Open speech recognition stopped being a Whisper monoculture some time in the last twelve months. In March 2026 Cohere released Transcribe, a 2B Apache 2.0 model that took the top of the Hugging Face Open ASR Leaderboard at 5.42% average word error rate. Five weeks later IBM shipped Granite Speech 4.1 2B at 5.33%. Since then ARK-ASR-3B and MOSS-Transcribe-preview-2B have posted lower numbers still.
The top of that leaderboard is now separated by less than one WER point. That has a specific consequence for anyone choosing a model: rank is no longer the deciding variable. License, language coverage, streaming support, and cost per audio-hour are. This roundup compares the field on all four.
First, a problem with the leaderboard number everyone quotes
The Open ASR Leaderboard average is not a single fixed quantity, and the models currently listed side by side were not all scored the same way.
Cohere’s 5.42% is an average across eight English test sets, including TED-LIUM. That checks out: AMI 8.13, Earnings-22 10.86, GigaSpeech 9.34, LibriSpeech clean 1.25, LibriSpeech other 2.37, SPGISpeech 3.08, TED-LIUM 2.49, VoxPopuli 5.87 averages to exactly 5.42.
ARK-ASR-3B’s 5.04% is an average across seven sets. TED-LIUM is absent. The MOSS-Transcribe-preview-2B card states this explicitly: TED-LIUM is not currently part of the leaderboard run and is therefore excluded.
TED-LIUM is one of the easier sets in the suite, so dropping it raises the average. Recompute Cohere’s published per-dataset scores over the same seven sets ARK reports and Cohere lands at 5.84, not 5.42. Do the same to Granite Speech 4.1 2B and it moves from 5.33 to 5.65. On a like-for-like basis ARK’s lead is larger than the headline numbers imply, not smaller — but the point is that you cannot subtract one published figure from another and get a meaningful answer.
Two further caveats belong on the same page:
Some scores are openly leaderboard-fitted: The MOSS-Transcribe-preview-2B card states the model was fine-tuned with reinforcement learning on the Open ASR Leaderboard training splits. That is disclosed, which is more than most, but it means the score measures the benchmark rather than the capability.
Private-track data reorders the board: Appen contributed held-back evaluation sets covering Australian, Canadian, Indian, and American accents in scripted and conversational conditions. When those private sets are toggled on, zoom/scribe_v1 moves from #4 to #1 and the public-leaderboard leader drops a position. Models tuned for clean read speech degrade disproportionately on spontaneous conversational audio.
Use the leaderboard to build a shortlist. Do not use it to pick a winner.
The accuracy tier
Cohere Transcribe (2B, Apache 2.0, 14 languages) is the model that actually shipped into production. It has been downloaded over 620,000 times in the past month and has runtime support across transformers, vLLM, mlx-audio for Apple Silicon, a Rust port, and a WebGPU build. It is a Conformer encoder with a lightweight Transformer decoder, trained from scratch. Cohere also ran human preference evaluation, where trained annotators scored transcripts for meaning preservation, hallucination, and named entities: a 61% average win rate, 78% against IBM Granite 4.0 1B Speech and 64% against Whisper large-v3.
The limitations section of its model card is unusually honest and should be read before committing. There is no automatic language detection, no timestamps, and no diarization, and the model is eager to transcribe silence, so Cohere recommends prepending a VAD or noise gate. The repo is also gated behind a contact-information agreement despite the Apache 2.0 license.
Granite Speech 4.1 2B (2B, Apache 2.0) is the better pick if you need capability rather than a lower number. Six languages for ASR plus bidirectional speech translation, keyword-list biasing for names and jargon, punctuation and truecasing including German noun capitalization. Trained on 174,000 hours. RTFx 231.29. IBM also ships two siblings: -plus adds speaker-attributed ASR and word-level timestamps, and -nar is discussed below.
Canary-Qwen-2.5B (2.5B, CC-BY-4.0, English) pairs a FastConformer encoder with a Qwen3-1.7B decoder and runs in two modes — pure transcription, or LLM mode where the decoder summarizes and answers questions about the transcript. 5.63% WER at RTFx 418. Note that AMI was oversampled to roughly 15% of training data, which biases output toward verbatim disfluency-preserving transcripts. That is a feature for legal work and a nuisance for meeting notes.
Qwen3-ASR-1.7B (Apache 2.0) covers 52 languages and dialects — 30 languages plus 22 Chinese dialects — at 5.76%. It ships with a full inference toolkit and a separate forced-alignment model for timestamps in 11 languages. For anything touching Mandarin or Chinese regional speech this is the obvious starting point.
The throughput tier
Accuracy across the top of the field now varies by about one WER point. Throughput varies by more than an order of magnitude, which means throughput usually decides the invoice.
Parakeet TDT 0.6B v3 (0.6B, CC-BY-4.0) is the throughput leader among multilingual open models at RTFx 3332.74 across 25 European languages with automatic language ID, handling up to 24 minutes in a single pass on an A100 80GB. It costs 6.32% WER — roughly one point more than Granite 4.1 2B for roughly fourteen times the audio per GPU-second.
Granite Speech 4.1 2B-NAR is the more interesting engineering result. It is non-autoregressive: it edits a CTC hypothesis in a single forward pass using a bidirectional LLM, reaching RTFx ~1820 on one H100 at batch size 128. It gives up Japanese, speech translation, and keyword biasing to get there.
Qwen3-ASR-0.6B keeps all 52 languages and reaches 2000× throughput at concurrency 128.
The streaming tier
A batch WER seems to be the wrong test for a streaming model, and the leaderboard scores them anyway. Voxtral Realtime sits at 7.68% and Kyutai STT 2.6B at 6.40% — both below Whisper — and neither number tells you anything useful about their intended use.
Voxtral Mini 4B Realtime 2602 (Apache 2.0, 13 languages) is a 3.4B language model along with a 970M causal audio encoder trained from scratch, with sliding-window attention on both halves for effectively unbounded streaming. Transcription delay is configurable in 80ms steps from 80ms to 1200ms, plus a standalone 2400ms option; Mistral recommends 480ms as the sweet spot and reports that at that setting it matches leading offline open models. It runs on a single 16GB GPU and had day-0 vLLM Realtime API support.
Kyutai STT (CC-BY-4.0) comes in two shapes: a ~1B English/French model with a 0.5s delay and a built-in semantic voice activity detector, and a 2.6B English-only model with a 2.5s delay. For voice agents, the semantic VAD matters more than the transcription delay — it predicts when the speaker has actually finished, which is what governs perceived turn-taking latency. An H100 serves 400 concurrent streams in real time.
The coverage tier
Meta’s Omnilingual ASR (Apache 2.0, corpus CC-BY) is not competing on WER and should not be evaluated as if it were. It covers 1,600+ languages natively and extends to 5,400+ through zero-shot in-context learning, built on a wav2vec 2.0 encoder scaled to 7B and pre-trained on about 4.3M hours. The 7B LLM-ASR variant achieves character error rate below 10% on 78% of supported languages, including 500+ never previously served by any ASR system. Encoder sizes run from 300M to 7B. Meta also released the Omnilingual ASR Corpus covering 350+ underserved languages.
Whisper large-v3 (1.55B, MIT, 99 languages) has been overtaken on accuracy by roughly ten open models, and it remains the correct default for a large class of projects. MIT is the least encumbered license in the field. The runtime ecosystem — whisper.cpp, faster-whisper, WhisperX — has no equivalent among the newer releases. If your requirement is “some language, some hardware, no license lawyer,” it is still the answer.
The research edge
diffusion-gemma-asr-small from YC startup Interfaze is the most architecturally creative release of the year. It produces transcripts by parallel diffusion denoising over a 256-token canvas in 8 to 16 steps, so decoding cost does not grow with transcript length. Only ~42M parameters were trained — 0.16% of the weights — on top of a frozen 26B DiffusionGemma and a frozen whisper-small encoder. It reaches 6.6% WER on LibriSpeech test-clean at roughly 11 to 17× realtime. It also reaches 15.7% on FLEURS English and 29.6% CER on FLEURS Mandarin, so treat the LibriSpeech figure as the ceiling. Nobody should deploy this. Everybody working on ASR should read it.
MOSS-Transcribe-Diarize 0.9B (Apache 2.0, 50+ languages) solves the problem most roundups ignore: it emits speaker labels, word timestamps, and transcript in one generation instead of chaining ASR to a separate diarization stack. 128k context, roughly 90 minutes of audio in one pass, RTF ~0.017 on an RTX 4090, with hotword biasing.
The license split nobody reads until deployment
This is the section that actually blocks shipping, and the field divides cleanly:
Apache 2.0: Cohere Transcribe, Granite Speech 4.1 (all three variants), Qwen3-ASR (both sizes), Voxtral Mini Realtime, Omnilingual ASR, ARK-ASR, MOSS-Transcribe. No attribution obligation, commercial use unrestricted. Note that Cohere’s repo is gated behind a contact-information agreement even though the license itself is Apache 2.0.
MIT: Whisper large-v3. The most permissive option in the field.
CC-BY-4.0: Canary-Qwen-2.5B, Parakeet TDT 0.6B v3, Kyutai STT. Commercially usable, but attribution is required. For an embedded product or a white-labeled API, that is a real compliance obligation, and it is the single most common reason teams end up shipping a model that is not the most accurate one they tested.
Meta’s Omnilingual ASR splits the two: Apache 2.0 for models, CC-BY for the corpus.
How to actually choose
Run this order, not just the leaderboard order:
- License: If attribution is a blocker, CC-BY-4.0 removes Parakeet, Canary-Qwen, and Kyutai before you benchmark anything.
- Language coverage: Cohere’s 14 languages, Granite’s 6, and Canary’s English-only are hard limits, not soft ones. Cohere additionally has no language auto-detection, so you must know the language in advance.
- Streaming or batch: This is architectural. No amount of tuning turns an offline encoder-decoder into a low-latency streaming model.
- Then measure WER on your own audio: The spread between the top ten models on the public benchmark is under one point. The spread on your accented, noisy, domain-specific audio will be several times that, and it will not rank the models in the same order.
- Then compute cost per audio-hour on your own GPUs: RTFx figures are measured at large batch sizes on datacenter hardware and do not transfer.
