Tasks and inputs
Jais-ASR is listed for Automatic speech recognition, Speech → Text. It accepts audio and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
SyncDev model directory · Profile updated Nov 17, 2026
Jais-ASR is IFM's open Arabic speech recognition model on Jais-3-4B, covering MSA and dialects of 19 countries, with code-switching into four languages.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Arabic speech recognition usually breaks on two things: dialect and code-switching. A Moroccan caller drops into French mid-sentence; a Gulf customer mixes in English product names. Jais-ASR, also called Jais Speech, launched alongside Jais 3 by MBZUAI's Institute of Foundation Models, is trained for exactly that mess.
It covers Modern Standard Arabic and dialects from 19 countries, with code-switching into English, French, Spanish and Italian. The weights are Apache-2.0.
Across the six test sets of an open Arabic ASR leaderboard, Jais-ASR posts an average word error rate of 22.85% and a character error rate of 8.37%, the lowest averages in its comparison. The next-best average WER, from Audar-ASR-V1-Turbo, is 23.17%.
Averages hide the spread, so here is the breakdown that matters:
| Test set | Jais-ASR WER | Note |
|---|---|---|
| MASC clean | 16.12 | Best in comparison |
| MASC noisy | 25.44 | Best in comparison |
| MGB-2 | 9.59 | Best in comparison |
| SADA | 30.85 | Audar-ASR-V1-Turbo better at 28.92 |
| Common Voice 18 | 8.56 | Voxtral-Small-24B better at 5.19 |
Some commercial comparison figures were measured by IFM itself. Treat the table as a publisher result.
[IFM-NoTranscript], when the audio contains no speech, rather than hallucinated text.That last one is worth a lot in call-centre pipelines, where silence and hold music are common.
At about 6B parameters (11.4 GB in BF16) and built on Jais-3-4B, it needs a GPU with bfloat16 support and roughly 12 GB of memory. Two routes:
automatic-speech-recognition pipeline, trust_remote_code=True. Batch transcription is supported./v1/audio/transcriptions endpoint.FP8, AWQ, GGUF and MLX builds are published too. No maximum audio length is stated; if a result returns truncated=True, raise max_new_tokens.
The repository appeared on 6 October; the model is dated to the Jais 3 launch on 15 October 2026.
If your audio is Arabic and your callers are not reading from a script, this belongs on your shortlist. Hosted alternatives include Cohere Transcribe Arabic; for multilingual work, see speech recognition models.
To evaluate, sample real recordings per dialect you serve, with at least 30 minutes each, transcribe them by hand, and measure WER per dialect rather than overall. A single average will hide the region where your customers are unhappy.
Compare the model on the work you need to complete.
Jais-ASR is listed for Automatic speech recognition, Speech → Text. It accepts audio and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
The weights are downloadable under Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
salesforce
SOV-Transcribe-1 is Salesforce's open-weight speech-augmented model for non-streaming transcription, with keyword biasing and style instructions; Apache-2.0.
ElevenLabs
Scribe v3 Live is ElevenLabs' streaming transcription model with turn prediction, mid-sentence language switching and a 10,000-term keyterm library.
NVIDIA
Nemotron 3.5 Transcribe 30B A4B is NVIDIA's open offline ASR model: 40 languages, hour-long files, word timestamps, diarisation and translation.
IBM
Granite Speech 5.0 2B is IBM's open speech-language model for multilingual ASR and translation with keyword biasing, runnable from Ollama to MLX.
Mistral AI
Voxtral Mini Transcribe 3 is Mistral's 4B batch transcription model with speaker labels, segment timestamps and custom vocabulary, replacing Mini 2.
Mistral AI
Voxtral Mini Transcribe Realtime 3 is Mistral's streaming speech-to-text model for voice agents: under 500 ms, tunable delay, built-in turn detection.