Nvidia AI Released Nemotron Speech ASR (huggingface.co)

NVIDIA has launched the Nemotron Speech Streaming ASR model, a significant advancement in speech-to-text technology that offers high-quality English transcription for both low-latency streaming and high-throughput batch processing. This first unified model in the Nemotron Speech family, known as nemotron-speech-streaming-en-0.6b, features a cache-aware architecture designed for efficient continuous processing of audio streams, thereby enhancing the operational efficiency of voice applications like virtual assistants and live captioning. It supports dynamic configurable chunk sizes, allowing developers to balance latency and accuracy based on specific needs without requiring model retraining. The core of the model is based on a FastConformer architecture with a Recurrent Neural Network Transducer (RNNT) decoder, boasting 600 million parameters. It processes single-channel audio sampled at 16,000 Hz and outputs text with integrated punctuation and capitalization, significantly boosting the usability of transcriptions. The adoption of continuous transcription through its innovative caching mechanism reduces computational redundancy and end-to-end delays, making it a game changer for real-time applications. With performance benchmarks showing impressive Word Error Rates (WER) across various datasets, this model positions itself strongly in the competitive field of AI/ML-driven speech recognition solutions.