The best ASR model in 2026 depends on three questions: which languages you need, whether you can send audio to an API or must run on your own hardware, and how much latency you can tolerate. With that said, the short list is stable. Open source: OpenAI's Whisper large-v3 for multilingual breadth, NVIDIA's Canary-Qwen-2.5B for top English accuracy (~5.6% average WER on the Hugging Face Open ASR Leaderboard), and Parakeet-TDT-0.6B-v3 for extreme throughput. Commercial APIs: Deepgram Nova-3, AssemblyAI Universal-3, Google Chirp, Azure AI Speech, and Amazon Transcribe, which now sit within a couple of WER points of each other on public benchmarks.
The more useful truth is that leaderboard accuracy has largely converged, and the gap between any of these models on your audio is usually smaller than the gap between benchmark audio and your audio. This guide compares the field as of mid-2026, then covers how to choose, and why fine-tuning on in-domain data matters more than the logo on the model.
If you want the fundamentals first, start with what is automatic speech recognition.
The Best Open-Source ASR Models in 2026
OpenAI Whisper large-v3 and large-v3-turbo. Still the reference point. Large-v3 (1.55B parameters, encoder-decoder Transformer, MIT license) covers 99 languages and is unmatched for ecosystem: faster-whisper, whisper.cpp, WhisperX, and first-class Hugging Face fine-tuning support. Large-v3-turbo (809M) cuts the decoder down for several-times-faster inference with a modest accuracy cost. Whisper no longer tops English leaderboards (roughly 7-8% average WER on the Open ASR Leaderboard versus ~5.6% for the leaders) and it can hallucinate on silence and long-form audio, but for multilingual production it remains the default.
NVIDIA Canary and Parakeet. NVIDIA's NeMo models bracket the accuracy-speed trade-off. Canary-Qwen-2.5B, a Conformer encoder feeding a Qwen-based LLM decoder, leads the established field on the Open ASR Leaderboard at ~5.63% average WER (English). Canary-1B-v2 extends coverage to 25 languages with translation support under a permissive license. Parakeet-TDT-0.6B-v3 (CC-BY-4.0) covers 25 European languages with automatic language detection and posts the highest throughput among multilingual models on the leaderboard (thousands of times faster than real time on a datacenter GPU), making it the go-to for bulk transcription. Limits: language coverage is Europe-centric, and the LLM-decoder Canary models are markedly slower than TDT/CTC decoders.
Mistral Voxtral. Mistral's Apache-2.0 speech models (released 2025, with a streaming-capable Voxtral Transcribe generation in early 2026) post Whisper-beating multilingual WER, but across a much narrower set of languages (around a dozen majors). Worth benchmarking if your languages are on the list.
Meta wav2vec 2.0, MMS, and Seamless. wav2vec 2.0 matters today mainly as the self-supervised pretraining backbone you fine-tune with CTC on a low-resource language. MMS scales that recipe to 1,100+ languages, by far the widest coverage anywhere, with the caveats of CC-BY-NC licensing and modest per-language accuracy. SeamlessM4T targets speech translation more than pure ASR. These are the tools for the long tail; see our guide to speech data for low-resource languages.
Kyutai STT. From the lab behind Moshi: streaming-native models (~1B-2.6B, CC-BY) with word-level timestamps and low latency, currently English and French only. Interesting for real-time voice agents you must self-host.
Vosk / Kaldi. The legacy hybrid stack. Choose it only for tiny-footprint offline deployment on constrained hardware; accuracy is well behind modern end-to-end models.
A note on leaderboards: the Open ASR Leaderboard reshuffles monthly. Several newer entrants (from Cohere and Chinese labs, among others) have posted sub-5.5% averages in 2026. Check it before committing, but read the next section before trusting it.
The Best Commercial ASR APIs in 2026
Deepgram Nova-3 / Flux. Consistently the latency leader, with batch pricing around $0.0043/min. Nova-3 posts ~5.3% WER on Deepgram's own real-world test suite (vendor-reported); Flux (2026) integrates end-of-turn detection, purpose-built for voice agents. Self-hosted deployment available on enterprise plans.
AssemblyAI Universal-3. Universal-2 was the streaming accuracy benchmark; Universal-3 Pro (February 2026) added a speech-LM architecture with natural-language "keyterm prompting" up to 1,500 terms (effectively lightweight domain adaptation without fine-tuning) at a vendor-reported ~5.6% mean WER. Strong audio-intelligence add-ons (diarization, summarization, PII redaction).
Google Cloud STT (Chirp). Chirp models descend from Google's USM work; 100+ languages and deep GCP integration, around $0.016/min. Accuracy on majors is competitive; the draw is language breadth plus enterprise compliance.
Microsoft Azure AI Speech. The widest locale coverage of any provider (140+), custom speech adaptation, containerized on-prem deployment, plus Microsoft's own MAI-Transcribe-1 model (April 2026) claiming state-of-the-art multilingual WER. The enterprise-compliance default.
Amazon Transcribe. Roughly 100 languages, ~$0.024/min, unremarkable benchmark accuracy but the path of least resistance inside AWS, with medical and call-analytics variants.
ASR Model Comparison Table
| Model | Params / architecture | Languages | License / access | WER positioning (mid-2026) | Deployment |
|---|---|---|---|---|---|
| Whisper large-v3 | 1.55B enc-dec Transformer | 99 | MIT | ~7-8% avg Open ASR (EN); strong multilingual | Self-host |
| Whisper large-v3-turbo | 809M (reduced decoder) | 99 | MIT | Slightly behind large-v3, much faster | Self-host |
| Canary-Qwen-2.5B | 2.5B Conformer + LLM decoder | English | Permissive (NVIDIA open) | ~5.6% avg; top of established open field | Self-host |
| Canary-1B-v2 | 1B Conformer enc-dec | 25 | CC-BY-4.0 | Near-SOTA at 3x smaller than rivals | Self-host |
| Parakeet-TDT-0.6B-v3 | 0.6B Conformer-TDT | 25 (EU) | CC-BY-4.0 | Good; fastest multilingual RTFx | Self-host |
| Voxtral Transcribe | ~4B | ~13 | Apache 2.0 | Beats Whisper on covered languages | Self-host / API |
| Meta MMS | ~1B wav2vec 2.0 + CTC | 1,100+ | CC-BY-NC | Modest per language; unmatched coverage | Self-host (non-commercial) |
| Kyutai STT | 1-2.6B streaming | EN, FR | CC-BY | Competitive streaming EN | Self-host |
| Vosk / Kaldi | Small hybrid | ~20+ | Apache 2.0 | Legacy tier | Embedded / offline |
| Deepgram Nova-3 | Proprietary | 30+ | API (+self-host option) | ~5.3% vendor real-world suite | API / on-prem |
| AssemblyAI Universal-3 | Proprietary speech-LM | 99 (batch) | API | ~5.6% vendor-reported | API |
| Google Chirp | Proprietary (USM lineage) | 100+ | API | Competitive on majors | API |
| Azure AI Speech | Proprietary | 140+ locales | API | Competitive; custom adaptation | API / container |
| Amazon Transcribe | Proprietary | ~100 | API | Mid-pack | API |
WER figures are averages on public or vendor test suites and are not directly comparable across rows; normalization and test material differ. Treat them as tiers, not rankings.
How to Choose an ASR Model by Use Case
- Multilingual product, self-hosted: Whisper large-v3 (or turbo for throughput). Nothing else combines coverage, license, and tooling.
- English-only, maximum accuracy, own GPUs: Canary-Qwen-2.5B; validate against the current leaderboard top.
- Bulk offline transcription at scale: Parakeet-TDT. Throughput is the whole game, and its WER is close enough for most pipelines.
- Real-time voice agent via API: Deepgram (Flux) or AssemblyAI streaming; latency and endpointing matter more than the last WER point.
- Regulated enterprise, many locales: Azure (containers for on-prem) or Google.
- Long-tail languages: MMS to bootstrap, then fine-tune wav2vec 2.0/Whisper on collected data.
- Embedded/offline devices: Vosk, or a quantized Whisper small/turbo via whisper.cpp.
Why Leaderboard WER Isn't Your WER
Every number above was measured on curated benchmarks: audiobooks, prepared talks, clean podcasts. Your audio is contact-center calls with hold music, accented speakers on cheap headsets, overlapping conversational turns, domain jargon, and code-switching. It is routine to see a model that scores 6% on benchmarks land at 15-25%+ WER on real conversational telephony, with the errors concentrated on the words you care about: names, products, amounts.
The ranking can also flip out of domain. A model that wins on read English may lose to a "worse" model on accented spontaneous speech, simply because of what each saw in training. The only evaluation that counts is a held-out test set of your own audio, transcribed carefully, scored under one normalization scheme. Build that before you build anything else.
The Real Differentiator: Fine-Tuning on In-Domain Data
Base-model accuracy among the top ten has converged to within a couple of points. What has not converged is performance on your domain, and that is bought with data, not model selection. Fine-tuning Whisper (or CTC-tuning a wav2vec/Parakeet checkpoint, or using Azure/AssemblyAI adaptation features) on even 10-50 hours of matched conversational audio typically yields relative WER reductions of 20-50% on that domain, a far larger swing than any choice between leading base models. How much data you need per goal is covered in how much audio data you need to train ASR.
The binding constraint is matched data: spontaneous, conversational, in your language and acoustic conditions, with accurate transcripts and a license that permits model training. Scripted read-speech corpora don't close the gap (the mismatch is the disease, not the cure), which is why conversational vs. scripted data is the first sourcing decision to get right. If the data doesn't exist off the shelf for your domain, custom collection is the remaining lever.
Need training data?
SpeechData.ai supplies fine-tuning-ready conversational speech datasets in 60 languages: 500-2,000 hours each, dual-channel audio with time-aligned transcripts and full commercial licensing. Browse the datasets or contact us to benchmark a sample against your current model.