Whisper vs Parakeet vs Apple Speech: Which Transcription Engine to Use

Whisperer is the only Mac dictation app that supports three transcription engines — Whisper, Parakeet, and Apple Speech. Each uses different hardware, different models, and has different strengths. This guide helps you choose the right engine for your workflow.

Engine Overview#

Whisper — The Default Engine#

Whisper is OpenAI's open-source speech recognition model, running locally via whisper.cpp. Whisperer vendors the C library directly — no Python, no dependencies.

How It Works#

Whisper processes your complete audio recording after you release the record button. It runs on Apple Silicon's Metal GPU, which provides fast inference without blocking the CPU or Neural Engine for other tasks.

Model Sizes#

When to Use Whisper#

• You need the highest accuracy available
• You're dictating in non-English languages (99+ supported)
• You want the widest model selection to trade speed for accuracy
• You're doing file transcription (Whisper handles long audio well)

Parakeet — The Fastest Engine#

Parakeet is NVIDIA's CTC-based speech recognition model, running on Apple Silicon's Neural Engine via CoreML. It's the fastest option available — and it has a unique feature: CTC vocabulary boosting.

How It Works#

Parakeet uses the dedicated Neural Engine (ANE), which is separate from the GPU. This means Parakeet can run simultaneously with Whisper without resource contention. It's also used for Whisperer's live preview — the EOU (End-of-Utterance) engine that provides ~300ms preview latency.

CTC Vocabulary Boosting#

This is Parakeet's killer feature. Your personal dictionary entries and prompt words directly bias the CTC decoder at the acoustic level. This means:

• Project-specific terms are recognized more accurately
• Names, acronyms, and jargon are decoded correctly
• The boost happens during decoding, not as a post-processing step

When to Use Parakeet#

• You want the fastest possible transcription
• You have a large personal dictionary (vocabulary boosting)
• You want to leave the GPU free for other tasks (video editing, ML training)
• You primarily dictate in English (v2) or one of the 25 supported languages (v3)

Apple Speech — The Native Option#

Apple Speech uses macOS's built-in SpeechAnalyzer framework, available on macOS 26 (Tahoe) and later. Zero setup — no model downloads required.

How It Works#

Apple Speech leverages the system's built-in speech recognition pipeline. It's deeply integrated with macOS and uses Apple's own ML infrastructure. No additional models to download or manage.

When to Use Apple Speech#

• You're on macOS 26+ (Tahoe)
• You want zero setup — no model downloads
• You need quick, lightweight dictation without maximum accuracy
• You're low on disk space and can't store Whisper models

Head-to-Head Comparison#

Accuracy#

For English dictation:

*With vocabulary boosting enabled and dictionary configured.

Speed (Apple M1 Pro, typical 10-second recording)#

Resource Usage#

Parakeet's Neural Engine usage means your GPU stays free for video editing, 3D rendering, or ML workloads. This makes Parakeet ideal for multitasking.

Choosing Your Engine#

Hot-Swapping Engines#

Whisperer lets you switch between engines without restarting. Open Settings → Engine and select your preferred backend. The switch is instant — no app restart, no re-configuration.

You can even use different engines for different purposes:

• Parakeet for live dictation (fastest feedback)
• Whisper Large V3 for file transcription (maximum accuracy)
• Apple Speech when you're low on battery

Related: Offline Transcription Engines, Best Whisper Model for Dictation, Whisper vs Parakeet Transcription. See all features and pricing.