What it is
Parakeet TDT 0.6B v3 is a 600-million-parameter speech-recognition model from NVIDIA, published on Hugging Face on August 14, 2025 under the CC BY 4.0 licence, which allows commercial use with attribution. It extends the English-only v2 to 25 European languages, from Bulgarian to Ukrainian, and detects the language automatically.
It adds punctuation, capitalisation and word, segment and character timestamps, and handles up to 24 minutes of audio with full attention on an A100 80GB, or up to 3 hours with local attention. NVIDIA says it needs at least 2GB of memory to load. It runs through NVIDIA's NeMo toolkit, Hugging Face Transformers or the lightweight NeMo-Speech.cpp runtime, and NeMo includes a chunked streaming script.
Our review
Parakeet v3 is the fastest transcription model Artificial Analysis measures. Its Together AI deployment runs at 617.7 times real time, ahead of Deepgram Nova-3 at 546.2, which means an hour of audio comes back in about six seconds. Accuracy is ordinary: a 4.5% AA-WER ties Whisper large-v3 on Together and sits 39th of 62 entries.
The hosted case is weaker than the self-hosted one. Together charges $1.50 per 1,000 minutes, while Whisper large-v3-turbo on Groq costs $0.67 at 4.6% error, so a buyer who only wants an API pays more than twice as much for speed they may not need. Accuracy leaders have also moved away: Microsoft's MAI-Transcribe-2 scores 2.0% at $1.67, and Mistral's Voxtral Small Transcribe 3 reaches 2.1%.
The model earns its place on your own hardware. At 600 million parameters and a 2GB memory floor it fits beside other workloads on a single GPU, timestamps come free, and NVIDIA's 3-hour local-attention mode handles a full meeting or court session without chunking logic. NVIDIA trained it on 128 A100s, and the Granary corpus behind it is published, which is more training transparency than most commercial rivals give.
The language list is the boundary. It covers 25 European languages and nothing else, so a contact centre taking calls in Arabic, Hindi or Mandarin needs Whisper or a hosted model alongside it, and the CC BY 4.0 attribution requirement is one more line of legal review than Whisper's MIT.
Where it earns its place
An ML engineer at a German media monitoring firm processing 10,000 hours of broadcast a month
Transcribe and timestamp European TV and radio on owned GPUs
Run Parakeet v3 in NeMo on existing GPUs. At the 617.7 times real time Artificial Analysis measured on Together, a month's 10,000 hours is about 16 hours of processing; the same 600,000 minutes on Together would cost about $900.
A product team building live captions for a pan-European video-conferencing app
Caption calls in the speaker's language without asking users to pick one
Automatic language detection across 25 languages and NeMo's chunked streaming script cover the basics; English, French and German calls work, but a participant speaking Turkish or Arabic will get nothing usable.
A legal-tech startup handling court recordings that cannot leave the EU
Produce timestamped transcripts of hearings up to three hours long
Local attention handles a 3-hour file in one pass on-premise, so audio never reaches a third-party API. The trade-off is a 4.5% error rate, so names and figures still need a human check.
Before you commit
- Setup
- pip install NeMo, load nvidia/parakeet-tdt-0.6b-v3 with from_pretrained and call transcribe with timestamps=True; NVIDIA lists Linux as the supported platform.
- Integrations
- NeMo, Hugging Face Transformers and NeMo-Speech.cpp; Together AI for a hosted endpoint. NVIDIA's NIM page covers v2, not v3.
- Data and privacy
- CC BY 4.0 allows commercial use with attribution and no revenue cap. Self-hosted, audio stays on your hardware; Together's retention terms apply to its hosted version.
- Cost at scale
- Self-hosted it needs only 2GB to load, so it runs on any recent NVIDIA GPU. Hosted on Together, 600,000 minutes a month costs about $900; Whisper turbo on Groq would be about $400 for the same volume.
- Limits
- 25 European languages only; 24 minutes per file with full attention on an A100 80GB, 3 hours with local attention; 4.5% AA-WER.
What's good
- 617.7 times real time on Together AI, first on Artificial Analysis's board
- 600M parameters, loads in 2GB of memory
- Word, segment and character timestamps; up to 3 hours per file with local attention
What's not
- 25 European languages only
- 4.5% AA-WER, about twice the error of Scribe v2 or Voxtral Small Transcribe 3
- Hosted on Together at $1.50 per 1,000 minutes, more than Whisper turbo on Groq at $0.67
- CC BY 4.0 requires attribution
Best for
Developers transcribing large volumes of European-language audio who prize speed and low cost.
Alternatives
Parakeet TDT 0.6B v3: common questions
- Is Parakeet TDT 0.6B v3 worth it?
- Parakeet v3 is the speed record in transcription, with Whisper-class accuracy and a hard border around Europe. We rate it 4.1 out of 5.
- How much does Parakeet TDT 0.6B v3 cost?
- Free under CC BY 4.0. Together AI's hosted version costs $0.0015 a minute ($1.50 per 1,000 minutes).
- What are the downsides of Parakeet TDT 0.6B v3?
- 25 European languages only; 4.5% AA-WER, about twice the error of Scribe v2 or Voxtral Small Transcribe 3; hosted on Together at $1.50 per 1,000 minutes, more than Whisper turbo on Groq at $0.67; cC BY 4.0 requires attribution.
- Who is Parakeet TDT 0.6B v3 best for?
- Developers transcribing large volumes of European-language audio who prize speed and low cost.
- How much does Parakeet TDT 0.6B v3 cost for a team or in production?
- Self-hosted it needs only 2GB to load, so it runs on any recent NVIDIA GPU. Hosted on Together, 600,000 minutes a month costs about $900; Whisper turbo on Groq would be about $400 for the same volume.
- Is Parakeet TDT 0.6B v3 safe for company data?
- CC BY 4.0 allows commercial use with attribution and no revenue cap. Self-hosted, audio stays on your hardware; Together's retention terms apply to its hosted version.
- What are the best Parakeet TDT 0.6B v3 alternatives?
- The closest alternatives we have reviewed are Whisper large-v3-turbo, VibeVoice-ASR and Deepgram Nova-3.
Answers checked Oct 23, 2026.
Make Parakeet TDT 0.6B v3? Show this rating on your site
