Canary-Qwen 2.5B
NVIDIA’s open-weight speech-to-text model that tops the Open ASR Leaderboard — WER 5.63% (English). 2.5B params, CC-BY-4.0, English-only, RTFx 418× open-source-stt-models.
The architecture that defines the 2026 trend
A SALM (Speech-Augmented Language Model): a FastConformer speech encoder feeding an unmodified Qwen3-1.7B LLM decoder. So ASR is reframed as language modeling over speech features — the recognition-side instance of the “speech on an LLM backbone” convergence the wiki tracks across text-to-speech (Llama-based TTS, neural-audio-codec) and music (suno). Siblings on the same idea: IBM Granite-Speech, Alibaba Qwen3-ASR.
Significance
- Beats whisper (7.4%) and the commercial batch leaders on English WER while staying open-weight.
- Direct cross-wiki bridge: its decoder is literally a Qwen3 LLM — the gemini/open-weight-Qwen lineage in llm-providers-wiki, now doing transcription.
- Tradeoff: English-only — Whisper still wins multilingual.
Displaced from the top (2026-07-24)
The “tops the Open ASR Leaderboard” claim above dates to 2026-06 and no longer holds. MarkTechPost’s July roundup lists four open models below its 5.63% — ARK-ASR-3B at 5.04, Granite Speech 4.1 2B at 5.33, Cohere Transcribe at 5.42, and MOSS-Transcribe-preview-2B under 5.33 — and all four are Apache-2.0 against Canary-Qwen’s CC-BY-4.0, so it now loses on licence as well as rank. Six weeks, four models past it, and the whole group inside 0.6 WER points. The SALM architecture story below is unaffected and still the point: three of the models that overtook it are SALM-style too. Both figures kept per the record-don’t-overwrite rule; the 5.63% is not disputed, its position is.
Related
speech-to-text · speech-audio-ai · whisper · open-source-stt-models · open-asr-models-2026-comparison · neural-audio-codec