Whisper Models and Local Transcription Benchmarks
Updated September 28, 2026
Compare local speech recognition models by speed, accuracy, language coverage, and Apple Silicon performance.
Every local speech model we ship or have benchmarked — what each is good at, how fast it runs on Apple Silicon, and which one to pick for your language. All numbers below come from our published benchmark articles.
Local Speech Models at a Glance (2026)
The short version: on a Mac, use Parakeet V3 for English and European languages (it's the default in Whisper Notes), SenseVoice for Chinese, Japanese, Korean, and Cantonese, and Whisper Large-v3 Turbo for everything else in its ~100-language long tail.
| Model | Best for | Accuracy | Speed on Apple Silicon | Speed vs Whisper Large V3 Turbo | Memory needed | Languages | In Whisper Notes |
|---|---|---|---|---|---|---|---|
| Parakeet V3 (NVIDIA, 0.6B) | English & European languages | 6.32% English WER (Open ASR Leaderboard) | ~10× faster than Whisper Turbo | 7.2× faster (103× vs 14.3× realtime, M4 Pro) | 0.09 GiB peak | 25 | Yes — default on Mac |
| Whisper Large-v3 Turbo (OpenAI, 809M) | Widest language coverage | 7.83% English WER (Open ASR Leaderboard) | ~5× faster than Large V3 | Baseline | 1.6 GB | ~100 | Yes — optional download (~1.6 GB) |
| SenseVoice Small (FunAudioLLM) | Chinese, Japanese, Korean, Cantonese | Purpose-built for CJK | 27-min Chinese podcast in 13.83 s (M4 Pro) | 9.0× faster (118× vs 13.1× realtime, M4 Pro) | 0.95 GiB peak | CJK focus | Yes — since v1.5.0 |
| Whisper Large V3 (OpenAI, 1.55B) | Reference baseline | Near-identical to Turbo | Baseline (slowest here) | About 5× slower than Turbo | Not measured — not shipped | ~100 | No — Turbo supersedes it |
| Voxtral (Mistral) | Speech + language understanding | See our benchmark article | Server-class models | Not measured | Not measured | Multilingual | No — covered in blog |
WER = word error rate; lower is better. English WER comes from the Hugging Face Open ASR Leaderboard; 25-language averages from FLEURS. Speed figures are our in-house tests on an M4 Pro running Whisper Notes — full methodology in the linked benchmark articles below. All models listed as "in Whisper Notes" are included in the one-time price. Memory is each engine's peak process footprint in our own FLEURS run on an M5 MacBook Air — that is how Parakeet V3 and SenseVoice Small were measured. The 1.6 GB for Whisper Large V3 Turbo is the working-set figure the app documents, not a peak we measured. The two speed ratios are computed from realtime factors measured on the same audio on an M4 Pro.
Highest-traffic model benchmarks
Start here when the searcher is comparing local ASR model speed, decoder layers, silence hallucinations, and Apple Silicon performance.
Whisper Transcription Guide
What Whisper transcription is, every way to run it (API, CLI, apps), accuracy by model size, and how to run it offline.
Parakeet V3 vs Whisper
A benchmark page for Parakeet V3, speed, silence hallucinations, and Apple Silicon transcription.
Whisper Large V3 Turbo vs V3
A benchmark of Whisper Large V3 Turbo against V3, including speed, decoder layers, and Apple Silicon trade-offs.
Mistral Voxtral Models
An overview of Voxtral speech models and how they compare with local transcription options.
Model-specific local transcription pages
These pages connect model searches to practical Mac, iPhone, and multilingual transcription workflows.
MacWhisper Alternative
Explains how Whisper Notes uses the same local Whisper engine while adding Fn key dictation and iPhone capture.
SenseVoice for CJK Transcription
A model-focused article for Chinese, Japanese, and Korean transcription speed and accuracy.
Best Whisper App for iPhone and Mac
A platform-focused guide for people searching for a practical Whisper app rather than a model paper.
Use models in real workflows
Where these models actually get used: iPhone recording, Mac Fn-key dictation, and end-to-end offline transcription workflows.
Voice to Text Offline
How local models power iPhone recording, Mac Fn key dictation, and offline voice-to-text workflows.
Offline Transcription Guides
The broader guide hub for users who arrived through a model search but need an end-to-end workflow.
Offline Speech to Text Complete Guide
A complete guide to local transcription, privacy, speed, supported devices, and when offline speech to text beats cloud transcription.
More Guide Hubs
Explore the other Whisper Notes topic hubs for models, alternatives, offline workflows, and Mac dictation.
Offline Transcription Hub
Every guide to transcribing audio offline — apps, workflows, and privacy fundamentals for Mac and iPhone.
Alternatives Hub
Whisper Notes vs Otter.ai, MacWhisper, Superwhisper, Notta, and Typeless — feature and pricing comparisons.
Mac Dictation Hub
System-wide voice typing on macOS — dictate into any app with the Fn key, fully offline.