Model Details

Domain:

Speech

Task:

Speech recognition ASR

Translation

Speech-to-text

Model Access:

Open weights (unrestricted)

AI Tools Usage

This model is commonly used behind the scenes in AI tools.

Introduction

This report introduces Canary-1B-v2, a fast, robust multilingual model for Automatic Speech Recognition (ASR) and Speech-to-Text Translation (AST). Built with a FastConformer encoder and Transformer decoder, it supports 25 languages primarily European. The model was trained on 1.7M hours of total data samples, including Granary and NeMo ASR Set 3.0, with non-speech audio added to reduce hallucinations for ASR and AST. We describe its two-stage pre-training and fine-tuning process with dynamic data balancing, as well as experiments with an nGPT encoder. Results show nGPT scales well with massive data, while FastConformer excels after fine-tuning. For timestamps, Canary-1B-v2 uses the NeMo Forced Aligner (NFA) with an auxiliary CTC model, providing reliable segment-level timestamps for ASR and AST. Evaluations show Canary-1B-v2 outperforms Whisper-large-v3 on English ASR while being 10x faster, and delivers competitive multilingual ASR and AST performance against larger models like Seamless-M4T-v2-large and LLM-based systems. We also release Parakeet-TDT-0.6B-v3, a successor to v2, offering multilingual ASR across the same 25 languages with just 600M parameters.

Training

Training Code Accessibilitycc-by-4.0 https://huggingface.co/nvidia/canary-1b-v2

HardwareNVIDIA A100

Hardware Quantity64

Size Notes: Total training hours: 1.7M ASR: 660,000 hrs X→En: 360,000 hrs En→X: 690,000 hrs Non-speech: 36,000 hrs "In the first stage, the Canary model was trained on X→En (360,000 hours), English ASR (285,000 hours), and non-speech data (1,200 hours) for 150,000 steps" "After completing the first stage, training continued from the resulting checkpoint on the full blend of data (ASR, X→En, and En→X), comprising roughly 1.7 million hours, for an additional 100,000 steps."