripts/

# Use English-only v2 model
./scripts/transcribe audio.wav --model nvidia/parakeet-tdt-0.6b-v2

# JSON output with metadata
./scripts/transcribe audio.wav --format json -o result.json

# Specify expected language (validation only; v3 auto-detects)
./scripts/transcribe audio.wav --language fr

Options

Input:
  AUDIO                 Audio file(s), directory, glob pattern, or URL
                        Native: .wav, .flac (16kHz mono preferred)
                        Converts via ffmpeg: .mp3, .m4a, .mp4, .mkv, .ogg, .webm, .aac, .wma, .avi, .opus
                        URLs auto-download via yt-dlp (YouTube, direct links, etc.)

Model & Language:
  -m, --model NAME      NeMo ASR model (default: nvidia/parakeet-tdt-0.6b-v3)
                        Also: nvidia/parakeet-tdt-0.6b-v2 (English), nvidia/parakeet-tdt-1.1b (larger)
  -l, --language CODE   Expected language code, e.g. en, es, fr (v3 auto-detects if omitted)
                        Used for validation; does not force the model

Output Format:
  -f, --format FMT      text | json | srt | vtt (default: text)
  --timestamps          Enable word/segment/char timestamps (auto-enabled for srt, vtt, json)
  --max-words-per-line N  For SRT/VTT, split segments into sub-cues of at most N words
  --max-chars-per-line N  For SRT/VTT, split lines so each fits within N characters
                        Takes priority over --max-words-per-line when both are set
  -o, --output PATH     Output file or directory (directory for batch mode)

Long-Form & Streaming:
  --long-form           Enable local attention for audio >24 min (up to ~3 hours)
                        Changes attention model to rel_pos_local_attn with [256,256] context
  --streaming           Print segments as they are transcribed (chunked inference)

Inference:
  --batch-size N        Inference batch size (default: 32; reduce if OOM)

Batch Processing:
  --skip-existing       Skip files whose output already exists (batch mode)

Device:
  --device DEV          auto | cpu | cuda (default: auto)
  -q, --quiet           Suppress progress and status messages

Utility:
  --version             Print version info and exit

Output Formats

Text (default)

Clean, automatically punctuated and capitalized transcript:

Hello, welcome to the meeting. Today we'll discuss the quarterly results
and our plans for next quarter.

JSON (--format json)

Full metadata including segments, word/char timestamps, and performance stats:

{
  "file": "audio.wav",
  "text": "Hello, welcome to the meeting.",
  "duration": 600.5,
  "segments": [
    {
      "start": 0.0,
      "end": 2.5,
      "text": "Hello, welcome to the meeting.",
      "words": [
        {"word": "Hello,", "start": 0.0, "end": 0.4},
        {"word": "welcome", "start": 0.5, "end": 0.9},
        {"word": "to", "start": 0.95, "end": 1.1},
        {"word": "the", "start": 1.15, "end": 1.3},
        {"word": "meeting.", "start": 1.35, "end": 2.0}
      ]
    }
  ],
  "words": [...],
  "chars": [...],
  "stats": {
    "processing_time": 0.18,
    "realtime_factor": 3380.0
  }
}

SRT (--format srt)

Standard subtitle format for video players:

1
00:00:00,000 --> 00:00:02,500
Hello, welcome to the meeting.

2
00:00:02,800 --> 00:00:05,200
Today we'll discuss the quarterly results.

VTT (--format vtt)

WebVTT format for web video players:

WEBVTT

1
00:00:00.000 --> 00:00:02.500
Hello, welcome to the meeting.

2
00:00:02.800 --> 00:00:05.200
Today we'll discuss the quarterly results.

Long-Form Audio

Parakeet v3 supports two attention modes:

Mode	Max Duration	Flag	Notes
Full attention (default)	~24 min	(none)	Best accuracy, requires more VRAM
Local attention	~3 hours	--long-form	Slightly reduced accuracy, much lower VRAM

For audio over 24 minutes, use --long-form:

./scripts/transcribe long-lecture.wav --long-form

This changes the attention model to rel_pos_local_attn with a context window of [256, 256], allowing the model to process very long audio without running out of memory.

URL & YouTube Input

Pass any URL as input — audio is downloaded automatically via yt-dlp:

# YouTube video
./scripts/transcribe https://youtube.com/watch?v=dQw4w9WgXcQ

# Direct audio URL
./scripts/transcribe https://example.com/podcast.mp3

# With options
./scripts/transcribe https://youtube.com/watch?v=... -l en --format srt -o subs.srt

Requires yt-dlp (checks PATH and ~/.local/share/pipx/venvs/yt-dlp/bin/yt-dlp).

Batch Processing

Process multiple files at once with glob patterns, directories, or multiple paths:

# All WAV files in current directory
./scripts/transcribe *.wav -o ./transcripts/

# Specific directory
./scripts/transcribe ./recordings/ -o ./transcripts/

# Resume interrupted batch
./scripts/transcribe *.wav --skip-existing -o ./transcripts/

# SRT subtitles for all files
./scripts/transcribe *.wav --format srt -o ./subtitles/

ETA is shown automatically for batch jobs after the first file completes.

Audio Format Support

Format	Support	Notes
.wav	✅ Native	Preferred; 16kHz mono
.flac	✅ Native	Lossless; works directly
.mp3	🔄 Converts	Requires ffmpeg
.m4a	🔄 Converts	Requires ffmpeg
.mp4	🔄 Converts	Requires ffmpeg
.ogg	🔄 Converts	Requires ffmpeg
.webm	🔄 Converts	Requires ffmpeg
.aac	🔄 Converts	Requires ffmpeg
.wma	🔄 Converts	Requires ffmpeg
.opus	🔄 Converts	Requires ffmpeg
.mkv	🔄 Converts	Requires ffmpeg

Non-native formats are automatically converted to 16kHz mono WAV using ffmpeg before transcription.

License

The Parakeet TDT v3 model is released under CC-BY-4.0 — free for commercial and non-commercial use.