nyra labs

nyra labs: nothing is noise

nyra labs, the speech intelligence lab of nyra health based in Vienna, conducts research and builds open models designed to preserve and measure human speech. Their flagship release, CrisperWhisper 2.0, is a speech-to-text model that transcribes speech exactly as it happens, capturing every word, pause, filler, repetition, false start, and vocal sound.

Key Features

• Verbatim Transcription: Captures spontaneous speech details including fillers, repetitions, and cut-offs across 10+ languages.

• Word-Level Timing: Provides highly accurate word boundaries to within 30 ms using guided cross-attention, eliminating the need for a separate aligner.

• Verbatimize Task: Allows users to convert existing spontaneous-speech corpora into faithful verbatim data while preserving trusted transcripts.

• Open and Production-Ready: Features open weights and code available on Hugging Face, optimized inference, seamless longform capabilities, and strong hallucination resistance.

• Industry-Leading Accuracy: Outperforms competitors in disfluency detection for both English and German.

Getting Started

Website: https://nyra-labs.com

nyra labs provides open, verifiable speech recognition models that prioritize the exact structure and timing of spoken language, ensuring that real human speech can be fully understood without losing critical disfluencies.