sailu4/lmaana-v0

Lmaana v0

Lmaana v0 is a 1B-parameter-class research checkpoint for automatic speech recognition in Moroccan Darija. It combines a Dataset13-adapted OmniASR model with mixed replay over Dataset13 and MoulSot. The objective is to learn the MoulSot domain without losing the speech representations already acquired on Dataset13.

Model Summary

Property	Value
Task	Automatic speech recognition
Language	Moroccan Darija (ary)
Architecture	OmniASR CTC 1b_v2 (wav2vec2_asr)
Framework	fairseq2
Tokenization	Character tokenizer (omniASR_tokenizer_written_v2)
Decoding used for evaluation	Greedy CTC, no external language model
Checkpoint type	Native resumable training checkpoint

Why Mixed Replay?

Adapting the Dataset13 model only on MoulSot improved MoulSot WER from 46.0435% to 43.9947%, but Dataset13 WER deteriorated from 44.1165% to 50.0524%. This is the catastrophic-forgetting problem: specialization on the new corpus damages performance on the original domain.

Lmaana v0 instead replays Dataset13 examples while learning from MoulSot. The result is a general-purpose checkpoint that improves MoulSot while preserving Dataset13 almost exactly.

Training Approach

1. Starting point: OmniASR CTC 1b_v2 after 6,000 Dataset13 adaptation steps.
2. Data construction: Dataset13 and MoulSot were exposed as one mixture-Parquet dataset. The local preparation uses hard links, avoiding a second physical copy of the audio payload.
3. Replay sampling: corpus and language sampling coefficients were both set to 0.5 (beta_corpus and beta_language).
4. Optimization: all encoder parameters remained trainable, with a 5e-7 learning rate and gradient accumulation over 8 batches.
5. Selection: validation and checkpointing ran every 250 steps. The retained checkpoint is step 1,000, selected using validation WER.

Main Training Parameters

Parameter	Value
Training steps	1,000
Learning rate	5e-7
Frozen encoder steps	0
Gradient accumulation	8 batches
Maximum audio length	320,000 elements
Maximum batch size	1,280,000 elements
Example shuffle window	1,000
Validation interval	250 steps
Checkpoint interval	250 steps
Audio normalization	Enabled

Evaluation Protocol

The selected checkpoint was evaluated separately on the held-out Dataset13 and MoulSot test partitions with the same fairseq2 evaluation pipeline. Reported WER and UER use greedy CTC decoding without an external language model. The test partitions were not used for checkpoint selection.

Final Test Results

Test set	CTC loss	UER	WER	Evaluated examples
Dataset13	183.9310	17.7329%	44.1211%	6,576
MoulSot	76.4528	14.8204%	45.4075%	1,960

Comparison With Previous Checkpoints

Checkpoint	Dataset13 WER	MoulSot WER	Interpretation
Dataset13 step 6,000	44.1165%	46.0435%	Strong starting baseline
MoulSot-only step 1,000	50.0524%	43.9947%	Better specialist, substantial forgetting
Lmaana v0	44.1211%	45.4075%	Best general-purpose trade-off

Against the starting checkpoint, Lmaana v0 changes Dataset13 WER by only +0.0046 point while improving MoulSot WER by 0.6360 point. The MoulSot-only checkpoint remains preferable only when MoulSot specialization matters more than cross-domain retention.

Intended Use

• Research and evaluation of Moroccan Darija speech recognition.
• Continued fairseq2/OmniASR training from the native checkpoint.
• Development of Darija transcription systems with domain-specific testing.

This checkpoint has not been validated for high-stakes or fully automated decision-making.

Checkpoint Format

This repository stores a complete native fairseq2/OmniASR training checkpoint, including optimizer and trainer state. It is not a Transformers from_pretrained() export.

from huggingface_hub import snapshot_download

local_path = snapshot_download(repo_id="sailu4/lmaana-v0")
print(local_path)

Load the directory under checkpoint/ with the same OmniASR fairseq2 recipe. The tokenizer, training configuration, evaluation summary, and detailed logs are available under metadata/.

Reproducibility

The exact mixed-replay configuration and evaluation outputs are included in metadata/. manifest.json records every uploaded file, its size, and its SHA-256 digest. These records can be used to verify a restored checkpoint before resuming training or running inference.

Repository Contents

• checkpoint/: resumable native fairseq2 checkpoint.
• metadata/: tokenizer, configuration, metrics, and evaluation logs.
• manifest.json: file sizes and SHA-256 integrity hashes.

Limitations

Performance can vary with recording conditions, speaker demographics, code switching, regional vocabulary, and domains not represented in the evaluation sets. WER values should not be assumed to transfer unchanged to other Darija corpora. The model may produce omissions, substitutions, or plausible but incorrect text. Review transcriptions before using them in consequential workflows. Dataset and deployment licensing must be reviewed separately for the intended use.

Version

This repository contains Lmaana v0, the retained general-purpose baseline for subsequent Lmaana experiments.