NVIDIA Releases 600M Parakeet for Speech Recognition
The new FastConformer model uses a specialized training technique to improve transcription accuracy in noisy, real-world environments.
NVIDIA has released a new open model for automatic speech recognition (ASR) called Parakeet TDT 0.6B. As part of its NeMo toolkit for conversational AI, this 600-million-parameter model is designed to transcribe speech across multiple languages with high accuracy.
The model's architecture and training method are key to its performance. It uses a FastConformer encoder, which is known for its efficiency in processing audio sequences. The "TDT" in its name signifies Transducer with Denoising Training, a technique that makes the model more robust by training it to ignore noise and focus on the primary speech signal, a common challenge in real-world applications.
This release provides developers with a powerful and relatively lightweight tool for building speech-enabled products. With a permissive CC-BY-4.0 license, Parakeet can be freely used and modified for both research and commercial projects. Its 0.6-billion-parameter size makes it more accessible to deploy than the massive, multi-billion-parameter systems that often dominate ASR research.
By open-sourcing a specialized model like Parakeet, NVIDIA is contributing a significant building block to the conversational AI ecosystem. Developers interested in experimenting with the model can find the weights and usage instructions on the official Hugging Face repository.
Sources
- Visit
nvidia/parakeet-tdt-0.6b-v3
Hugging Face
More from NVIDIA
All NVIDIA releases →
NVIDIA releases compact Dory 1.0 4B text model
The new Nemotron Labs model packs a 4-billion-parameter language model trained on NVIDIA's Nemotron datasets.
NVIDIA's Nemotron 3.5 Super VL Pairs Vision With MoE Efficiency
The 120B-parameter vision-language model activates just 12B per token and ships in NVIDIA's NVFP4 format for leaner inference.
NVIDIA's Nemotron 3.5 Super VL Pairs Vision With MoE
The new vision-language model carries 120B total parameters but activates just 12B per token, aiming to keep multimodal inference efficient.
More in Speech → Text
All Speech → Text →IBM's Granite Speech 5.0 Adds Speaker Attribution
The Apache-2.0 speech model handles transcription, translation, and figuring out who said what.
Qwen-Audio-3.1-ASR Adds Speaker-Aware Transcription
Alibaba's new open audio-language model transcribes and diarizes speech across 30 languages under an Apache 2.0 license.
NVIDIA's Nemotron 3.5 Transcribe Brings Promptable ASR
A sparse 30B mixture-of-experts model handles multilingual transcription, translation, and speaker diarization from a single system.
0 comments
No comments yet. Be the first to weigh in.