Introducing fihrIA-ASR: SOTA open-weight Darija speech recognition, better and 7 times faster than Voxtral
fihrIA-ASR is our lab’s first product release: a pair of open-weight models (0.6B and 1.7B) that transcribe Moroccan Darija. The 1.7B model makes slightly fewer meaning errors than Voxtral Mini 4B, and the 0.6B model slightly more. Both make clearly fewer than MoulSot, an open Darija model built on the same Qwen3-ASR base. Both run about 7 times faster than Voxtral, and the small one runs on a laptop CPU.
Introduction
Tens of millions of people speak Darija, but there is no standard way to write it. Speakers move between Darija, French, and English within one sentence. The same word can appear in Arabic or in Latin script, attached to its neighbour or apart, with or without vowel marks.
That causes two problems for speech recognition. There is no large set of transcribed Darija speech to train on. And there is no reliable way to score a model: the usual metric counts every spelling difference as a mistake. At fihrIA, we had to solve both before we could build and evaluate fihrIA-ASR. This product blog tells that story in order. Each section links to a technical page with the full details.
Data collection
We collected 213,089 public videos from 460 Moroccan channels: news, talk shows, cooking, religion, comedy, regional outlets. We cut the audio into clips of at most 30 seconds, and Gemini transcribed about 10,000 hours of them. After cleaning, 1.51 million clips remain.
The decision that mattered most was how to write Darija. We gave Gemini two rules: Darija words in Arabic script, and French or English words in Latin script, as the speaker says them (هادي هي la discipline). With these rules, the model learns one consistent style.
The labels come from a machine, not from people. At this scale there was no practical alternative, but it means that our models learn Gemini’s habits, including its mistakes.
The training corpus stays private. The held-out test set—the 508 clips and the expert references—will be released as a public benchmark.
Tokenizer
We started from Qwen3-ASR, an open speech recognition backbone. It writes a transcript one token at a time, where a token is a word or a piece of a word. Its vocabulary is shared across many languages, and Darija is not one of the main ones, so it breaks Darija words into small fragments:
Original Qwen tokenizer (151,936 tokens): 18 pieces
Darija BPE (89,000 tokens): 9 pieces
On all our training transcripts, the Qwen vocabulary needs an average of 2.29 tokens for each word. The Darija vocabulary needs 1.32 tokens (42% fewer). Each token needs one decoder step.
Each token costs one step of the model, so fragments make transcription slow. For fihrIA-ASR, we replaced the vocabulary with one trained on Darija text, which needs 42% fewer tokens for the same text, and then trained the model on our clips. The result is 1.65 times faster than the same model with the original vocabulary. Accuracy did not change: a control model that keeps Qwen’s vocabulary has the same WER within 0.3 points and the same meaning-error rate within 0.1 points.
Evaluation
The usual metric is the word error rate (WER): the share of words that differ from a reference transcript written by a person. fihrIA-ASR 1.7B has a WER of 31%, which suggests that one word in three is wrong. Native speakers do not see anywhere near that many errors. Because Darija has no standard spelling, a correct transcript can differ from the reference in almost every word. Switch between the views below to see which words each method marks.
The hypothesis writes some attached words as separate words (فداركم is written في داركم). It also removes some hamzas. The meaning does not change.
Human SER
0%
LLM SER
3.8%
WER
82.3%
CER
17.3%
Words that WER counts as errors
Reference
وحنا محتاجين مرا متعلمة، ولكن التعلم ماشي معناه هو الاختلاط. اليوم راه يمكن ليك تقرا فداركم فستانفورد، وكامبريدج، وإم آي تي، وأوبن كورس وير، وإيديكس، وكورسيرا، وݣاع المنصات العالمية بأرقى الشواهد وأنت جالس فدارك.
Hypothesis · Gemma-4-E4B
و حنا محتاجين مرا متعلمة ولكن التعلم ماشي المعنى هو الاختلاط اليوم راه يمكن ليك تقرا في داركم في ستانفورد و كامبريدج و ام آي تي و اوبن كورس وير و ايديكس و كورسيرا وكاع من المنصات العالمية بارقى الشواهد و نتا جالس في دارك
These pairs come from our study with native speakers. The reference is the human transcript. The hypothesis is the transcript of an ASR system. The human and LLM scores are SER. Select a view to see the words that each method marks as errors.
In the first example, the annotator found no errors. In the last, the annotator found errors in 60% of the words. WER gives the two transcripts almost the same score, 82% and 81%. A metric that cannot separate these two cases cannot tell us which of two models is better.
So we ask a different question. We give a large language model the reference, the transcript, and written instructions, and it sorts every difference into one of two groups. A variation leaves the meaning intact: spelling, script, spacing, or a number written as digits. An error is a missing, added, or replaced word that changes what was said. The semantic error rate (SER) is the share of error words.
We did not take the judge on trust. 7 native Darija speakers marked the errors in 2,006 transcripts by hand, with the same instructions, and we compared each metric with their marks:
The judge is not perfect. It is a little strict, on average 1.9 points above the native speakers, and its absolute values depend on which model is the judge. So we report the mean of four judges from four companies, and we report WER and CER next to it.
Results
Every system transcribed the same 508 test clips, about 2 hours of audio from videos that are not in the training data. The closest comparisons are two open-weight models: MoulSot, a Darija model from AtlasIA with the same base model and the same size as fihrIA-ASR 1.7B, and Voxtral Mini 4B Realtime Arabic, the prior open-weight state of the art on Darija, which is larger than either of ours.
Swipe horizontally to compare all four systems →
| MoulSot1.7B · Qwen tokenizer | VoxtralMini 4B · Realtime Arabic | fihrIA-ASR 1.7B1.7B · Darija tokenizer | fihrIA-ASR 0.6B0.6B · Darija tokenizer | |
|---|---|---|---|---|
| Parameters | 1.7B | 4B | 1.7B | 0.6B |
| Meaning errors (mean SER) | 11.2% | 8.4% | 6.4% | 7.6% |
| Median WER | 42.3% | 36.5% | 30.9% | 31.6% |
| Median CER | 13.0% | 10.6% | 10.3% | 10.7% |
| H100, × real time | 147× | 34× | 243× | 255× |
fihrIA-ASR 1.7B
1.7×fewer meaning errors than MoulSot, at the same size
1.3×fewer meaning errors than Voxtral
1.65×faster than MoulSot
7.1×faster than Voxtral on one H100
fihrIA-ASR 0.6B
1.5×fewer meaning errors than MoulSot, with 2.8× fewer parameters
1.1×fewer meaning errors than Voxtral
2.0×faster than MoulSot
7.5×faster than Voxtral on one H100
Meaning errors under each judge (SER %, lower is better)
Swipe horizontally for every judge →
| Judge | fihrIA 1.7B | fihrIA 0.6B | MoulSot | Voxtral |
|---|---|---|---|---|
| GPT-5.6 Luna | 5.8 | 6.9 | 10.2 | 7.5 |
| Gemini 3.8 Flash | 7.1 | 8.2 | 11.9 | 9.8 |
| Claude Sonnet 5 | 7.0 | 8.0 | 13.4 | 9.7 |
| GLM-5.3 | 5.8 | 7.2 | 9.2 | 6.5 |
MoulSot is the other Qwen3-ASR 1.7B fine-tune for Darija. Voxtral Mini 4B Realtime Arabic is a separate open-weight model. All four were tested on the same 508 clips. Meaning errors and the multiples are the mean of the four judges' SER. The strip below shows each judge, and they do not all rank Voxtral the same way. H100 speed is batch size 32. MoulSot's H100 figure is the control with MoulSot's architecture and tokenizer, and that is also the "faster than MoulSot" ratio for our 1.7B model. For our 0.6B model, "faster than MoulSot" is the laptop comparison, both models in 4-bit.
On the mean of the four judges, fihrIA-ASR 1.7B makes meaning errors on 6.4% of words, Voxtral on 8.4%, fihrIA-ASR 0.6B on 7.6%, and MoulSot on 11.2%. On that mean, the only system ahead of fihrIA-ASR 1.7B is Gemini, a closed commercial system, at 5.0%.
The gaps to Voxtral are small, and the judges do not all agree on them. Two judges put both of our models ahead of Voxtral, one puts Voxtral between them, and one puts Voxtral ahead of both. The gap to MoulSot is larger, and it holds under every judge.
The difference in cost is not small. On one H100, fihrIA-ASR 0.6B and 1.7B transcribe at 255× and 243× real time, against 34× for Voxtral and 147× for a model with MoulSot’s architecture.
DarijaCS benchmark
Our test set and our judges could favour our own models, so we looked for numbers we did not measure. DarijaCS is a Darija benchmark published on 10 November 2026 by Morocco’s Ministry of Artificial Intelligence, Digital Transformation and Administration Reform. It has two test sets, and Voxtral publishes its character error on both: 6.68% on Prism (a.k.a. ISMA) and 9.69% on Probe (a.k.a. Darija in the Wild). We scored fihrIA-ASR on the same clips with the same protocol.
Prism (a.k.a. ISMA)
- fihrIA-ASR 1.7B6.05%
- Voxtral Mini 4B, as published6.68%
- fihrIA-ASR 0.6B7.90%
- MoulSot10.27%
Probe (a.k.a. Darija in the Wild)
- fihrIA-ASR 1.7B8.52%
- Voxtral Mini 4B, as published9.69%
- fihrIA-ASR 0.6B9.79%
- MoulSot14.93%
Voxtral’s bars use the figures from its own chart. The other bars are our measurements at 16-bit precision, on the same clips and with the same protocol as DarijaCS.
The order is the one we found ourselves: fihrIA-ASR 1.7B slightly ahead of Voxtral, fihrIA-ASR 0.6B slightly behind on one set and level on the other, and both well ahead of MoulSot.
Local inference
fihrIA-ASR 0.6B does not need a GPU. On disk it is 721 MB. Through llama.cpp on an ordinary laptop it runs about 6 times faster than real time: one minute of audio takes about 10 seconds.
Every claim in this product blog has a technical page behind it, and some of those pages go one level further. The map of all the details shows them in the order of the article.