Cohere Transcribe 2B

This first-party Soniqo page documents Cohere Transcribe 2B from the local speech-swift / speech-core implementation. Hugging Face bundles are linked below after the integration notes.

Internal Page First

Landing cards and docs menus now point here first; source model and bundle links remain available from this page.

At a Glance

Model	Cohere Transcribe 2B
Role	High-accuracy multilingual offline speech-to-text
Backend	Native MLX on Apple Silicon
Output	Plain-text transcription
Languages	14 languages
License	Apache-2.0
Status	Published FP16, INT5, and INT8 bundles; INT5 is the default
Source	Cohere Transcribe
Swift product	`CohereTranscribeASR`
CLI / runtime	`speech transcribe --engine cohere`

Use

The snippet below mirrors the current speech-swift API or command exposed by the repo.

# INT5 is the default.
speech transcribe recording.wav --engine cohere

# Select another published precision and pass a language hint.
speech transcribe recording.wav --engine cohere --model int8 --language de

# --model also accepts a Hugging Face model ID or a local directory.
speech transcribe recording.wav --engine cohere --model /models/cohere/int5

Benchmark

Measured on 2026-07-22 over 80 English FLEURS read-speech utterances totalling 759.56 seconds, on an Apple M5 Pro with 48 GB of memory running macOS 26.5.2. Each variant ran in its own process from a release build.

Variant	Bundle	WER	Δ WER	Mean RTF	Overall ×RT	Footprint
FP16	3.85 GiB	6.178%	—	0.0444	23.59×	6,724 MiB
INT5	1.62 GiB	6.288%	+0.110 pp	0.0150	69.64×	2,582 MiB
INT8	2.25 GiB	6.178%	+0.000 pp	0.0159	65.74×	3,292 MiB

Physical footprint is the deployment-relevant unified-memory figure, because MLX may memory-map weight files. FLEURS is English read speech: these numbers do not claim parity on conversational, telephone, noisy, or speaker-heavy audio.

Model Links

Implementation Notes

The runtime resamples mono Float32 PCM to 16 kHz and handles long recordings in overlapping chunks.
INT5 is the practical default: on the validated English FLEURS run it used a 1.62 GiB bundle, 2,582 MiB physical footprint, and 0.0150 mean RTF.
MLX affine quantization supports 2, 3, 4, 5, 6, and 8 bits, not INT7; choose INT8 when prioritizing quality.
This is a non-streaming engine. The CLI rejects --stream instead of silently changing behavior.