NVIDIA open-sources massive multilingual speech dataset and models

NVIDIA just teamed up with Carnegie Mellon University and Fondazione Bruno Kessler to release Granary, a massive open-source dataset, alongside two new multilingual speech models: Canary-1b-v2 and Parakeet-tdt-0.6b-v3.

The entire collection is already available on Hugging Face, with the accompanying data pipeline hosted on GitHub. If you are building multilingual voice bots, customer support agents, or real-time transcription tools, this is an off-the-shelf foundation ready for production.

Granary packs around a million hours of audio, including nearly 650,000 hours for automatic speech recognition (ASR) and over 350,000 hours for speech translation (AST). It covers 25 European languages, tackling low-resource languages like Croatian, Estonian, and Maltese, covering nearly all official EU languages alongside Russian and Ukrainian.

How Granary was built without human labeling

The clever part of Granary is how it sidesteps manual tagging. The researchers built an automated processing pipeline powered by the NVIDIA NeMo Speech Data Processor toolkit to turn raw, unlabeled public audio into structured, high-grade training material.

The team ran NeMo Curator to filter out synthetic samples so that only clean audio made the cut. The toolkit then automatically aligned audio files with text and converted everything into ready-to-train formats, skipping expensive human annotation entirely.

In their paper heading to Interspeech 2025, the team showed that hitting a target accuracy level for speech recognition and translation took only about half as much data with Granary compared to other standard datasets.

Canary for precision, Parakeet for raw speed

The two companion models tackle two distinct production needs, and both return outputs complete with punctuation, capitalization, and word-level timestamps.

Canary-1b-v2 has 1 billion parameters and ships under a permissive license. It expands language support from four languages to 25, handling transcription across Europe and two-way translation between English and every supported language.

Canary currently tops Hugging Face's open leaderboard for multilingual speech recognition accuracy. NVIDIA claims it delivers transcription and translation quality on par with models three times its size while running inference up to 10x faster.

Parakeet-tdt-0.6b-v3 trims the footprint down to 600 million parameters, built strictly for high throughput and low latency. It boasts the highest throughput among multilingual models on the Hugging Face leaderboard, capable of transcribing a 24-minute audio clip in a single inference pass while automatically detecting the spoken language without prompt engineering.

Why it matters

If you build voice applications across European markets, this release is worth an immediate test. Small-market languages usually suffer from poor open-source accuracy, forcing startups onto expensive proprietary speech APIs. With lightweight 600M and 1B architectures, these models make high-accuracy transcription affordable and fast enough for edge or real-time server deployments.

Beyond weights and checkpoints, open-sourcing the NeMo cleaning pipeline is the real force multiplier. Any developer or research team can now replicate this exact workflow to process low-resource dialects across Asia, Africa, or the Americas, accelerating multilingual speech AI across the board.