mqqz/csm-1b-arabic

CSM 1B Arabic Finetune

This model is a full weight finetune of sesame/csm-1b on ~1000 hours of Arabic clips from various sources.

CSM (Conversational Speech Model) is a speech generation model from Sesame that generates RVQ audio codes from text and audio inputs. The model architecture employs a Llama backbone and a smaller audio decoder that produces Mimi audio codes.

Audio examples

Short Modern Standard Arabic examples generated with this checkpoint, without a reference voice.

Example 1

أهلاً بك، يسعدني أن أساعدك في إكمال طلبك.

Example 2

يساعدنا التعاون على تجاوز الصعوبات وتحقيق أهدافنا المشتركة.

Example 3

في الصباح، فتحت النافذة ودخل الهواء العليل إلى الغرفة.

Data

Source	Data used	~ hrs.
pain/MASC (https://huggingface.co/datasets/pain/MASC)	Broad Arabic speech	800
khaledalganem/sada2022 (https://huggingface.co/datasets/khaledalganem/sada2022)	Khaliji-labelled segments only	10
MBZUAI/ClArTTS (https://huggingface.co/datasets/MBZUAI/ClArTTS)	Classical Arabic TTS	10
google/fleurs (https://huggingface.co/datasets/google/fleurs)	Arabic configuration ar_eg	7
Common Voice Arabic 27.0	Arabic speech from supplied archive	120
Tunisian MSA — OpenSLR 46 (https://www.openslr.org/46/)	Modern Standard Arabic speech collected in Tunisia	10
MediaSpeech — OpenSLR 108 (https://www.openslr.org/108/)	Arabic subset	10
Arabic Speech Corpus (ASC) (https://en.arabicspeechcorpus.com/)	Arabic read speech	4

Audio was standardized to mono at 24 kHz, with basic transcript normalization. We removed unusable and silent recordings, restricted clips to 0.5–20 seconds, and trimmed leading and trailing silence using voice activity detection. The corpus combines broad Arabic speech with Khaliji-labelled segments from SADA. Training, validation, and test splits were kept separate, with checks for duplicate audio across splits. Audio tokens were precomputed for efficient training.

Training

The model was trained (full-weight finetune) from the original Sesame published 1B checkpoint for 5 epochs on Nvidia B200s. Hyperparameters were selected after an initial Optuna sweep.

Previous Sesame finetunes on other languages and our empirical tests showed better results with a full-weight finetune as opposed to LoRA.

Limitations

A lot can be improved. More training/data is the obvious one. Cleaner data with manual review would've helped.

See Also

• sesame/csm-1b Original model checkpoint

• MAdel121/Seasmed-Fine-Tuned-on-Common-Voice-17-Arabic Previous attempt for an Arabic finetune

  It probably used LoRA on a smaller dataset hence the results were noisy and it wasn't able to properly learn Arabic.