Open speech recognition is no longer dominated by Whisper in the way it was a year earlier. In March 2026, Cohere released Transcribe, a 2B-parameter Apache 2.0 model that reached the top of the Hugging Face Open ASR Leaderboard with a 5.42% average word error rate, or WER. WER counts substitutions, deletions and insertions against a reference transcript, so it is useful for comparing recognition accuracy but does not capture latency, speaker attribution, timestamp quality or formatting. Five weeks later, IBM released Granite Speech 4.1 2B at 5.33%. Since then, ARK-ASR-3B and MOSS-Transcribe-preview-2B have reported still lower numbers.

The leading models on that leaderboard are now separated by less than one WER point. For teams selecting an automatic speech recognition model, that changes the decision process: leaderboard rank is no longer the only or even the decisive variable. License terms, language support, streaming capability and cost per audio-hour now carry comparable weight. This comparison reviews the open ASR field across those factors.

The leaderboard figure has important limits

The Open ASR Leaderboard average is not a single fixed measure, and models shown side by side on the board have not all been evaluated in exactly the same way.

Cohere’s 5.42% score is an average across eight English test sets, including TED-LIUM. The per-dataset numbers are 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 and VoxPopuli 5.87, which average to exactly 5.42.

ARK-ASR-3B’s 5.04% figure is averaged across seven sets. TED-LIUM is not included. The MOSS-Transcribe-preview-2B model card states this explicitly: TED-LIUM is not currently part of the leaderboard run and is therefore excluded.

Because TED-LIUM is one of the easier datasets in the suite, excluding it raises the average. If Cohere’s published per-dataset scores are recalculated over the same seven datasets reported by ARK, Cohere’s result becomes 5.84 rather than 5.42. Applying the same method to Granite Speech 4.1 2B moves its figure from 5.33 to 5.65. On a like-for-like basis, ARK’s lead is larger than the headline figures suggest, not smaller. The broader point is that subtracting one published leaderboard number from another does not necessarily produce a meaningful comparison.

Two additional caveats are also material.

First, some scores are openly leaderboard-fitted. The MOSS-Transcribe-preview-2B card says the model was fine-tuned with reinforcement learning on the Open ASR Leaderboard training splits. That disclosure is useful, but it also means the score is measuring benchmark performance rather than general capability alone.

Second, private-track data can reorder the board. Appen contributed held-back evaluation sets covering Australian, Canadian, Indian and American accents in both scripted and conversational conditions. When those private sets are enabled, zoom/scribe_v1 moves from No. 4 to No. 1, while the public-leaderboard leader falls by one position. Models tuned for clean read speech can degrade disproportionately on spontaneous conversational audio.

The leaderboard is therefore useful for building a shortlist. It should not be treated as a complete selection mechanism.

Accuracy-focused models

Cohere Transcribe (2B, Apache 2.0, 14 languages) is the model in this group that has already shipped into production. It has been downloaded more than 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. The model uses a Conformer encoder with a lightweight Transformer decoder and was trained from scratch. Cohere also conducted human preference evaluation in which trained annotators scored transcripts for meaning preservation, hallucination and named entities. The model recorded a 61% average win rate, including 78% against IBM Granite 4.0 1B Speech and 64% against Whisper large-v3.

Its limitations are significant and are clearly stated in the model card. Cohere Transcribe does not provide automatic language detection, timestamps or diarization. The model is also eager to transcribe silence, so Cohere recommends placing a VAD or noise gate in front of it. In addition, the repository is gated behind a contact-information agreement despite the Apache 2.0 license.

Granite Speech 4.1 2B (2B, Apache 2.0) is stronger when the requirement is broader capability rather than the lowest leaderboard number. It supports ASR in six languages, bidirectional speech translation, keyword-list biasing for names and jargon, punctuation and truecasing, including German noun capitalization. IBM trained it on 174,000 hours of data and reports RTFx 231.29. RTFx is a throughput measure: higher values mean more audio processed per unit of wall-clock time, usually under a specified hardware and batching setup. IBM also ships two related models: the -plus variant adds speaker-attributed ASR and word-level timestamps, while the -nar variant is discussed in the throughput section.

Canary-Qwen-2.5B (2.5B, CC-BY-4.0, English) combines a FastConformer encoder with a Qwen3-1.7B decoder. It runs in two modes: standard transcription, and an LLM mode in which the decoder summarizes and answers questions about the transcript. It reports 5.63% WER at RTFx 418. AMI was oversampled to about 15% of the training data, which biases the output toward verbatim transcripts that preserve disfluencies. That behavior can be useful for legal work and less desirable for meeting notes.

Qwen3-ASR-1.7B (Apache 2.0) covers 52 languages and dialects, comprising 30 languages plus 22 Chinese dialects, and reports 5.76% WER. It ships with a full inference toolkit and a separate forced-alignment model that provides timestamps in 11 languages. For use cases involving Mandarin or Chinese regional speech, it is a natural starting point.

Throughput-focused models

Accuracy among the leading open ASR models now differs by roughly one WER point. Throughput differs by more than an order of magnitude, which means throughput often determines the operating cost. This matters most for archives, call-center analytics and media pipelines, where the workload is measured in thousands or millions of audio minutes rather than live conversations.

Parakeet TDT 0.6B v3 (0.6B, CC-BY-4.0) is the throughput leader among multilingual open models. It reaches RTFx 3332.74 across 25 European languages, includes automatic language ID and can process up to 24 minutes in a single pass on an A100 80GB. The tradeoff is 6.32% WER, roughly one point higher than Granite 4.1 2B, in exchange for about fourteen times the audio per GPU-second.

Granite Speech 4.1 2B-NAR is the more notable engineering result. It is non-autoregressive: it edits a CTC hypothesis in a single forward pass using a bidirectional LLM. IBM reports RTFx of about 1820 on one H100 at batch size 128. To achieve that throughput, it gives up Japanese, speech translation and keyword biasing.

Qwen3-ASR-0.6B retains all 52 languages and reaches 2000× throughput at concurrency 128.

Streaming-focused models

Batch WER is an awkward measure for streaming models, although the leaderboard scores them anyway. Voxtral Realtime is listed at 7.68% and Kyutai STT 2.6B at 6.40%. Both are below Whisper on that metric, but the numbers say little about their intended use.

Voxtral Mini 4B Realtime 2602 (Apache 2.0, 13 languages) combines a 3.4B language model with a 970M causal audio encoder trained from scratch. Both parts use sliding-window attention, allowing effectively unbounded streaming. Transcription delay can be configured in 80ms steps from 80ms to 1200ms, with an additional standalone 2400ms option. Mistral recommends 480ms as the optimal setting and says that at that setting the model 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) is available in two forms: a roughly 1B-parameter 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 can matter more than the transcription delay because it predicts when the speaker has actually finished, which is what governs perceived turn-taking latency. An H100 can serve 400 concurrent streams in real time.

Coverage-focused models

Meta’s Omnilingual ASR (Apache 2.0 for the model, CC-BY for the corpus) is not competing primarily on WER and should not be evaluated as though it were. It covers more than 1,600 languages natively and extends to more than 5,400 through zero-shot in-context learning. The system is 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 more than 500 languages not previously served by any ASR system. Character error rate is commonly used where word boundaries are not consistent or where low-resource language evaluation is more practical at the character level. Encoder sizes range from 300M to 7B. Meta also released the Omnilingual ASR Corpus, covering more than 350 underserved languages.

Whisper large-v3 (1.55B, MIT, 99 languages) has been surpassed on accuracy by roughly ten open models, but it remains a strong default for many projects. MIT is the least encumbered license in the field. Its runtime ecosystem — whisper.cpp, faster-whisper and WhisperX — remains unmatched among newer releases. For requirements that amount to broad language support, flexible hardware support and minimal licensing friction, Whisper large-v3 is still a practical answer.

Research-focused models

diffusion-gemma-asr-small from YC startup Interfaze is one of the year’s most architecturally unusual releases. It generates transcripts through parallel diffusion denoising over a 256-token canvas in 8 to 16 steps, so decoding cost does not grow with transcript length. Only about 42M parameters were trained, equal to 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 records 15.7% on FLEURS English and 29.6% CER on FLEURS Mandarin, so the LibriSpeech number should be treated as the ceiling. The model is not presented as deployment-ready, but it is important research for ASR practitioners.

MOSS-Transcribe-Diarize 0.9B (Apache 2.0, 50+ languages) addresses a problem many ASR comparisons leave aside: it outputs speaker labels, word timestamps and the transcript in one generation rather than chaining ASR to a separate diarization system. It supports 128k context, roughly 90 minutes of audio in one pass, RTF of about 0.017 on an RTX 4090 and hotword biasing.

License differences can decide deployment

Licensing is often the issue that determines whether a model can actually ship. The field divides clearly.

Apache 2.0 covers Cohere Transcribe, Granite Speech 4.1 in all three variants, Qwen3-ASR in both sizes, Voxtral Mini Realtime, Omnilingual ASR, ARK-ASR and MOSS-Transcribe. These models have no attribution obligation, and commercial use is unrestricted. Cohere’s repository, however, is gated behind a contact-information agreement even though the license itself is Apache 2.0.

MIT applies to Whisper large-v3, making it the most permissive option in the field.

CC-BY-4.0 applies to Canary-Qwen-2.5B, Parakeet TDT 0.6B v3 and Kyutai STT. These models can be used commercially, but attribution is required. For an embedded product or white-labeled API, that attribution requirement is a meaningful compliance obligation and is often the reason teams deploy a model other than the most accurate one they tested.

Meta’s Omnilingual ASR is split: the models are Apache 2.0, while the corpus is CC-BY.

A practical selection order

A model-selection process should not simply follow the leaderboard order.

Start with the license. If attribution is not acceptable, CC-BY-4.0 removes Parakeet, Canary-Qwen and Kyutai before benchmarking begins.

Next, check language coverage. Cohere’s 14 languages, Granite’s 6 and Canary’s English-only coverage are fixed constraints. Cohere also lacks automatic language detection, so the language must be known in advance.

Then decide whether the application requires streaming or batch operation. That is an architectural distinction: tuning does not turn an offline encoder-decoder model into a low-latency streaming system.

After that, measure WER on the actual audio to be processed. The public benchmark spread among the top ten models is under one point, but the spread on accented, noisy or domain-specific audio will be several times larger and will not necessarily rank models in the same order. If downstream systems depend on diarization, timestamps, punctuation or hotword handling, those outputs need to be tested directly rather than inferred from WER.

Finally, calculate cost per audio-hour on the hardware that will actually be used. RTFx figures are generally measured at large batch sizes on datacenter hardware and do not transfer directly to every deployment setting.

The central development in 2026 is not that a single model has won the ASR market. It is that a 2B open-weight model with a permissive license can now exceed the performance level for which closed APIs were charging eighteen months earlier, while the remaining choice increasingly resembles a procurement decision rather than a research decision.

Sources cited in the original article include Cohere Transcribe blog and model card; IBM Granite Speech 4.1 2B and 4.1 2B-NAR; ARK-ASR-3B; MOSS-Transcribe-preview-2B and MOSS-Transcribe-Diarize; NVIDIA Canary-Qwen-2.5B and Parakeet TDT 0.6B v3; Qwen3-ASR; Voxtral Mini 4B Realtime 2602; Kyutai STT; Meta Omnilingual ASR and paper; Whisper large-v3; diffusion-gemma-asr; the Open ASR Leaderboard paper; the leaderboard dataset; and Appen’s private-track announcement.

Leaderboard positions are live and change frequently. All figures were verified against primary sources on 23 July 2026.