Meet Cohere Transcribe Arabic

Today, we are releasing Cohere Transcribe Arabic as an open-source model. Based on our 2B frontier Automatic Speech Recognition (ASR) model released earlier this year, Cohere Transcribe Arabic is built for the realities of Arabic in business and developer settings: dialect variation, and bilingual Arabic-English speech and code-switching.

It is the most accurate, open-source, Arabic speech-to-text model to date, outperforming leading alternatives, including Whisper v3 Large and OmniASR LLM 7B, across dialects and common speech patterns. It also delivers substantial gains over Cohere Transcribe on both Arabic and bilingual Arabic-English audio.

Cohere Transcribe Arabic is available under the Apache 2.0 license. Visit the model card to download the weights and read our quickstart implementations, or access the hosted model through the Cohere API or Model Vault.

Architecture

Cohere Transcribe Arabic builds on the same 2B-parameter encoder-decoder architecture described in our Cohere Transcribe release: a FastConformer encoder handling acoustic modelling, paired with a lightweight autoregressive Transformer decoder.

Training Data

Our training data is a mix of Arabic, Arabic-accented English and L1 English datasets. It emphasises three dimensions:

• Dialect diversity: among the Arabic samples we carefully resample across dialect groups to prevent any single variety from dominating and ensure the model generalizes broadly.
• Arabic-English Code-switching: real bilingual speech where Arabic and English are mixed naturally, reflecting how Arabic professionals actually speak in business contexts.
• Acoustic variety: speech recorded across different devices, environments, and noise conditions, augmented with TTS-generated samples for underrepresented conditions.

Results

Cohere Transcribe Arabic achieves the lowest average word error rate of any open-source model on the Hugging Face Arabic ASR Leaderboard, with a WER of 25.87. This is a 2.45-point improvement over the previous leader, Meta’s OmniASR-LLM-7B, and an 11-point improvement over OpenAI’s Whisper v3 Large.

Table 1: the Open Universal Arabic ASR leaderboard as of 23.06.26. This public benchmark evaluates zero-shot multi-dialect generalization across six test sets spanning MSA, Egyptian, Gulf, Levantine, and Maghrebi dialects. See the latest leaderboard and details on the benchmark methodology.

The gains are broad-based. Cohere Transcribe Arabic delivers the best overall WER and ranks first on four of the six composite task sets. On Casablanca, for example, which evaluates conversational Arabic across eight dialects, it improves on OmniASR by nearly six points. On Common Voice, a crowd-sourced dataset covering 25 dialects, it reduces WER by more than two points from a low previous base.

Internally, we have evaluated the model on test sets containing speech in specific Arabic dialects. Arabic dialect handling is a clear gap across the field (learn more). Existing models flatten dialectal speech into Modern Standard Arabic, struggle with Arabic-English code-switching, and underserve the dialects that hundreds of millions of people actually speak. Cohere Transcribe Arabic is our contribution to solving this problem.

Table 2: WER by Arabic dialect and code-switching following Cohere internal evaluations. Dialectic and code-switched speech were grouped across various open-source and internal benchmarks to measure performance gains in each category.

The performance carried through to human evaluations with native Arabic speakers. Evaluators assessed transcription quality across three dimensions:

• Overall accuracy: How well the transcript captured both the form and semantic meaning of the audio.
• Dialect faithfulness: Whether the model preserved the speaker’s dialect rather than converging toward Modern Standard Arabic
• Robustness to code-switching: Whether English terms were correctly transcribed in Latin script.

Cohere Transcribe Arabic scored highest on all three dimensions compared with Whisper v3 Large and Cohere Transcribe. In head-to-head evaluations, it was preferred over Whisper in 95.8% of tests.

Plot 1: human preference evaluations of Arabic transcripts, according to overall quality, faithfulness to the dialect used, and ability to handle code-switching elements. Scores were rated on a scale of 1-5, with 5 marking the highest quality.

The model also improved significantly on Cohere Transcribe for English spoken with an Arabic accent, covering many workplace and second-language English use cases. Human evaluators preferred Cohere Transcribe Arabic over Cohere Transcribe in 77.2% of tests, and found it broadly comparable to Whisper on these inputs, with Cohere Transcribe Arabic preferred in 52.6% of tests.

Plot 2: human preference scores of Arabic transcripts in pairwise comparisons. A score of 50% or higher indicates that outputs from Cohere Transcribe Arabic were considered more accurate on average.

Normalization

We normalize hypotheses and references using an Arabic ASR normalizer developed internally and adapted from the text normalization utilities in OpenAI's Whisper. The normalizer applies full orthographic canonicalization: diacritics (harakat) removal, hamza/alef variant collapsing (آأإٱ→ا, ؤ→و, ئ→ي, ء deleted), alif maqsura folding (ى→ي), Persian/Urdu letter folding (ک→ك, ی→ي, پ→ب, ڤ→ف), Eastern Arabic numeral transliteration (٠–٩ → 0–9), tatweel removal, punctuation normalization, and Unicode NFKC cleanup. As a result, orthographic variation that is irrelevant to transcription quality is collapsed.

Efficiency

Cohere Transcribe Arabic is built for high-throughput serving in production environments, where performance under concurrent demand matters as much as model quality.

We optimized the system around vLLM to handle high-volume speech workloads, even when audio inputs vary in length. Further updates to the runtime and model stack helped yield up to 2x higher throughput.

Overall, Cohere Transcribe Arabic achieves an RTFx (real-time factor multiple) score of 525 versus 146 for Whisper v3 Large and 66 for OmniASR-LLM-7B.

Image 5: throughput (RTFx) vs accuracy (WER) plot for three leading models. RTFx (real-time factor multiple) measures how fast an audio model processes its input relative to real time. Source: Open ASR Leaderboard.

Limitations & Future Work

This model has some known limitations:

• Cohere Transcribe Arabic is trained to expect a language tag (Arabic or English). Code-switching performance is optimised for the selected language tag being the matrix language, and the other being the embedded language.
• Like many AED speech models, Cohere-transcribe is eager to transcribe, even non-speech sounds. The model thus benefits from prepending a noise gate or VAD (voice activity detection) model in order to prevent low-volume, floor noise from turning into hallucinations.
• Some common ASR model features - namely, output timestamps and speaker diarization - are not supported by this model.

Conclusion

We are thrilled to share Cohere Transcribe Arabic with the community and hope it inspires more open work on localisation of AI technology.

Model weights are available today on Hugging Face. You can test the model beforehand in our Hugging Face Space.

Cohere Transcribe Arabic is also available through the Cohere API with free access subject to rate limits. Get a production key and refer to the API documentation to get started.

For managed production deployment without rate limits, provision a dedicated Model Vault from your Cohere dashboard. Pricing is calculated per instance-hour, with discounted plans available for longer-term commitments.