Skip to content
ZenVoice

Models

14 models. Zero servers.

A speech engine turns audio into text. An optional Smart formatting model turns spoken drafts into clean writing. Both run on your Mac, and ZenVoice downloads only the one you pick.

English or a European language on Apple Silicon

NVIDIA Parakeet TDT 0.6B V3

Auto-detect, or a language outside Parakeet's 25

Whisper Large V3 Turbo

Hindi and English, written as Hinglish

Hinglish Prime

macOS 26 and nothing to download

Apple Speech Analyzer

Setup recommends an engine for your language and your Mac. Switch any time in Models.

Speech to text

The engine that listens. Pick one; ZenVoice downloads only its file.

11 models
  • Apple Speech Analyzer

    macOS 26+No download

    Apple

    On-device recognition in macOS. Nothing to download. Requires macOS 26+.

    System languagesSpeechAnalyzerApple SLA

    Built in

    no download

  • NVIDIA Parakeet TDT 0.6B V3

    RecommendedApple Silicon

    NVIDIA

    The default for English and European languages. 6.9% WER at 73× real time.

    25 Europeanparakeet.cpp · MetalCC-BY-4.0

    0.9 GB

    download

  • NVIDIA Parakeet TDT 0.6B V2

    English

    NVIDIA

    The English-only Parakeet, on the same open parakeet.cpp runtime.

    Englishparakeet.cppCC-BY-4.0

    0.8 GB

    download

  • NVIDIA Nemotron 3.5 Multilingual 0.6B

    Streaming

    NVIDIA

    Streaming ASR for English, Spanish, French, German, Chinese, Japanese, Korean, Portuguese.

    8 languagesparakeet.cppOpenMDW-1.1

    0.9 GB

    download

  • Qwen3-ASR 0.6B

    MLX

    Qwen

    Compact multilingual recogniser running natively on MLX.

    MultilingualMLX · 6-bitApache-2.0

    0.8 GB

    download

  • Cohere Transcribe

    2B params

    Cohere Labs

    A 2B-parameter Conformer encoder-decoder. The largest engine in the catalogue.

    14 languagesONNX · INT8Apache-2.0

    2.9 GB

    download

  • Whisper Large V3 Turbo

    Auto-detectIntel + Apple Silicon

    OpenAI

    Auto-detect and the widest language coverage. 8.2% WER at 11× real time.

    99 languageswhisper.cppMIT

    0.5 GB

    download

  • Whisper Small

    Intel

    OpenAI

    Fallback for Macs that cannot run Turbo comfortably.

    99 languageswhisper.cppMIT

    0.5 GB

    download

  • Whisper Tiny (English)

    Smallest

    OpenAI

    The smallest download. Good for trying ZenVoice in a minute.

    Englishwhisper.cppMIT

    0.1 GB

    download

  • Whisper Tiny (Multilingual)

    Smallest

    OpenAI

    Tiny footprint with multilingual coverage.

    99 languageswhisper.cppMIT

    0.1 GB

    download

  • Hinglish Prime

    Hinglish

    Oriserve · ZenVoice

    A full Large V3 fine-tune that writes Hindi and English speech as romanized Hinglish.

    Hindi & Englishwhisper.cppApache-2.0

    1.5 GB

    download

Dictation enhancement

Used by Smart formatting: fillers and restarts out, numbers, dates and money written the way you'd type them.

3 models
  • Apple Intelligence

    Supported MacsNo download

    Foundation Models

    On-device Foundation Models built into macOS. Nothing is downloaded or uploaded. Ready on supported Macs.

    FoundationModels

    Built in

    no download

  • ZenPolish 1.7B v2.2

    NewRecommended

    Qwen3-1.7B · fine-tuned

    ZenVoice's own writing model: a fifth of the word errors of its stock base model, and 90% fewer words you never said, in 1.8 GB of memory.

    MLX · 6-bit + LoRA

    1.5 GB

    download

  • Qwen 3.5 2B

    Offline

    Qwen3.5-2B

    On-device Smart formatting for Macs without Apple Intelligence. 4-bit MLX, runs offline.

    MLX · 4-bit

    1.6 GB

    download

See the benchmark tables →
Ask AI about this pageChatGPT ↗Claude ↗Perplexity ↗Grok ↗Markdown