> ## Content Index
> Fetch the complete content index at: https://globalfeed.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Mistral releases an open Voxtral model for real-time Arabic speech transcription
- URL: https://globalfeed.ai/en/mistral-releases-an-open-voxtral-model-for-real-time-arabic-speech-transcription/
- Published: 2026-10-10T18:55:30.000Z
- Updated: 2026-10-10T18:55:30.000Z
- Description: Mistral AI released Voxtral Mini 4B Realtime Arabic, a model that transcribes Arabic and its dialects in real time, under the Apache 2.0 license. Alongside it comes LIDstral Arabic, which identifies the dialect or language of Arabic-script text. Both were built with Morocco's digital transition mini
- Author: GlobalFeed Editor
- Tags: Mistral, Voxtral, speech recognition, Arabic, x-MistralAI, dil-en, elle

**Mistral AI** released two Arabic models on Hugging Face on October 8, 2026\. **Voxtral Mini 4B Realtime Arabic** is a speech recognition model that transcribes Modern Standard Arabic and Arabic dialects while people are still speaking, for uses such as live captions and voice interfaces. **LIDstral Arabic** is a fast classifier that tells whether a text written in Arabic script is Modern Standard Arabic, which dialect it is, or another language that uses the same script. Both are released under Apache 2.0 with open weights, so organizations can run them on their own servers.

According to the model cards, both models were developed with Morocco's Ministry of Digital Transition and Administrative Reform (MTNRA) as part of the strategic partnership Mistral and Morocco signed in January 2026\. Mistral describes them as a "sovereign AI asset" under that partnership.

## Transcription while the speaker talks

Voxtral Mini 4B Realtime Arabic has about 4.4 billion parameters and was fine-tuned for Arabic from Mistral's multilingual Voxtral Mini 4B Realtime. It streams text as the audio arrives, and its transcription delay can be configured for different uses. Training put particular weight on code-switching, where speakers move between Arabic and another language mid-sentence. In Mistral's measurements, the model reaches an average character error rate of 8.82 percent across seven Arabic benchmarks at a 480-millisecond delay, 0.91 points behind the 7.91 percent of the company's Voxtral Transcribe Arabic model. Alongside FLEURS, the benchmarks include two new sets, Darija in the Wild, which covers Moroccan Arabic, and ISMA.

The model can be served with vLLM, with audio streamed over a WebSocket to a realtime endpoint, and it also runs with the Transformers library. The model card does not give a GPU memory figure. The model is built for Arabic only; for other languages Mistral points to the original Voxtral Mini 4B Realtime.

## A classifier for the dialect

LIDstral Arabic picks one of 51 classes: Modern Standard Arabic, country dialects such as those of Morocco, Algeria, Mauritania, Libya, Iraq and Egypt, and other languages written in Arabic script. It combines a fastText classifier with a small neural network and a calibration step, and it runs on a CPU with no GPU needed. In Mistral's evaluation of 84,870 examples, it scores an F1 of 88.67 percent on Moroccan Darija, against 72.65 percent for LahjatBERT and 71.28 percent for GlotLID v3 on the same test.

The speech model's error rates were computed with a text normalization method developed together with MTNRA, and the model card does not give a detailed per-dialect breakdown. Because LIDstral always assigns one of its 51 classes, it can give wrong results for unsupported languages or mixed-language text. For a fuller picture, follow evaluations from independent researchers as well.

Sources: [Hugging Face: Voxtral Mini 4B Realtime Arabic model card](https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-Arabic?ref=globalfeed.ai), [Hugging Face: LIDstral Arabic model card](https://huggingface.co/mistralai/LIDstral-Arabic?ref=globalfeed.ai).