Choose a model

TranscribeX runs several families of speech recognition models on your Mac. Which one is best depends on your language, how fast you need results, and whether you need timestamps and speaker labels. Start with the quick picks, then use the table for details.

Quick picks

• English, free: Apple On-Device if your Mac runs macOS 26 or later. Otherwise Whisper Tiny.
• English, Pro: Parakeet V2. Accurate and much faster than Whisper Large V3.
• European languages (Pro): Parakeet V3. It detects the language automatically.
• Any of 97 languages, or an accent other models struggle with (Pro): Whisper Large V3 Turbo, or Large V3 for the highest Whisper accuracy.
• German, Chinese, Korean, French and other tuned languages (Pro): the matching language-tuned Whisper model.
• Live transcription: Parakeet EOU 160ms for English (free), Nemotron Streaming 560ms for other languages (Pro).

Set the model new jobs use in Settings › Transcription › Default Model, or change it for one file in the job’s model menu.

Our speed and accuracy test

We ran the same 8-minute English recording through five models on an M4 Pro Mac. Lower WER (word error rate) is better.

Our own test, run October 17, 2026 on a Mac with an Apple M4 Pro chip and 24 GB of memory, macOS 27.0.1. Audio: 73 clean read-speech clips from the LibriSpeech validation set (hf-internal-testing/librispeech_asr_dummy on Hugging Face), joined into one 8-minute file, 1,150 reference words. Each model was run twice through the TranscribeX command line and the second run is shown; times are end to end, including loading the model. Word error rate uses simple normalization (lowercase, punctuation removed). Clean read speech is easier than a real meeting, so expect higher error rates on your own calls, and treat differences under one point as noise on a sample this size. Qwen3 ASR was not timed in this round; Whisper Large V3 Turbo, Parakeet V3 and Nemotron were not part of it.

For leaderboard numbers and deeper comparisons, see speech recognition models.

Every model for file transcription

Sizes are download sizes as shown in the app. “Recommended” matches the badge in the model list.

Cohere supports English, French, German, Spanish, Italian, Portuguese, Dutch, Polish, Greek, Arabic, Japanese, Chinese, Korean and Vietnamese. The language menu in each job shows exactly what the selected model supports.

Language-tuned Whisper models (Pro)

These are Whisper Large V3 models fine-tuned by the community for one language. Use them when the general model makes too many mistakes in that language.

Need another language? The Whisper model pane says to email requests for fine-tuned models to support@transcribex.io.

Live-only models

These stream text as you speak and appear in the model menu of the live transcription page. You can’t use them for files.

Online models (Pro)

If you prefer a cloud model, TranscribeX can use your own ElevenLabs or Mistral account. These send your audio to that company’s servers, and you pay them for usage.

• ElevenLabs: scribe_v1, scribe_v2 (99 languages) and Scribe v2 Realtime for live use. Turn on Settings › ElevenLabs Online › Enable ElevenLabs and enter your API Key.
• Mistral: voxtral-mini-2602 and voxtral-mini-latest (13 languages). Turn on Settings › Mistral Online › Enable Mistral and enter your API Key.

Both have a Diarize option that lets the service label speakers instead of TranscribeX.

Timestamps, speakers and reflow

Some features need timestamps from the model:

• Speaker diarization (Pro) needs segment timestamps. Qwen3 0.6B, Qwen3 1.7B and both Cohere models don’t produce them (the model list marks them Timestamps not supported), so diarization is skipped and timestamps are hidden for those jobs.
• Reflow Transcripts…, which re-splits subtitles by characters or words per line, needs word-level timings. Whisper, Parakeet and Qwen3 ForcedAligner provide them. Apple On-Device transcripts don’t, and Mistral with Diarize on doesn’t either.

For subtitles and speaker labels, pick Whisper, Parakeet or Qwen3 ForcedAligner.

Download and manage models

Download models in Settings › Whisper Model, Parakeet Model, Qwen Model and Cohere Model, or click More Models... in a job’s model menu. See manage models for storage, deleting models and hardware settings.