Cohere Transcribe
Cohere Transcribe is a 2-billion-parameter speech model and the most accurate one FreeFlow ships, at a 5.4% average English word error rate. It covers 14 languages including Arabic, Japanese, Chinese, Korean and Vietnamese, runs as a 2.2 GB INT8 CoreML model on macOS 15 or later, and costs a couple of seconds a take. It transcribes the language the mode names — there is no auto-detect.
| Size | About 2 billion parameters; 2.2 GB download as an INT8 CoreML port |
|---|---|
| Languages | 14: English, French, German, Spanish, Italian, Portuguese, Dutch, Polish, Greek, Arabic, Japanese, Chinese, Korean, Vietnamese |
| Average English WER | 5.4% (Cohere's Open ASR Leaderboard figure for the model, checked 2026-11-09) |
| Speed on Apple Silicon | 1.5 to 3 seconds a take; a fixed 35-second encoder window is the reason |
| Language detection | None — set the language in the mode |
| Runs through | FluidAudio |
| Needs | macOS 15+, Apple Silicon |
When the extra second is worth it
Accuracy you can measure: a point below Parakeet V3's 6.3% is roughly one fewer wrong word in a hundred. For long-form dictation — a report, a brief, an essay — that adds up, and the wait is tolerable because you are not reading each sentence as it lands. For chat, prompts and commands, Parakeet's speed wins. The common setup is a Parakeet mode for everyday dictation and a Cohere mode for long careful work, a menu-bar click apart.
The 14 languages, and the ones only it has
Arabic, Japanese, Chinese, Korean and Vietnamese are not in Parakeet V3's 25, so for those languages the choice is Cohere Transcribe or a Whisper model — and Cohere is the more accurate of the two. Because there is no auto-detect, each language gets its own mode with the language set; a mode set to a language outside the 14 says so in red, and before 1.5.1 quietly typed English instead.
The fixed window
Cohere's encoder processes a fixed 35-second window regardless of how long you spoke, which is why a two-word take costs about as long as a two-sentence one. FreeFlow pads short takes and prefetches the model's 2 GB of weights while you are still speaking if it has been idle, so the first take after a break is no slower than the rest (about a third of a second to re-read, since 1.5.2).
Requirements
macOS 15 Sequoia or later, Apple Silicon, and 2.2 GB of disk. The first load after download takes a moment to prepare the model; after that it is resident until you switch engines.
Questions people ask
Is Cohere Transcribe the most accurate model in FreeFlow?
Yes, at 5.4% average English WER against Parakeet V3's 6.3% and Whisper Large's 7.4%.
Why is it slower than Parakeet?
A fixed 35-second encoder window and a 2-billion-parameter model. Expect 1.5 to 3 seconds a take.
Does it detect the language?
No. Set it in the mode; a mode per language is the usual setup.
Does it work on macOS 14?
No. It needs macOS 15 or later. Parakeet and Whisper work on macOS 14.