Parakeet (NeMo Speech)
NVIDIA's speech models and toolkit: Apache-2.0 code, and the Parakeet TDT 0.6B v3 model under CC-BY-4.0 with 25 European languages, automatic language detection, punctuation and word timestamps.
What it does
NeMo Speech is NVIDIA's toolkit "for researchers and PyTorch developers working on Speech models", covering speech recognition, text to speech and speech language models. The model most people mean when they say Parakeet is parakeet-tdt-0.6b-v3: "a 600-million-parameter multilingual automatic speech recognition (ASR) model designed for high-throughput speech-to-text transcription", covering 25 European languages and detecting the language automatically.
The model card lists automatic punctuation and capitalisation, word-level and segment-level timestamps, and long audio: up to 24 minutes with full attention on an A100 80 GB, or up to 3 hours with local attention. The README also names newer models: Nemotron ASR Streaming with 40 languages and controllable latency from 80 milliseconds to 1 second, a unified English model that does offline and streaming inference in one network, and Canary for recognition plus translation.
What it replaces
Parakeet replaces Deepgram and AssemblyAI for high-volume transcription of European languages on your own NVIDIA hardware, with the timestamps and punctuation those APIs return.
The label is open weights rather than open source because the weights are CC-BY-4.0, a Creative Commons licence that is not on the OSI list, and the card says training combined a public corpus with an in-house NVIDIA dataset. CC-BY-4.0 does permit commercial use, with attribution. The code around it is Apache-2.0.
Where it falls short: the toolkit expects an NVIDIA GPU. Speaker labelling and quantisation are not described in the README. The desktop dictation apps on this site (Handy, VoiceInk, OpenWhispr, Vibe) already bundle Parakeet for CPU use, which is the easier route for a single machine.
How it runs
pip install 'nemo-toolkit[asr,tts]', a source install with uv, or the NGC Docker container run with all GPUs attached. Python 3.12 or later and PyTorch 2.7 or later. The README states "NVIDIA GPU + CUDA (required for training; recommended for inference)". The model card documents a Transformers path with a CPU fallback and a NeMo-Speech.cpp runtime. NVIDIA links a hosted NIM for the streaming model; no prices are stated.
Who it fits
A team with NVIDIA GPUs transcribing many hours of European-language audio a day. A team on Macs or CPUs should use whisper.cpp, or one of the desktop apps that already ships Parakeet.
Limits the README and card state
Some checkpoints need loading with weights_only off, which the README warns "can have the risk of arbitrary code execution" from untrusted files. The card: "Transcripts may be not 100% accurate", and words absent from the vocabulary "not likely to be recognized".
Sources
- NeMo Speech README and Apache-2.0 LICENSE: github.com/NVIDIA-NeMo/Speech, read 2026-09-16.
- Parakeet TDT 0.6B v3 model card, CC-BY-4.0: huggingface.co/nvidia/parakeet-tdt-0.6b-v3, read 2026-09-16.
More speech-to-text engines
OpenAI's general-purpose speech recognition model, released with code and weights under MIT, in six sizes from 39 million to 1.55 billion parameters, with multilingual transcription, translation to English and language detection.
A dependency-free C and C++ port of Whisper under MIT that runs on CPUs, Apple Silicon and phones, with quantised models, an HTTP server and bindings for a dozen languages.
Batched Whisper transcription under BSD-2-Clause with word-level timestamps from forced alignment and speaker labels from pyannote, the closest open-source match to a hosted transcription API's output.
An offline speech recognition toolkit under Apache-2.0 with streaming, models of about 50 MB, 20+ languages, a reconfigurable vocabulary and speaker identification, from small boards to clusters.
A runtime under Apache-2.0 from the next-generation Kaldi team that runs speech to text, text to speech, speaker diarization and voice activity detection locally on CPUs, NPUs and phones, with no internet connection.
An on-device speech toolkit under MIT built for live streaming, with speech-to-text models trained from scratch in sizes down to 1 MB, one library across Python, the browser, phones and desktops.
Added September 16, 2026. Found an error on this page? Write to reveneau@licheo.com and it gets fixed in the next weekly pass.
Running one of these inside your own environment, with your own data and your own security rules, is the kind of work Reveneau does. Read how a forward deployed engagement works.