NVIDIA has made significant strides in the field of speech recognition through the release of cutting-edge models that set new standards for accuracy and performance. At the forefront is the Parakeet TDT 0.6B v2, a 600-million-parameter automatic speech recognition model. This model currently leads the Hugging Face ASR leaderboard with a remarkable word error rate of 6.05%, offering ultra-fast inference that is 50 times quicker than alternative options. Its capabilities also include precise timestamps and song-to-lyrics transcription, making it an invaluable tool in various applications.
NVIDIA’s speech AI portfolio is extensive, featuring multilingual models such as Parakeet and NeMo Canary, which support 25 languages and boast built-in noise robustness, enabled by the Silero voice activity detection (VAD). These models cater to the needs of both developers and enterprises, available as open-source solutions and enterprise-ready NVIDIA NIM microservices via NVIDIA Riva. This flexibility targets a broad range of use cases from edge devices to extensive enterprise voice solutions.
Key to NVIDIA’s approach is the prioritization of models that are robust in noisy environments, have wide domain and language coverage, are highly personalized to meet specific customer requirements, and are optimized for deployment. The architecture underpinning these models centers around the Fast Conformer, with notable offerings such as Riva Parakeet for streaming applications and Riva Canary for nuanced multitask scenarios. NVIDIA emphasizes the customization capabilities of their models, including fine-tuning of acoustic and language models, as well as word boosting for domain-specific vocabulary needs.
Consistently ranking in the top 5 on Hugging Face’s Open ASR leaderboard, NVIDIA’s models exemplify the blend of high performance and adaptability. They ensure scalable, low-latency inference across diverse platforms, facilitated through NVIDIA NIM containers, thereby addressing the diverse needs of modern speech AI applications. This comprehensive approach promises to end the era of awkward AI transcriptions, delivering precise and high-quality speech recognition solutions across various industries.

Collections

NVIDIA has released advanced speech recognition models including the Parakeet TDT 0.6B v2, which achieves a 6.05% word error rate and leads the Hugging Face ASR leaderboard. The company offers multilingual models supporting 25 languages with noise robustness, available as open-source solutions and enterprise-ready NIM microservices. These models feature Fast Conformer architecture, customization capabilities, and optimized deployment for various applications from edge devices to enterprise solutions.

NVIDIA Advances Speech AI with High-Performance Models