Fact · Speech recognition
5.63percent average WER
On the same leaderboard track, NVIDIA Canary Qwen 2.5B, a speech encoder with a language-model decoder, had an average word error rate of 5.63 percent and an RTFx of 418.
Checked against its source on October 16, 2026
In the source's own words
“NVIDIA Canary Qwen 2.5B Yes 5.63 418 FastConformer”
Open ASR Leaderboard: Towards Reproducible and Transparent Multilingual and Long-Form Speech Recognition Evaluation, Hugging Face and collaborators, arXiv, 2025, Table 3.
Open the original- What it measures
- One of the most accurate open models listed; it is slower than the transducer models.
- What it does not show
- The leaderboard is updated often and this row is a snapshot of the paper's table.
Source
Used to calculate
Cited on
- Speech Recognition: What It Is and How Well It WorksTopic
- What Is Whisper AI? OpenAI's Speech Model ExplainedGuide
- Word Error Rate (WER): How to Calculate and Read ItGuide
- Speech to Text and LLMs: What Language Models ChangeGuide
- Speech Recognition Models: How Neural ASR WorksGuide
- Open-Source Speech Recognition: The Best ModelsGuide
- Speech Recognition Benchmarks and LeaderboardsGuide
- Offline Speech to Text on Windows: What WorksGuide
- Can large language models do speech recognition?Question
- What are the best alternatives to Whisper?Question