Open speech recognition ceased to be a monoculture at Whisper sooner or later within the final 12 months. March 2026 Cohere releases Transcribethe 2B Apache 2.0 mannequin got here out on high. Hug Face Open ASR Leaderboard The typical phrase error charge is 5.42%. 5 weeks later IBM ships Granite Speech 4.1 2B It was 5.33%. since then ARK-ASR-3B and MOSS-Transcription Preview-2B Nonetheless posting low numbers.
The distinction on the high of that leaderboard is now lower than 1 WER level. This has particular implications for these selecting a mannequin. Rating is not the figuring out variable. Licensing, language help, streaming help, and price per audio hour. This abstract compares all 4 fields.
First, the issue with the leaderboard numbers that everybody cites.
The Open ASR Leaderboard common is just not a single fastened amount, and the fashions at the moment listed facet by facet weren’t all scored in the identical method.
Cohere’s 5.42% is the typical throughout eight English check units, together with TED-LIUM. In consequence, the typical for AMI 8.13, Earnings-22 10.86, GigaSpeech 9.34, LibriSpeech clear 1.25, LibriSpeech different 2.37, SPGISpeech 3.08, TED-LIUM 2.49, VoxPopuli 5.87 is precisely 5.42.
ARK-ASR-3B’s 5.04% is the general common Seven set. TED-LIUM was absent. The MOSS-Transcribe-preview-2B card states this explicitly. TED-LIUM is at the moment not included within the leaderboard run and is subsequently excluded.
TED-LIUM is among the best units within the suite, so eradicating it would improve your common. After we recalculated the scores for every dataset that Cohere publishes in opposition to the identical seven units that ARK stories, Cohere now scores 5.84 as an alternative of 5.42. Doing the identical for Granite Speech 4.1 2B strikes it from 5.33 to five.65. ARK leads by related standards: greater It is not as small because the headline quantity suggests, however the level is that subtracting one printed quantity from one other does not provide you with a significant reply.
Two further notes are listed on the identical web page.
Some scores are brazenly tailored to leaderboards: The MOSS-Transcribe-preview-2B card states that the mannequin was fine-tuned utilizing reinforcement studying on the Open ASR Leaderboard coaching break up. That is public and is larger than most numbers, however it means the rating measures a benchmark fairly than capacity.
Personal monitor information modifications board order:Appen Pending Evaluation Set Contributed Covers Australian, Canadian, Indian, and American accents with scripts and dialog necessities. When these non-public units are turned on, zoom/scribe_v1 strikes from #4 to #1 and the general public leaderboard chief strikes down. Fashions tuned for clear spoken speech degrade disproportionately for spontaneous conversational speech.
Create a shortlist utilizing leaderboards. Do not use it to choose winners.
accuracy degree
Kohia transcription (2B, Apache 2.0, 14 languages) is the mannequin that was truly shipped into manufacturing. It has been downloaded over 620,000 occasions within the final month and has runtime help. transformers,vLLM, mlx-audio For Apple Silicon, Rust ports, and WebGPU builds. This can be a Conformer encoder with a light-weight Transformer decoder skilled from scratch. Cohere additionally carried out human choice assessments, with skilled annotators scoring transcripts for semantic preservation, hallucinations, and named entities. The typical win charge was 61%, 78% in opposition to IBM Granite 4.0 1B Speech, and 64% in opposition to Whisperlarge-v3.
The constraints part of the mannequin card could be very sincere and ought to be learn earlier than committing. There isn’t a computerized language detection, timestamps, or diarization, and the mannequin tries to transcribe silence, so Cohere recommends including a VAD or noise gate in entrance. This repository is gated by the Contact Data Settlement regardless of the Apache 2.0 license.
Granite Speech 4.1 2B In case you want performance, we suggest selecting (2B, Apache 2.0) over a decrease quantity. Six languages for ASR, plus two-way voice translation, key phrase record bias for names and jargon, and true casing together with punctuation and German noun capitalization. I acquired 174,000 hours of coaching. RTFx 231.29. IBM additionally ships two siblings: -plus Add ASR and word-level timestamps for speaker attributes. -nar I’ll clarify under.
Canary-Quen-2.5B (2.5B, CC-BY-4.0, English) combines a FastConformer encoder with a Qwen3-1.7B decoder and runs in two modes: pure transcription, or LLM mode, the place the decoder summarizes and solutions questions concerning the transcript. RTFx 418 with 5.63% WER. Be aware that AMI oversamples the coaching information by roughly 15%, which biases the output towards transcripts that protect verbatim disfluency. This can be a perform of authorized follow and is a nuisance for assembly minutes.
Quen 3-ASR-1.7B (Apache 2.0) covers 52 languages and dialects (30 languages and 22 Chinese language dialects) at 5.76%. It comes with an entire inference toolkit and separate pressured alignment fashions for timestamps in 11 languages. That is an apparent start line relating to publicity to Mandarin or Chinese language regional languages.
throughput tier
The accuracy throughout the highest of the sphere modifications by about 1 WER level. Throughput usually determines your invoice as a result of it might differ by greater than an order of magnitude.
Parakeet TDT 0.6B v3 (0.6B, CC-BY-4.0) delivers the very best throughput of any multilingual open mannequin with RTFx 3332.74 throughout 25 European languages with computerized language ID, as much as 24 minutes on a single cross of the A100 80GB. It prices 6.32% WER, about 14x extra audio per GPU second, and about 1 level greater than Granite 4.1 2B.
Granite Speech 4.1 2B-NAR is a extra fascinating engineering outcome. That is non-autoregressive. Edit the CTC speculation in a single ahead cross utilizing bidirectional LLM to succeed in RTFx ~1820 in a single H100 with batch measurement 128. To get there, abandon Japanese, voice translation, and key phrase bias.
Quen 3-ASR-0.6B Retains all 52 languages and reaches 2000x throughput at 128 concurrency.
streaming layer
Batch WER appears to be the fallacious check for streaming fashions, and leaderboards rating them anyway. Voxtral Realtime is 7.68% and Kyutai STT 2.6B is 6.40%, each under Whisper, however neither quantity offers any helpful details about its meant use.
Voxtral Mini 4B Real Time 2602 (Apache 2.0, 13 languages) is a 3.4B language mannequin and a 970M causal audio encoder skilled from scratch, with sliding window consideration on each halves to successfully obtain limitless streaming. Transcription delay is configurable from 80 ms to 1200 ms in 80 ms increments, with a standalone 2400 ms choice added. Mistral recommends 480ms as a candy spot and stories that that setting matches the main offline open fashions. Runs on a single 16GB GPU, Day 0 of vLLM Real-Time API Support.
Nine body STT (CC-BY-4.0) is available in two codecs. One is the ~1B English/French mannequin with a 0.5 second delay and a built-in semantic speech exercise detector, and the opposite is the two.6B English-only mannequin with a 2.5 second delay. For voice brokers, semantic VAD is extra necessary than transcription delay. Semantic VAD predicts when a speaker will truly end and controls the latency of the flip at which it’s acknowledged. H100 processes 400 simultaneous streams in actual time.
protection layer
All language ASR in Meta (Apache 2.0, Corpus CC-BY) is just not in competitors with WER and shouldn’t be evaluated as such. It natively covers over 1,600 languages and extends to over 5,400 by zero-shot in-context studying, scaled to 7B and constructed on a wav2vec 2.0 encoder pre-trained with roughly 4.3 million hours. The 7B LLM-ASR variant achieves a personality error charge of lower than 10% in 78% of supported languages, together with over 500 languages by no means earlier than supplied on an ASR system. Encoder measurement is 300M to 7B. Meta additionally launched an omnilingual ASR corpus protecting over 350 underserved languages.
Whisper large-v3 The accuracy of (1.55B, MIT, 99 languages) has been surpassed by about 10 open fashions, however stays the right default for a big class of tasks. MIT is the least burdensome license on this space. There isn’t a equal new launch within the runtime ecosystem (whisper.cpp, faster-whisper, WhisperX). In case your necessities are “no language, no {hardware}, no lawyer {qualifications},” you will nonetheless get the reply.
Analysis benefit
Diffusion Gemma ASR Small YC startup Interfaze is probably the most architecturally artistic launch of the yr. We generate transcripts with 8 to 16 steps of parallel diffuse denoising on a canvas of 256 tokens, so the decoding price doesn’t improve with the size of the transcript. Solely about 42 million parameters, representing 0.16% of the weights, had been skilled on the frozen 26B DiffusionGemma and the frozen Whisper miniature encoder. The LibriSpeech check clear achieves a real-time WER of roughly 11-17x at 6.6%. Additionally, since FLEURS English reaches a CER of 15.7% and FLEURS Mandarin reaches 29.6% CER, we deal with the LibriSpeech quantity as an higher restrict. Nobody ought to introduce this. Everybody engaged on ASR ought to learn it.
MOSS-Transcribe-Diarize 0.9B (Apache 2.0, 50+ languages) solves an issue that’s ignored in most summaries. Moderately than chaining ASR to a separate diarization stack, it outputs speaker labels, phrase timestamps, and transcripts in a single era. 128k context, ~90 minutes of audio per cross, RTF ~0.017 (on RTX 4090, with hotword bias).
License division that nobody reads till implementation
That is the part that really blocks transport, and the fields are neatly separated.
Apache 2.0: Cohere Transcribe, Granite Speech 4.1 (all three variants), Qwen3-ASR (each sizes), Voxtral Mini Realtime, Omnilingual ASR, ARK-ASR, MOSS-Transcribe. No attribution obligation, limitless industrial use. Be aware that Cohere’s repositories are gated by contact info agreements, despite the fact that the license itself is Apache 2.0.
Massachusetts Institute of Know-how: Whisper Dai-v3. Probably the most forgiving choice on this space.
CC-BY-4.0:Canary-Qwen-2.5B, Parakeet TDT 0.6B v3, Kyutai STT. Business use out there, however attribution required. For embedded merchandise and white-label APIs, this can be a actual compliance obligation and the most typical cause groups ship fashions that aren’t probably the most correct of these examined.
Meta’s omnilingual ASR splits the next into two: Apache 2.0 for the mannequin, CC-BY for the corpus.
Find out how to truly select
Execute the following order in addition to the leaderboard order.
- license: If attribution is a deterrent, CC-BY-4.0 removes Parakeet, Canary-Qwen, and Kyutai earlier than benchmarking.
- Supported languages: 14 languages in Cohere, 6 languages in Granite, and solely English in Canary are exhausting limits fairly than smooth limits. Moreover, Cohere doesn’t have computerized language detection, so you could know the language upfront.
- streaming or batch: That is an architectural factor. No quantity of tuning can rework an offline encoder/decoder right into a low-latency streaming mannequin.
- Subsequent, measure the WER of your audio: The unfold between the highest 10 fashions on public benchmarks is lower than 1 level. The unfold of accented, noisy, domain-specific audio can be a number of occasions that, and the fashions is not going to be ranked in the identical order.
- Subsequent, calculate the fee per hour of audio by yourself GPU.: RTFx numbers are measured in giant batch sizes on information middle {hardware} and are usually not transferred.
A helpful abstract of 2026 is that nobody mannequin received. That stated, the 2B open weight mannequin of permissive licensing now exceeds the costs closed APIs had been charging 18 months in the past, and the remaining choices are procurement points, not analysis points.
supply of data:
Leaderboard positions are real-time and alter steadily. All figures had been verified in opposition to major sources on July 23, 2026.
Asif Razzaq is the CEO of Marktechpost Media Inc. As a visionary entrepreneur and engineer, Asif is dedicated to harnessing the potential of synthetic intelligence for social good. His newest endeavor is the launch of Marktechpost, a man-made intelligence media platform. It stands out for its thorough protection of machine studying and deep studying information, which is technically sound and simply understood by a large viewers. The platform boasts over 2 million views monthly, demonstrating its reputation amongst viewers.
