How much does Cohere Transcribe Arabic cost?

Cohere Transcribe Arabic is Apache-2.0 open weights, so running it yourself is free of licensing cost. Cohere also offers hosted API access with standard rate limits and a Model Vault deployment option; both use Cohere's usual pay-as-you-go pricing rather than a bespoke rate for this model.

Which Arabic dialects does Cohere Transcribe Arabic handle?

Cohere Transcribe Arabic covers Modern Standard Arabic plus regional dialects, and is tuned for Arabic-English code-switching that comes up in enterprise call flows. Its 25.87 average WER is measured across Common Voice, SADAWER, MASC, MGB-2 and Casablanca, which together sample MSA and a wide dialect mix.

How does Cohere Transcribe Arabic compare to Whisper Large v3?

Cohere Transcribe Arabic cuts word error rate from Whisper Large v3's 36.86 to 25.87 on the same Arabic dialect benchmarks and reaches 525 RTFx on H100 versus Whisper Large v3's 146 — about 3.6x higher throughput. Human evaluators also preferred Cohere Transcribe Arabic in 95.8% of paired comparisons.

What is Cohere Transcribe Arabic not good at?

Cohere Transcribe Arabic focuses on single-language audio, so it can slip on heavy code-switched clips, and it ships with no automatic language detection, no timestamps and no speaker diarization. It can also transcribe non-speech noise without a VAD front end, so production stacks should add voice activity detection.

Where can I try Cohere Transcribe Arabic?

Cohere Transcribe Arabic weights are on Hugging Face at CohereLabs/cohere-transcribe-arabic-07-2026, with native Transformers and vLLM inference paths plus Colab and Kaggle notebooks. There is also a hosted Hugging Face Space demo, and enterprise access via the Cohere API and Model Vault for private deployments.

Cohere · 2026-07-07 · major

Cohere Transcribe Arabic — 2B open-weight ASR beats Whisper on dialect and code-switching

Cohere Transcribe Arabic is a 2B Conformer encoder-decoder speech model, Apache-2.0. It posts 25.87 average WER across five Arabic dialect benchmarks, beating OmniASR 7B (28.32) and Whisper Large v3 (36.86), and runs 3.6x faster on H100.

Cohere Transcribe Arabic announcement card

The open-source Arabic speech model that finally beats Whisper on dialect audio.

Key specs

Parameters	2B
Avg wer	25.87
Human preference vs whisper	95.8%
Rtfx (h100)	525

Quick facts

Maker	Cohere Labs
Parameters	2B Conformer encoder-decoder
License	Apache-2.0
Languages	Arabic (MSA + dialects) and English
Availability	Hugging Face weights + Cohere API + Model Vault
Best WER	25.87 avg vs Whisper Large v3 36.86

Benchmarks

Average WER across Arabic ASR benchmarks (Common Voice, SADAWER, MASC, MGB-2, Casablanca)

Cohere Transcribe Arabic 07-2026		25.87WER
OmniASR LLM 7B		28.32WER
Cohere Transcribe 03-2026		30.67WER
Whisper Large v3		36.86WER

source ↗

What is it?

Cohere Transcribe Arabic is a 2B-parameter Conformer speech recognition model tuned for Arabic dialects and Arabic-English code-switching. Cohere Labs released the weights under Apache-2.0 today alongside a hosted API and Hugging Face demo. It targets enterprise call-center and media workloads where Modern Standard Arabic and regional variants coexist.

How does it work?

A large Conformer encoder feeds a lightweight Transformer decoder that Cohere Labs trained specifically on Arabic dialect audio plus English. The compact 2B footprint pushes throughput to 525 RTFx on a single H100, about 3.6x Whisper Large v3, and vLLM and native Transformers are both supported for serving.

Why does it matter?

Arabic ASR is a long-tail gap: Whisper Large v3 sits around 36.86 average WER on the dialect benchmarks, which is too high for search, subtitling or call analytics. Dropping that to 25.87 with open weights lets teams self-host Arabic transcription instead of paying per minute to closed vendors, and the Apache-2.0 license clears the way for on-prem regulated deployments.

Who is it for?

teams building Arabic ASR pipelines, dialect researchers, and media / call-center engineers

Frequently asked questions

How much does Cohere Transcribe Arabic cost?: Cohere Transcribe Arabic is Apache-2.0 open weights, so running it yourself is free of licensing cost. Cohere also offers hosted API access with standard rate limits and a Model Vault deployment option; both use Cohere's usual pay-as-you-go pricing rather than a bespoke rate for this model.
Which Arabic dialects does Cohere Transcribe Arabic handle?: Cohere Transcribe Arabic covers Modern Standard Arabic plus regional dialects, and is tuned for Arabic-English code-switching that comes up in enterprise call flows. Its 25.87 average WER is measured across Common Voice, SADAWER, MASC, MGB-2 and Casablanca, which together sample MSA and a wide dialect mix.
How does Cohere Transcribe Arabic compare to Whisper Large v3?: Cohere Transcribe Arabic cuts word error rate from Whisper Large v3's 36.86 to 25.87 on the same Arabic dialect benchmarks and reaches 525 RTFx on H100 versus Whisper Large v3's 146 — about 3.6x higher throughput. Human evaluators also preferred Cohere Transcribe Arabic in 95.8% of paired comparisons.
What is Cohere Transcribe Arabic not good at?: Cohere Transcribe Arabic focuses on single-language audio, so it can slip on heavy code-switched clips, and it ships with no automatic language detection, no timestamps and no speaker diarization. It can also transcribe non-speech noise without a VAD front end, so production stacks should add voice activity detection.
Where can I try Cohere Transcribe Arabic?: Cohere Transcribe Arabic weights are on Hugging Face at CohereLabs/cohere-transcribe-arabic-07-2026, with native Transformers and vLLM inference paths plus Colab and Kaggle notebooks. There is also a hosted Hugging Face Space demo, and enterprise access via the Cohere API and Model Vault for private deployments.

Try it

huggingface.co/CohereLabs/cohere-transcribe-arabic-07-2026