Spokenly logoSpokenly Docs

Speech-to-Text Models Compared

Compare speech-to-text models by accuracy and speed: ElevenLabs Scribe, Parakeet Ultra, Apple Speech Analyzer, Whisper and more, cloud and on-device.

Speech-to-text models, also called ASR models, turn your voice into text. Spokenly rates the accuracy and the speed of each model on a scale of 5. Higher is better. On Mac, Windows, Linux and in the Chrome extension, the model list shows these ratings in the Accuracy and Speed columns.

Quick picks

  • Most accurate: ElevenLabs Scribe, Microsoft MAI-Transcribe 2, Grok STT, Voxtral Small Transcribe 3, Gemini 3.5 Transcribe, Fun-ASR Flash and AssemblyAI Universal 3.5 Pro. These cloud models have accuracy 5.0. ElevenLabs Scribe is the default model.
  • Accurate and fast: Grok STT (accuracy 5.0, speed 5.0) and Microsoft MAI-Transcribe 2 (accuracy 5.0, speed 4.5).
  • Text while you speak: Soniox Realtime (accuracy 4.5, speed 5.0). For other options, see Real time.
  • Offline on any device: Parakeet Ultra (accuracy 4.5, speed 4.5, 25 languages). It runs on Mac, iPhone, Windows, Linux and Android.
  • Most accurate offline on Mac and iPhone: Cohere Transcribe Q4 (accuracy 5.0, speed 4.0, 14 languages).
  • Offline, English only: Bark 3.0, Parakeet V2 and Parakeet Unified (accuracy 4.5, speed 4.5).
  • Medical terms: Nova-3 Medical (Real-time) (accuracy 4.0, speed 5.0). It understands English only.
  • Built into Apple devices: Apple Speech Analyzer (accuracy 4.0, speed 5.0). It needs macOS 26 or iOS 26.

How the ratings work

  • Accuracy shows how few words a model gets wrong. Speed shows how fast the text appears.
  • The accuracy rating is empirical: it comes from feedback from Spokenly users, not only from word error rates in public speech recognition tests, because word error rate alone does not show how good a dictated text is. These tests use English speech, so in other languages the order of models can be different. Try two or three models in your language.
  • For cloud models, speed is based on how fast Spokenly users get their text. On-device speed also depends on your computer or phone.
  • The ratings are for the default mode of each model. In real time, some models work differently.

Cloud models

Cloud models need internet. Spokenly sends your audio through its servers to the model provider and shows the text. These models come with Spokenly Pro, and you can try them with a one-time free trial before you buy. Some of them also work with your own API key, see AI Models.

ModelAccuracySpeedReal timeLanguages
ElevenLabs Scribe5.04.0With Live transcriptionMany
Microsoft MAI-Transcribe 25.04.5With Live transcription60
Grok STT5.05.0NoMany
Voxtral Small Transcribe 35.04.0No50
Gemini 3.5 Transcribe5.04.0With Live transcription85+
Fun-ASR Flash5.03.0No30
AssemblyAI Universal 3.5 Pro5.01.5Live mode on Mac and iPhoneMany
Soniox Realtime4.55.0Yes60+
Soniox Async4.52.0No60+
GPT Transcribe4.53.5NoMany
Parakeet V34.03.0No25
Nova-3 Medical (Real-time)4.05.0YesEnglish
Real-time Nova-33.55.0YesMany
Real-time Nova-3 (English Only)3.55.0YesEnglish
Whisper Large v33.54.5No99+
Whisper v3 Turbo3.04.5No99+
Cartesia Ink3.05.0YesEnglish

On-device models

On-device models run on your computer or phone. They are free, work without internet and keep your audio on your device. Download one in AI Models > Local Models.

ModelAccuracySpeedSizeLanguagesDevices
Cohere Transcribe Q45.04.01 GB14Mac, iPhone
Cohere Transcribe (INT8)5.04.04.7 GB14Mac, iPhone
Parakeet Ultra4.54.5632 MB25Mac, iPhone, Windows, Linux, Android
Bark 3.04.54.5436 MBEnglishMac, iPhone, Windows
Parakeet Unified4.54.51.2 GBEnglishMac, iPhone, Windows, Android
Parakeet V24.54.5496 MBEnglishMac, iPhone, Windows, Linux, Android
Parakeet V34.04.5496 MB25Mac, iPhone, Windows, Linux, Android
Parakeet Redux4.04.5220 MB25Mac, iPhone, Windows
Phonon-24.04.5358 MBEnglishMac, iPhone, Windows, Linux, Android
Nemotron 3.5 Multilingual 0.6B4.04.5665 MB100+Mac, iPhone
Nemotron 3.5 Latin 0.6B4.04.5665 MB6Mac, iPhone
Parakeet EOU 120M3.55.085 MBEnglishMac, iPhone
Distil-Whisper Large v3.5 (English Only)3.53.01.5 GBEnglishMac, iPhone
Whisper Large v33.52.52.9 GB99+Mac, iPhone
Whisper Large v3 Turbo3.03.01.5 GB99+Mac, iPhone, Windows, Linux
Whisper Large v23.03.52.9 GB99+Mac, iPhone
Whisper Small3.04.0465 MB99+Mac, iPhone, Windows, Linux
Whisper Large v3 Turbo (Quantized)2.53.5547 MB99+Mac, iPhone, Windows, Linux
Distil-Whisper Medium (English Only)2.53.5794 MBEnglishMac, iPhone
Distil-Whisper Small (English Only)2.54.0336 MBEnglishMac, iPhone
Whisper Base2.04.5141 MB99+Mac, iPhone, Windows, Linux
Whisper Tiny1.05.075 MB99+Mac, iPhone, Windows, Linux
Whisper Tiny (English)1.05.075 MBEnglishMac, iPhone
  • Sizes are for Mac and iPhone. On Windows, Linux and Android, some models use other files, so their sizes can be different.
  • On Mac and iPhone, Parakeet V3, Parakeet V2 and Parakeet Unified are named Parakeet TDT 0.6B V3, Parakeet TDT 0.6B V2 and Parakeet Unified 0.6B.
  • Parakeet V3, Parakeet Ultra and Parakeet Redux understand 25 European languages: Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Spanish, Swedish and Ukrainian.
  • Nemotron 3.5 Latin 0.6B understands English, Spanish, French, German, Italian and Portuguese.
  • Parakeet V2, Parakeet Unified, both Nemotron models and Cohere Transcribe Q4 need a Mac with Apple silicon.
  • Bark 3.0, Phonon-2, Parakeet Redux and both Cohere models need macOS 15 or iOS 18. Models other than Whisper need macOS 14 or iOS 17.
  • On Android, Parakeet Unified shows only on phones whose processor supports it.

Apple models

These models are part of macOS and iOS. Spokenly shows them on Mac and iPhone.

  • Apple Speech Analyzer: accuracy 4.0, speed 5.0. It runs on your device and needs macOS 26 or iOS 26. Before you use it, Spokenly downloads Apple's language pack for your language. To see the text while you speak, turn on Real-time transcription in its settings.
  • Apple Speech Recognition: Apple's older speech engine. Apple processes the audio.

Real time

Most models transcribe your speech after you stop. Real-time models show the text while you speak.

  • Always in real time: Soniox Realtime, Real-time Nova-3, Real-time Nova-3 (English Only), Nova-3 Medical (Real-time) and Cartesia Ink.
  • In real time when you turn it on: ElevenLabs Scribe, Microsoft MAI-Transcribe 2 and Gemini 3.5 Transcribe. Open the settings of the model and turn on Live transcription. On Mac and iPhone, AssemblyAI Universal 3.5 Pro has Mode > Live in its settings.
  • On device: Parakeet Unified, Nemotron 3.5 Multilingual 0.6B, Nemotron 3.5 Latin 0.6B, Parakeet EOU 120M, and Apple Speech Analyzer with Real-time transcription on. On Windows and Android, Parakeet Unified.

In real time, these models use their streaming versions, so their accuracy can be different from the tables. In English, Artificial Analysis measures a word error rate of 2.5% for Microsoft MAI-Transcribe 2 in real time, 2.9% for ElevenLabs Scribe v3 Live and 4.6% for Soniox v6 Real-Time (November 2026). Source: Artificial Analysis (artificialanalysis.ai)

Common comparisons

Apple Speech Analyzer vs Parakeet Ultra

Parakeet Ultra is a little more accurate: 4.5 against 4.0. Apple Speech Analyzer is a little faster: 5.0 against 4.5. With Real-time transcription on, it shows the text while you speak. Both run on your device and work offline.

Apple Speech Analyzer needs macOS 26 or iOS 26. Parakeet Ultra is a 632 MB download, understands 25 languages and also runs on Windows, Linux and Android.

Parakeet vs Whisper

Parakeet models are more accurate and faster than Whisper models of a similar size. Parakeet Ultra has accuracy 4.5 and speed 4.5. Whisper Large v3 Turbo has accuracy 3.0 and speed 3.0. Whisper understands about 100 languages, and Parakeet Ultra understands 25. For a language that Parakeet does not support, use Whisper. For details, read Parakeet vs Whisper.

Cloud vs on-device

The most accurate cloud models have accuracy 5.0. They need internet and Spokenly Pro or your own API key. On-device models are free, work offline and keep your audio on your device. To keep Spokenly off the internet, use Local Only Mode.

ElevenLabs Scribe vs GPT Transcribe

ElevenLabs Scribe has accuracy 5.0 and speed 4.0. GPT Transcribe has accuracy 4.5 and speed 3.5. Scribe is the default model and can work in real time with Live transcription. GPT Transcribe shows the text after you stop speaking.

ElevenLabs Scribe vs Deepgram Nova-3

ElevenLabs Scribe is more accurate: 5.0 against 3.5 for Real-time Nova-3. Nova-3 is faster: 5.0 against 4.0, and it always shows the text while you speak. Scribe does this when you turn on Live transcription.

Soniox vs Deepgram Nova-3

Soniox Realtime and Real-time Nova-3 both show the text while you speak and have speed 5.0. Soniox Realtime is more accurate: 4.5 against 3.5. Both come with Spokenly Pro. Nova-3 also works with your own Deepgram API key.

Deepgram Nova-3 vs Whisper

Real-time Nova-3 and the cloud Whisper Large v3 have the same accuracy, 3.5. Nova-3 shows the text while you speak, and Whisper shows it after you stop. Whisper also runs on your device and understands 99+ languages. For English medical terms, use Nova-3 Medical (Real-time), with accuracy 4.0.

On this page