nvidia /

parakeet-tdt-0.6b-v2

A 600M-parameter English speech-to-text model with punctuation, capitals and word timestamps.

1.6k likes
182.6k downloads a month
Some experience
Speech

Open on Hugging Face (opens in a new tab)

Parakeet TDT 0.6B V2 is an automatic speech recognition (ASR) model with 600 million parameters, built on NVIDIA's FastConformer encoder and a TDT decoder. You give it English audio (.wav or .flac, 16kHz, single channel) and it returns text with punctuation and capital letters. It can handle audio segments up to 24 minutes in one pass, and it can give word-level timestamps.

It was trained on about 120,000 hours of English speech, mostly pseudo-labeled data plus 10,000 hours of human-transcribed audio. Its average word error rate on the Open ASR Leaderboard datasets is 6.05%, and it does best on clean read speech (1.69% on LibriSpeech test-clean). It does worse on meetings (11.16% on AMI), and noise hurts it: average WER rises to 20.26% at an SNR of -5. The card says transcripts may not be 100% accurate, and that accuracy varies with accent, noise, domain and speech type. Words missing from its vocabulary are unlikely to be recognized. It is English only, designed for NVIDIA GPU systems, with Linux as the preferred operating system.

What it does well

Punctuation and capitalization built in
The output text already includes punctuation and capital letters, so you don't need a separate step to tidy it up.
Word-level timestamps
You can ask for timestamps at character, word and segment level, which is useful for subtitles or jumping to a point in audio.
Long audio in one pass
With full attention, it can transcribe audio segments up to 24 minutes in a single pass. The card notes that more RAM supports longer input.
Strong scores on clean speech
Average WER is 6.05% across eight benchmark datasets, with 1.69% on LibriSpeech test-clean and 2.17% on SPGI Speech.
Holds up reasonably on phone-style audio
With 8kHz μ-law telephony-style audio, average WER moves from 6.05% to 6.32%.

What you could build

Generate subtitles for a video by using the word or segment timestamps
Transcribe recorded lectures or meetings into searchable text
Add a speech-to-text step to a voice assistant or conversational app
Fine-tune it on your own English audio, since the card says it can be used as a pre-trained checkpoint for fine-tuning
Compare its error rates on noisy versus clean recordings as a small experiment

Get started

Install NeMo after PyTorch, load the model with from_pretrained, then pass a list of audio file paths to transcribe and print the text. Add timestamps=True if you want word or segment times.

pip install -U nemo_toolkit["asr"]

asr_model = nemo_asr.models.ASRModel.from_pretrained(model_name="nvidia/parakeet-tdt-0.6b-v2")

output = asr_model.transcribe(['2086-149220-0033.wav'])

print(output[0].text)

Commands are copied from the model card. Read it on Hugging Face for the full setup.