Best open ASR models in 2026 — WER, languages, latency, licence (MarkTechPost)
A secondary roundup of 15 open ASR models, published 2026-07-23. T3: an editorial comparison that aggregates other people’s numbers, from the same outlet as tts-models-2026-benchmark. Its value here is the selection argument, not the figures.
The argument
“Rank is no longer the deciding variable.” The article’s case is that the spread between the top ten models on the public [Open ASR] leaderboard is under one point, and that the ranking is unstable anyway — it notes that toggling on Appen’s private-track data reorders the board. So it treats WER as a filter you pass rather than a contest you win, and pushes the decision onto licence, language coverage, streaming-vs-batch, and cost per audio-hour. Its closing line: “the remaining decision is a procurement question rather than a research one.”
Its five-step selection order: filter by licence first (drop CC-BY-4.0 if attribution is blocking), check language coverage (a hard architectural constraint, not a tunable), decide streaming or batch (also architectural), benchmark WER on your own audio because leaderboard rankings don’t transfer, then compute real cost per audio-hour since RTFx figures assume datacenter hardware.
The field, as it lists it
| Model | WER | Params | Langs | Speed | Licence |
|---|---|---|---|---|---|
| ARK-ASR-3B | 5.04% (7 sets, TED-LIUM excluded) | 3B | — | — | Apache-2.0 |
| Granite Speech 4.1 2B | 5.33% (8 sets) | 2B | 6 | RTFx 231 | Apache-2.0 |
| Cohere Transcribe | 5.42% (8 EN sets) | 2B | 14 | RTFx 231 | Apache-2.0 |
| Canary-Qwen 2.5B | 5.63% | 2.5B | 1 (EN) | RTFx 418 | CC-BY-4.0 |
| Qwen3-ASR-1.7B | 5.76% | 1.7B | 52 + dialects | — | Apache-2.0 |
| Parakeet TDT 0.6B v3 | 6.32% | 0.6B | 25 (European) | RTFx 3333 | CC-BY-4.0 |
| Kyutai STT 1B | 6.40% (batch) | 1B / 2.6B | 2 | 0.5 s / 2.5 s delay | CC-BY-4.0 |
| Voxtral Mini 4B Realtime | 7.68% (batch) | 3.4B LM + 970M enc | 13 | 80–1200 ms configurable | Apache-2.0 |
| diffusion-gemma-asr-small | 6.6% (LibriSpeech clean) | 42M trained on a 26B frozen base | EN | 11–17× realtime | not stated |
| MOSS-Transcribe-preview-2B | below 5.33% (7 sets) | 2B | — | — | Apache-2.0 |
| MOSS-Transcribe-Diarize 0.9B | — | 0.9B | 50+ | RTF ~0.017 (RTX 4090) | Apache-2.0 |
| Qwen3-ASR-0.6B | — | 0.6B | 52 | ~2000× throughput | Apache-2.0 |
| Granite Speech 4.1 2B-NAR | — | 2B | 4 | RTFx ~1820 | Apache-2.0 |
| Meta Omnilingual ASR | ~10% CER on 78% of languages | 300M–7B | 1600+, 5400+ zero-shot | — | Apache-2.0 (CC-BY corpus) |
| Whisper large-v3 | baseline | 1.55B | 99 | — | MIT |
The WER column is not comparable across rows and the article says so: 8 sets for some, 7 with TED-LIUM dropped for others, LibriSpeech-only for one. That inconsistency is the piece’s own evidence for distrusting the ranking.
What this changes here
Canary-Qwen has been displaced. This spoke has recorded it since 2026-06 as the model that tops the Open ASR Leaderboard at 5.63%. Four models here are listed below it — ARK-ASR-3B 5.04, Granite Speech 4.1 5.33, Cohere Transcribe 5.42, and MOSS-Transcribe-preview under 5.33 — and it carries the more restrictive licence (CC-BY-4.0) against their Apache-2.0. Six weeks, four models past it, under 0.6 WER points separating them. That is the leaderboard-churn argument demonstrated rather than asserted.
Three tiers the WER framing hid. The article separates accuracy from throughput (Parakeet TDT at RTFx 3333, Qwen3-ASR-0.6B at ~2000×) and from streaming latency (Voxtral Mini configurable 80–1200 ms, Kyutai STT at a 0.5 s delay) — and streaming-vs-batch is presented as an architecture you choose up front, not a setting. Latency had been a commercial-API property on stt-apis-comparison; here it is an open-weight property with published numbers.
Language coverage has an extreme now. Meta’s Omnilingual ASR claims 1600+ languages and 5400+ zero-shot, against Whisper‘s 99 — an order of magnitude past what this wiki had recorded as the multilingual ceiling, at ~10% CER on 78% of them. Coverage and accuracy are being traded against each other explicitly.
Diarization and non-autoregressive decoding appear as product axes, via MOSS-Transcribe-Diarize 0.9B and Granite’s 2B-NAR variant. The wiki has no page for either mechanism yet.
Licence is the first filter, not a footnote. Apache-2.0 dominates this list; the CC-BY-4.0 models (Canary-Qwen, Parakeet, Kyutai) are flagged as the ones attribution requirements can disqualify. The open-weight wedge in synthesis has been tracked mostly as free-vs-paid; this source splits the open side itself by how usable the licence is.
Numbers are the article’s, mostly restated from vendor cards and the leaderboard. Dated snapshot.
Related
speech-to-text · open-source-stt-models · canary-qwen · whisper · vibevoice-asr-bitnet · stt-apis-comparison · open-weight-tts