Voxtral-Mini-4B-Realtime-Arabic
by mistralai
Mistral's streaming speech-to-text for Modern Standard Arabic and 15 dialects, 8.82% average CER at 480 ms delay
mistralai/Voxtral-Mini-4B-Realtime-ArabicDeploy Voxtral-Mini-4B-Realtime-Arabic
Single-tenantMixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.
Overview
Voxtral Mini 4B Realtime Arabic is a streaming speech-to-text model from Mistral AI, released in October 2026 under the Apache 2.0 license. It is fine-tuned from Voxtral Mini 4B Realtime for Modern Standard Arabic and dialects from the Maghreb, the Gulf, the Nile Valley and the Levant, and for code-switching, where a speaker moves between Arabic and another language mid-sentence. It produces text as the audio arrives.\n\nAt a transcription delay of 480 ms, the card reports an average character error rate of 8.82% across seven Arabic benchmarks, 0.91 points behind Mistral's offline Voxtral Transcribe Arabic at 7.91%. Two of the benchmarks, ISMA and Darija in the Wild, were built with Morocco's digital transition ministry, and scores use the card's own Arabic normalization. For other languages, the card points to the base model.
Architecture
About 4.4B parameters in BF16: a causal audio encoder reading 16 kHz audio and a language decoder that emits text while audio streams in, with a configurable transcription delay. Fine-tuned from Voxtral-Mini-4B-Realtime-2602. Served through vLLM's /v1/realtime WebSocket endpoint or Transformers 5.2+ (VoxtralRealtimeForConditionalGeneration).
Mixpeek SDK Integration
# Voxtral Arabic runs on your side. Transcribe each segment and store the
# text next to the audio, so a retriever can search what was said.
import requests
for seg in segments: # e.g. {"key": "calls/2026-10-13/0412.wav", "start": 24.0, "text": transcript}
requests.post(
"https://api.mixpeek.com/v1/buckets/bkt_your_bucket/objects",
headers={"Authorization": "Bearer API_KEY", "X-Namespace": "ns_your_namespace"},
json={"blobs": [{"property": "audio", "type": "audio", "data": f"s3://your-bucket/{seg['key']}"}],
"metadata": {"start": seg["start"], "transcript_ar": seg["text"], "language": "ar"}},
)Capabilities
- Streaming transcription with a configurable delay
- Modern Standard Arabic and 15 regional dialects
- Arabic code-switching within one conversation
- vLLM realtime WebSocket serving and Transformers support
- Apache 2.0 license
Use Cases on Mixpeek
Benchmarks
| Dataset | Metric | Score | Source |
|---|---|---|---|
| Seven Arabic benchmarks (average) | CER | 8.82% | Model card: mistralai/Voxtral-Mini-4B-Realtime-Arabic (self-reported, 480 ms delay, the card's Arabic normalization) |
| Same benchmarks, Voxtral Transcribe Arabic (offline) | CER | 7.91% | Model card (self-reported comparison) |
Performance
Scores use normalization co-developed with Morocco's MTNRA, covering dialect spelling, clitics, loanwords and spoken numbers; the card says scores may differ under other normalization. We have not measured it.
Common Pipeline Companions
Frequently Asked Questions
What is Voxtral Mini 4B Realtime Arabic?
An open streaming speech-to-text model from Mistral AI for Modern Standard Arabic and regional dialects. It is a 4.4B-parameter fine-tune of Voxtral Mini 4B Realtime, released under Apache 2.0.
Which Arabic dialects does it cover?
The card lists Modern Standard Arabic plus Moroccan, Libyan, Tunisian, Algerian, Hassaniya, Gulf, Najdi, Omani, Sanaani, Egyptian, Sudanese, Mesopotamian, South Levantine, North Levantine and Chadian Arabic.
How accurate is it?
The card reports an 8.82% average character error rate across seven Arabic benchmarks at a 480 ms delay, against 7.91% for Mistral's offline Voxtral Transcribe Arabic. The scores use the card's own Arabic normalization, so test on your own recordings.
Can it transcribe English or French?
It handles Arabic speakers switching into other languages mid-conversation. For audio that is mainly in another language, the card recommends the original Voxtral Mini 4B Realtime.
Specification
Research Paper
Voxtral Mini 4B Realtime Arabic model card
arxiv.orgBuild a pipeline with Voxtral-Mini-4B-Realtime-Arabic
Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.
Run it on your own data, free