// catalog

Dictation models, published as data.

Vox reads https://voxd.cc/data/models.json. Built-in families load without extra install. New families ship as plugins.

catalog v2 · updated 2026-10-31

vox plugins
$ vox plugins list
mlx-vlm asr available bundle
$ vox plugins install mlx-vlm
Installed plugin mlx-vlm
$ vox transcribe file --model gemma-4-e2b-it clip.wav
restart voxd first, then transcribe

// models

21 entries

whistle

Cactus Whistle

on-device · file · finalized · word times

Optional separate-process CPU provider with word timings. File transcription up to 30 seconds; install and warm-up are explicit. Uses Cactus's prebuilt Needle runtime.

cactus-whistle

needle-cpu

plugin

parakeet:v3

Parakeet TDT v3

on-device · file · finalized · word times

Default file-transcription model. Multilingual FastConformer-TDT; pinned August 19 Core ML conversion. For model-backed live partials use the native streaming models.

parakeet-tdt

coreml

default

parakeet:v2

Parakeet TDT v2

on-device · file · finalized · word times

English-only TDT with a tighter vocabulary and better English recall.

parakeet-tdt

coreml

ready

apple:speech-transcriber

Apple SpeechTranscriber

on-device · file · finalized · word times

Core on-device parity option. Apple manages the model assets and audio format negotiation; Vox currently returns finalized file transcription and word timings.

apple-speech

speech-analyzer

ready

mlx-community/parakeet-tdt-0.6b-v3

Parakeet TDT v3 (MLX)

on-device · file · finalized

mlx-audio

mlx-audio

ready

mlx-community/parakeet-tdt-0.6b-v2

Parakeet TDT v2 (MLX)

on-device · file · finalized

mlx-audio

mlx-audio

ready

mlx-community/whisper-large-v3-turbo-asr-fp16

Whisper Large v3 Turbo

on-device · file · finalized

mlx-audio

mlx-audio

ready

mlx-community/Qwen3-ASR-1.7B-8bit

Qwen3-ASR 1.7B

on-device · file · finalized

Recommended Qwen3 ASR size on Apple Silicon. The current Vox mlx-audio bridge is file-based.

mlx-audio

mlx-audio

ready

mlx-community/Qwen3-ASR-0.6B-8bit

Qwen3-ASR 0.6B

on-device · file · finalized

mlx-audio

mlx-audio

ready

mlx-community/cohere-transcribe-03-2026-mlx-8bit

Cohere Transcribe 03-2026

on-device · file · finalized

Local MLX conversion of Cohere Transcribe. Offline rather than realtime; requires the mlx-audio ASR provider.

mlx-audio

mlx-audio

ready

mlx-community/nemotron-3.5-asr-streaming-0.6b-8bit

Nemotron 3.5 ASR Streaming 0.6B

on-device · file · finalized

Cache-aware streaming FastConformer-RNNT with configurable latency. The model is streaming-capable; the current Vox mlx-audio bridge is file-based.

mlx-audio

mlx-audio

ready

gpt-transcribe

GPT Transcribe

remote · file · finalized

Current OpenAI file-transcription model.

openai-transcribe

remote

ready

gpt-4o-transcribe

GPT-4o Transcribe

remote · file · finalized

openai-transcribe

remote

ready

gpt-4o-mini-transcribe

GPT-4o Mini Transcribe

remote · file · finalized

openai-transcribe

remote

ready

whisper-1

Whisper-1

remote · file · finalized

Legacy OpenAI Whisper API. The current Vox bridge uses finalized file transcription and does not yet request its timestamp or caption response formats.

openai-transcribe

remote

ready

gemma-4-e2b-it

Gemma 4 E2B

on-device · file · finalized

Install with `vox plugins install mlx-vlm`. The plugin speaks the Vox provider protocol and runs Gemma through mlx-vlm.

mlx-vlm

mlx-vlm

plugin

nemotron:streaming-560ms

Nemotron 3.5 Streaming Latin 560 ms

on-device · file · live · word times

Cache-aware native streaming alternative to Unified. Cache-aware ANE encoder, cumulative partials and greedy emission-frame word timings. End-of-utterance uses 1120 ms of RNNT blank duration, not a trained EOU token. English-US is the companion default; language-selectable embedded preload.

coreml-streaming

coreml

ready

nemotron:en-streaming-560ms

Nemotron English Streaming 560 ms

on-device · file · live · word times

Native cache-aware streaming, int8 encoder. RNNT blank-duration endpointing.

coreml-streaming

coreml

ready

parakeet:eou-320ms

Parakeet Realtime EOU 120M 320 ms

on-device · file · live · word times

Native streaming with trained EOU token. 630 ms input window, 320 ms stride, approximate word timestamps.

coreml-streaming

coreml

ready

parakeet:eou-160ms

Parakeet Realtime EOU 120M 160 ms

on-device · file · live · word times

Native streaming with trained EOU token. 160 ms input window, 80 ms stride; accuracy tradeoff at smallest context.

coreml-streaming

coreml

ready

nemotron:unified-560ms

Nemotron Labs Unified 0.6B 560 ms

on-device · file · live · word times

Default native live model. October 28 Core ML int8 conversion. Buffered streaming: 160 ms stride, 400 ms right context, up to 6.16 s rolling encoder window. Model-backed cumulative partials and approximate RNNT word timestamps; blank-duration endpointing. English-US companion default; language-selectable embedded preload.

coreml-streaming

coreml

ready

// plugins

New families install as plugins.

A plugin is an external JSON-RPC provider. Catalog refresh does not run it. Install writes ~/.vox/plugins/<id>/provider.json, then restart voxd.

whistle

asr · bundle

Optional Rust CPU speech provider for Apple Silicon. Register with `vox plugins install whistle`, restart voxd, then run `vox models install whistle` and `vox models preload whistle`.

mlx-vlm

asr · bundle

External JSON-RPC provider for Gemma 4 and other mlx-vlm audio models. Install with `vox plugins install mlx-vlm`, then restart voxd.

Models and plugins guide