Pick your language, hardware and workload. Get a practical shortlist of local ASR models with license notes, streaming support and links to primary sources.
You do not need to understand every ASR family first. Jump straight to the constraint that can eliminate the wrong models fastest.
The score is deterministic and explainable. It is a fit score for your selected constraints, not a universal accuracy leaderboard.
Use a preset to fill the constraints that usually matter first, then adjust any field before you decide.
Default setup shown: English, long recordings, 8–12 GB NVIDIA GPU, balanced priority. Change the fields to re-rank. Scores are deterministic guidance, not benchmark claims.
“Best ASR model” is usually the wrong question. A better question is which model is a good fit for your language, hardware, latency and licensing needs.
Official language lists differ sharply between model families. The matcher distinguishes explicit support, adaptation-only coverage and languages that are not listed by the reviewed source.
Some models have native streaming architectures, while others are streamed through runtimes or chunking wrappers. The matcher treats those differently.
Each model keeps its own license and source. We do not collapse custom, attribution and permissive licenses into a single “commercial” badge.
Speech Model Match is designed as a sequence: filter by hard constraints, compare documented capabilities, test the finalists on the same audio, then inspect the errors that matter to your product.
Use language, hardware, workload and feature requirements to remove obvious mismatches.
Run the matcher →Put two or three reviewed models side by side before you spend time installing runtimes.
Compare models →Record WER, real-time factor and memory for multiple runs under the same conditions.
Open benchmark workspace →Measure substitutions, deletions and insertions with a human-verified reference transcript.
Calculate WER →: 6 open-weight models added with provisional ratings — Granite Speech 5.0 2B, Granite Speech 5.0 470M TurboCTC, Jais-ASR, Parakeet Redux, VibeVoice-ASR-Streaming-7B and Whistle. What changed.
Hugging Face downloads over the last 30 days, as of 11 Nov 2026. Full popularity tracker
The matcher's top three for typical situations, computed from the same reviewed data and scoring. Open a model for its full profile, languages and license, or adjust the matcher above for your exact case.
| Setup | #1 | #2 | #3 |
|---|---|---|---|
| Long recordings on an 8–12 GB NVIDIA GPU, balanced | Whisper turbo 79/100 | Parakeet TDT 0.6B v3 79/100 | Orukeet 79/100 |
| Live captions or voice input (native streaming + timestamps, speed first) | Nemotron 3.5 ASR Streaming 0.6B 91/100 | Moonshine Tiny Streaming (English) 86/100 | Moonshine Medium Streaming (English) 86/100 |
| Subtitles with timing, accuracy first | Whisper medium 87/100 | Whisper turbo 87/100 | Distil-Whisper large-v3 87/100 |
| Meetings with speaker labels (diarization + timestamps) | VibeVoice-ASR 7B 86/100 | Whisper medium 84/100 | Whisper turbo 84/100 |
| CPU-only machine, low memory use | VibeVoice-ASR-BitNet 78/100 | Parakeet Redux 78/100 | Whisper tiny 74/100 |
| Apple Silicon with 8–16 GB unified memory, long recordings | Whisper small 75/100 | VibeVoice-ASR-BitNet 73/100 | Whisper tiny 72/100 |
| Japanese live transcription (native streaming, speed first) | Nemotron 3.5 ASR Streaming 0.6B 91/100 | Voxtral Mini 4B Realtime 2602 86/100 | VibeVoice-ASR Streaming 1.5B 83/100 |
Need another language? The language support matrix shows which models list each of 104 languages. Every model also has its own profile page.
These head-to-head guides focus on the point where one model family becomes a better first benchmark than another.
Pillar guide to the main local ASR families and the constraints that should decide your shortlist.
Native streaming versus chunked transcription for live dictation, captions and voice agents.
How to compare explicit language lists, tiered support and very broad multilingual models.
Compare real replacement model families and learn when a different Whisper runtime is the better answer.
Choose between model-size flexibility and a modern high-throughput transcription family.
Choose by timestamps, alignment, punctuation and long-audio editing workflow.
No. This matcher does not accept audio files. Your selections are evaluated in your browser against the local model database.
No. It is a deterministic fit score for language, hardware and use case. Benchmark the shortlisted models on your own representative audio.
Maybe. The site shows each model’s reviewed license, but the actual current license text controls. Re-check it before production or redistribution.
It is not the same as ready-to-use transcription. For Nemotron 3.5, NVIDIA distinguishes transcription-ready locales from adaptation-ready locales that require fine-tuning.
They combine broad local-runtime availability with explicit language coverage. The best choice still depends on your hardware and audio.
They can cover far more languages, but coverage alone does not guarantee equal accuracy. Treat them as strong candidates to test, not automatic winners.