For years "open speech recognition" meant Whisper. That is no longer true. By November 2026, more than twenty open-weight models beat Whisper large-v3 on the standard English benchmark, several of them at a fraction of its size, and OpenAI itself has deprecated Whisper in its paid API in favour of a closed successor. This article surveys the open models that matter, what makes each different, and which are practical to run on a laptop or phone.
All accuracy figures are word error rates (WER) from the Hugging Face Open ASR Leaderboard's default English average as of 9 November 2026 (leaderboard version 1.15.1), unless stated otherwise. Lower is better.
How to read the leaderboard in 2026
The Open ASR Leaderboard's English average now covers nine test sets. Seven are public: meetings (AMI), earnings calls, podcasts and YouTube (GigaSpeech), financial talks (SPGISpeech), parliamentary speech (VoxPopuli), Indian-accented conversations (Voice Arena Monsoon) and, since 9 October 2026, noisy and reverberant speech (URGENT24). The other two are private scripted and conversational sets from data providers, which model makers cannot train on. On 2 November 2026 the two LibriSpeech audiobook sets, the easiest in the list, were removed from the average.
Each of these changes made the test harder, so every score went up. A number from a 2025 model card, or from an earlier edition of this article, is not comparable with the current leaderboard. Qwen3-ASR 1.7B averaged 4.31% on the earlier eight public sets, 5.40% on the October average with LibriSpeech, and 6.15% now. Same model, different scoring.
The leaderboard also reports RTFx, how many seconds of audio a model transcribes per second on a data-centre GPU. It is a relative speed measure, not what you will see on a laptop.
The top open models for English
A selection of the leading open-weight models; a few less widely used entries (Hojo-ASR, Higgs Audio, ARK-ASR) also sit in this range.
| Model | Maker | Size | WER | Speed (RTFx) | Licence |
|---|---|---|---|---|---|
| Nemotron 3.5 Transcribe | NVIDIA | 34B (4.3B active) | 6.13 | 473 | OpenMDW 1.1 |
| Qwen3-ASR 1.7B | Alibaba Qwen | 1.7B | 6.15 | 836 | Apache 2.0 |
| Granite Speech 5.0 2B | IBM | 2B | 6.68 | 548 | Apache 2.0 |
| Cohere Transcribe 03-2026 | Cohere | 2B | 6.72 | 924 | Apache 2.0 |
| TheWhisper large-v3-turbo | TheStage AI | 0.8B | 6.74 | 1,184 | CC-BY-4.0 |
| MOSS-Transcribe-Diarize | OpenMOSS (Fudan) | 0.9B | 6.91 | 387 | Apache 2.0 |
| Canary-Qwen 2.5B | NVIDIA | 2.5B | 6.92 | 882 | CC-BY-4.0 |
| Parakeet TDT 0.6B v2 | NVIDIA | 0.6B | 6.93 | 6,148 | CC-BY-4.0 |
| Granite Speech 4.1 2B | IBM | 2B | 7.03 | 547 | Apache 2.0 |
| Parakeet TDT 0.6B v3 | NVIDIA | 0.6B | 7.04 | 6,322 | CC-BY-4.0 |
| Qwen3-ASR 0.6B | Alibaba Qwen | 0.6B | 7.07 | 750 | Apache 2.0 |
| Granite Speech 5.0 TurboCTC | IBM | 0.47B | 7.41 | 21,435 | Apache 2.0 |
| Voxtral Small | Mistral | 24B | 7.44 | 101 | Apache 2.0 |
| Phonon-2 | Fermion Research | 0.6B | 7.86 | 1,977 | CC-BY-4.0 |
| Whisper large-v3 | OpenAI | 1.55B | 8.24 | 471 | MIT |
| Whisper large-v3-turbo | OpenAI | 0.8B | 8.75 | 809 | MIT |
| VibeVoice-ASR | Microsoft | ~9B | 9.54 | 218 | MIT |
The top eight overall are all closed models. Zoom's scribe_v2.1_pro, added on 30 October, leads at 4.87%, followed by Otter.ai's speech-v3.1 at 5.21% and Azure's September speech model at 5.55%; ElevenLabs, Modulate and Reson8 are just under 6%. The best open model is about 1.3 points behind the leader. OpenAI's closed gpt-transcribe scores 6.80%, behind Nemotron 3.5 Transcribe, Qwen3-ASR 1.7B, Granite Speech 5.0 2B and Cohere Transcribe.
The spread among the top open models is small, about 0.8 points across the first ten. In practice, a difference of 0.3 points on this benchmark is less important than how a model handles your audio, how fast it runs on your hardware, and whether it invents text over silence.
The models worth knowing
NVIDIA Nemotron 3.5 Transcribe (October 2026)
Released on 28 October 2026 and, at 6.13, the most accurate open-weight model on the English leaderboard by a margin too small to matter (Qwen3-ASR 1.7B is at 6.15). What sets it apart is the rest of the job. It transcribes 40 languages, takes recordings of up to an hour in one pass, and labels up to eight speakers with word timestamps. It does that by pairing a 600M speech encoder with NVIDIA's Nemotron 3.5 Lightning language model and feeding in speaker activity from a frozen Nemotron 3 Diarization encoder. Among open models it leads the leaderboard's multilingual track (4.69% across German, French, Italian, Spanish, Portuguese, Dutch and Hindi) and is second on long recordings (10.52%).
The cost is size. The checkpoint holds 34B parameters, of which about 4.3B are active for any one token, and NVIDIA's quick start targets a data-centre GPU with 80 GB of memory. It works offline on finished recordings only, with no streaming mode. The licence is OpenMDW 1.1, which allows commercial use.
Qwen3-ASR (Alibaba, January 2026)
Second among open models on English by 0.02 points, at a twentieth of the size, and one of the broadest: 52 languages and dialects. It is built on the audio encoder of Qwen3-Omni with a language-model decoder, and ships with a separate forced aligner for word timestamps in 11 languages. The 0.6B version answers in about 92 ms and scores 7.07, still ahead of Whisper large-v3.
IBM Granite Speech 5.0 2B (October 2026)
Released on 20 October 2026 and now the third most accurate open model on the leaderboard, at 6.68. It pairs a 16-block conformer encoder with a language model built from Granite 4.0 1B. It transcribes and translates English, French, German, Spanish, Portuguese and Japanese, and can favour up to 2,000 keywords you give it, such as names or jargon. IBM recommends feeding it segments of up to about two minutes and splitting longer recordings. Its model card documents builds for llama.cpp and for mlx-audio on Apple silicon.
NVIDIA Parakeet TDT v2 and v3 (2025)
No longer the most accurate, but still the best balance of accuracy and speed. At 600 million parameters, both are within about 0.9 points of the best open model and around seven times faster than Qwen3-ASR 1.7B. Their Token-and-Duration Transducer decoder predicts each token together with how many audio frames to skip, which is what makes them so fast. Version 3 covers 25 European languages with automatic detection and is the best small NVIDIA model on long recordings (12.60% on the long-form track). We compare it with Whisper in Whisper vs Parakeet.
Parakeet v3 now has several derivatives aimed at smaller devices, all under CC-BY-4.0:
- Parakeet Ultra (moondream, September 2026) is a post-trained full-precision version of v3 that its authors say beats the original on every benchmark they report.
- Parakeet Redux (moondream, September 2026) stores every encoder weight as -1, 0 or +1, which brings it to 178 MB.
- Phonon-2 (Fermion Research, September 2026) is an English model trained to hold its encoder weights in about 2.1 bits, a 164 MB download, 15 times smaller than the original. It scores 7.86 on the leaderboard, still ahead of Whisper large-v3. Phonon-2.5, released on 24 October 2026, extends the approach to 25 languages in a 156 MB download; it is not on the leaderboard.
NVIDIA Canary-Qwen 2.5B and Canary-1B-v2 (2025)
Canary-Qwen pairs a FastConformer speech encoder with a Qwen3 1.7B language model and was the most accurate NVIDIA model on the leaderboard until Nemotron 3.5 Transcribe arrived. Canary-1B-v2 does speech recognition and translation between English and 24 European languages.
NVIDIA Nemotron 3.5 ASR (June 2026)
A 600M streaming model for 40 locales, grouped by how ready each language is. It is cache-aware: each new chunk of audio reuses the model's previous state instead of recomputing overlapping context, which suits live captions. It pairs with the new Nemotron 3 Diarization model for speaker labels.
IBM Granite Speech 4.1 and 5.0 TurboCTC (2026)
Besides the new 5.0 2B, IBM has two very different earlier models. Granite Speech 4.1 2B couples a conformer encoder with a small Granite language model for six languages. Granite Speech 5.0 TurboCTC, released in August 2026, is an English-only 470M model with no language-model decoder at all, just an encoder trained with CTC. After a speed-up on the leaderboard in October it runs at about 21,000 RTFx and, at 7.41, still beats Whisper; a non-commercial variant scores 6.98. Only a 290M Zipformer model under a non-commercial licence is faster. IBM pitches TurboCTC at laptops and phones.
Cohere Transcribe (March 2026)
A 2B model for 14 languages, and one of the most accurate open models on the leaderboard's long-form track (11.88% across earnings calls and the CORAAL interview recordings), behind a long-form fine-tune of Parakeet (9.81%) and Nemotron 3.5 Transcribe (10.52%). If you mostly transcribe hour-long recordings, that track matters more than short clips.
TheWhisper large-v3-turbo (TheStage AI)
A fine-tuned version of Whisper large-v3-turbo, optimised for NVIDIA GPUs, Apple silicon and Android phones. Added to the leaderboard in October 2026, it scores 6.74, well ahead of the original turbo (8.75). It shows how much accuracy fine-tuning can still add to Whisper's architecture.
Mistral Voxtral (2025 and 2026)
Voxtral Small (24B) and Mini (3B) combine transcription with audio understanding: you can ask questions about a recording. Voxtral Transcribe 2, from February 2026, adds an open 4B realtime model with a delay you can set between 80 ms and 2.4 s. In October 2026 Mistral released an Arabic version of that realtime model, covering Modern Standard Arabic and regional dialects.
Microsoft VibeVoice-ASR (January 2026) and MOSS-Transcribe-Diarize (2026)
Both are part of 2026's biggest trend: one model that transcribes and labels speakers in a single pass. VibeVoice-ASR takes up to 60 minutes of audio in one 64K-token context and outputs who spoke, when, and what. MOSS-Transcribe-Diarize does the same in 50+ languages at under 1B parameters. Nemotron 3.5 Transcribe, above, is the newest and most accurate model of this kind.
Meta Omnilingual ASR (November 2025)
The widest language coverage by far: more than 1,600 languages, over 500 never covered by any system before, with a character error rate under 10% for 78% of them. It scales a wav2vec 2.0 encoder to 7B parameters and includes a zero-shot variant that learns a new language from a few examples. Not the choice for English accuracy, but unmatched for low-resource languages.
Jais-ASR (IFM, October 2026)
An Arabic model from the Institute of Foundation Models, built on the Jais 3 4B language model. It covers Modern Standard Arabic and dialects from 19 Arab countries, follows speakers who switch into English, French, Spanish or Italian, and can add diacritics. Its model card reports the lowest average error on the Open Universal Arabic ASR Leaderboard (22.85% WER, 8.37% CER). Apache 2.0.
Cactus Whistle (September 2026)
At the other end of the scale: a speech-to-text model in a single 16.9 MB file that runs on a CPU with no GPU. It transcribes English, German, French, Spanish, Italian, Dutch and Polish with word timestamps, and its makers say silence and steady noise return an empty transcript instead of invented text. It is not on the Open ASR Leaderboard. Apache 2.0.
Kyutai STT (2025)
Streams text with a fixed delay by modelling audio and text as parallel token streams. The 1B model supports English and French with half a second of delay. Kyutai's models no longer appear on the current Open ASR Leaderboard.
What happened to Whisper
On English, Whisper large-v3 now ranks below more than twenty open models, including two models derived from Whisper itself (TheWhisper large-v3-turbo and distil-large-v3.5). It remains useful for its language breadth (about 99 languages), its translation-to-English mode, and the large ecosystem of community ports around it. OpenAI released no open successor. On 26 August 2026 it deprecated whisper-1 and its gpt-4o-transcribe models in the API in favour of a new closed gpt-transcribe model, with shutdown scheduled for February 2027. The open Whisper weights remain available.
Closed models set the pace
The fastest movement in 2026 is in commercial APIs. Zoom's scribe_v2.1_pro now tops the English leaderboard, ahead of Otter.ai's speech-v3.1, and ElevenLabs Scribe v2 leads long-form. Google's Gemini 3.5 Transcribe (August 2026), Microsoft's MAI-Transcribe-2 (September 2026) and its streaming version (October 2026), and Meta's Muse Voice Transcribe (September 2026) add streaming, speaker labels or wide language coverage. OpenAI's gpt-transcribe replaces Whisper in its API. They are accurate, but your audio goes to the provider. For the trade-offs, see on-device vs cloud transcription.
Which ones run well on a Mac or iPhone
Leaderboard speed is measured on data-centre GPUs. On Apple devices, what matters is whether a model has a good conversion for Apple hardware, fits in memory next to everything else, and copes with long recordings.
- Parakeet TDT v2/v3 are the practical sweet spot at 600M parameters, and conversions for Apple silicon already exist. What running Parakeet on Apple silicon involves covers the trade-offs.
- Low-bit Parakeet versions such as Phonon-2 (164 MB) and Parakeet Redux (178 MB) cut the download to under a tenth of the original; Phonon-2 still beats Whisper large-v3 on the English leaderboard.
- Whisper has the most mature community ports and the widest language range, at a cost in speed and a tendency to invent text over silence. See running Whisper locally on a Mac and why Whisper hallucinates.
- Apple's built-in speech recognition in iOS 26, macOS 26 and later needs no download and works offline for the languages Apple supports.
- Qwen3-ASR 0.6B, Granite 5.0 TurboCTC and Parakeet Ultra are the candidates to watch for on-device use; their size fits, and conversions are appearing.
- 2B-plus models with LLM decoders (Granite 5.0 2B, Canary-Qwen, Granite 4.1, Cohere, VibeVoice) are more demanding and better suited to a Mac with plenty of memory than a phone.
- Nemotron 3.5 Transcribe is a server model. NVIDIA lists only NVIDIA GPUs as supported hardware and documents no Apple silicon build; its own llama.cpp-based runtime defaults to a GPU with 32 GB of memory.
Frequently asked questions
What is the most accurate open-source speech recognition model in 2026?
On the Hugging Face Open ASR Leaderboard for English (9 November 2026), NVIDIA's Nemotron 3.5 Transcribe leads the open models with a 6.13% average word error rate, just ahead of Alibaba's Qwen3-ASR 1.7B at 6.15%, followed by IBM's Granite Speech 5.0 2B at 6.68% and Cohere Transcribe at 6.72%. The two leaders are effectively tied on accuracy, but Qwen3-ASR 1.7B is about a twentieth of the size.
Why are the Open ASR Leaderboard scores higher than earlier in 2026?
The test got harder. The English average now includes Indian-accented conversations, two private sets that model makers cannot train on, and, since 9 October 2026, a set of noisy and reverberant speech (URGENT24). On 2 November 2026 the two LibriSpeech audiobook sets, where every model scores well, were dropped from the average. Every model's average rose, so compare models only within the same version of the leaderboard.
Is Whisper still the best open speech model?
No, not for English accuracy. More than twenty open models now beat Whisper large-v3 on the Open ASR Leaderboard, including smaller and much faster ones like Parakeet TDT. Whisper is still valuable for its roughly 99 languages, translation to English, and mature tooling.
Which open speech model is fastest?
IBM's Granite Speech 5.0 TurboCTC runs at about 21,000 times real time on the leaderboard's GPU. NVIDIA's Parakeet TDT models run at about 6,000 with standard decoding and up to about 17,000 with a faster GPU decoder. Both are also more accurate than Whisper large-v3. A 290M Zipformer model is slightly faster but has a non-commercial licence.
Which open model supports the most languages?
Meta's Omnilingual ASR covers more than 1,600 languages. Among models focused on accuracy, Qwen3-ASR covers 52 languages and dialects, Parakeet TDT v3 and Canary-1B-v2 cover 25 European languages, and Whisper covers about 99.
Can I use these models commercially?
Most can. Qwen3-ASR, Granite, Cohere Transcribe, Voxtral and Jais-ASR use Apache 2.0; Parakeet, Canary and their derivatives use CC-BY-4.0, which requires attribution; Whisper and VibeVoice use MIT. Some variants, such as Granite 5.0 TurboCTC-nc, are non-commercial. Check each model card.
Sources
- Hugging Face Open ASR Leaderboard (all WER and RTFx figures, version 1.15.1 of 9 November 2026; dataset list and change log in the leaderboard's About tab)
- Hugging Face: Open ASR Leaderboard private data
- NVIDIA Nemotron 3.5 Transcribe model card and technical report
- Qwen3-ASR technical report, arXiv:2601.21337
- IBM Granite Speech 5.0 2B model card
- "Canary-1B-v2 & Parakeet-TDT-0.6B-v3", arXiv:2509.14128
- Granite Speech 5.0 TurboCTC announcement
- moondream Parakeet Redux model card and Parakeet Ultra model card
- Fermion Research Phonon-2 model card and Phonon-2.5 model card
- TheStage AI TheWhisper large-v3-turbo model card
- Mistral: Voxtral Transcribe 2 and Voxtral Mini 4B Realtime Arabic model card
- VibeVoice-ASR, arXiv:2601.18184
- Omnilingual ASR, arXiv:2511.09690
- IFM Jais-ASR model card
- Cactus Whistle model card
- OpenAI API changelog (Whisper deprecation)
- Microsoft AI: MAI-Transcribe-2 and MAI-Transcribe-2-Streaming
- Model details for Parakeet TDT v2, Canary-Qwen 2.5B, Nemotron 3.5 ASR, Granite Speech 4.1, Cohere Transcribe, MOSS-Transcribe-Diarize and Kyutai STT come from each maker's published model card.