Lmaana/lmaana-2.4

Lmaana 2.4

Lmaana 2.4 is a Moroccan Darija automatic speech recognition model based on OmniASR CTC 3B v2. It continues Lmaana 2.2 with a controlled mixed-corpus adaptation using Dataset13 clean v4 and corrected MoulSot (Lmaana clean v1).

The selected checkpoint is mixed-adaptation step_4500. It was selected on the separate Dataset13 and MoulSot validation sets before the held-out tests were evaluated.

Model summary

Field	Value
Release	Lmaana 2.4
Base checkpoint	Lmaana 2.2 step_10000
Architecture	OmniASR CTC 3B v2
Objective	CTC full fine-tuning
Target language	Moroccan Darija (ary-Arab)
Adaptation data	80% Dataset13, 20% corrected MoulSot
Adaptation schedule	2,000 steps, selected at step 4,500
Learning rate	2e-6
Tokenizer	OmniASR written tokenizer v2
Decoder	Greedy CTC

Results

All values use the same frozen splits, tokenizer, normalization and greedy CTC decoding protocol as Lmaana 2.2.

Corpus	Split	Examples	CTC loss	UER (%)	CER (%)	WER (%)
Dataset13	Validation	6,893	154.0409	15.4345	Not measured	37.9410
Dataset13	Test	6,576	159.4960	15.6460	16.1553	38.1100
MoulSot	Validation	5,187	65.3884	11.5805	Not measured	35.2742
MoulSot	Test	1,957	65.2453	12.6190	12.5567	39.5332

Improvement over Lmaana 2.2

Held-out test	Lmaana 2.2 WER	Lmaana 2.4 WER	WER gain	CER gain
Dataset13	38.4762	38.1100	0.3662	0.1510
MoulSot	40.7776	39.5332	1.2444	1.3718

Training data

• Dataset13 clean v4: the primary training and evaluation corpus.
• Corrected MoulSot: stored as omniasr_lmaana_clean_v1, with 91.68 hours of training audio. Validation and test sources are held out from its training split.

The mixed sampler targets 80% Dataset13 and 20% corrected MoulSot. MoulSot is also called Lmaana clean v1 in configuration files; these names refer to the same corrected corpus in this release.

Evaluation and selection

The mixed run started from Lmaana 2.2 and used a reinitialized optimizer. The Dataset13 validation WER guard was 38.5618%, and the MoulSot validation target was the Lmaana 2.2 baseline of 37.2417%. Mixed step_4500 passed both guards with 37.9410% and 35.2742%, respectively.

Test CER is computed at corpus level after removing whitespace from references and hypotheses. UER is the tokenizer-unit error rate reported by fairseq2 and is not identical to CER.

Files

Path	Description
model/consolidated.pt	Consolidated fairseq2 checkpoint
tokenizer/omniASR_tokenizer_written_v2.model	Character tokenizer
config/mixed_80_20_train_fsdp.yaml	Mixed adaptation recipe
config/test_mixed_step4500_dataset13.yaml	Dataset13 test recipe
config/valid_mixed_step4500_lmaana_clean.yaml	MoulSot validation recipe
config/test_mixed_step4500_lmaana_clean.yaml	MoulSot test recipe
results/metrics_summary.json	Machine-readable release metrics
SHA256SUMS	Checksum for the consolidated checkpoint

Intended use and limitations

The model is intended for research and evaluation of Moroccan Darija ASR. It uses greedy decoding and may benefit from a language model or lexicon-aware decoder. Performance can vary across speakers, recording conditions, regions, code-switching patterns and writing conventions. Validate the model on the target domain before production use.

This repository contains a native fairseq2 checkpoint, not a Transformers from_pretrained() package.