--- language: - ar license: apache-2.0 base_model: openbmb/VoxCPM2 tags: - text-to-speech - tts - arabic - najdi - saudi-arabic - voice-cloning - voxcpm2 - fasee7 pipeline_tag: text-to-speech model_name: Fasee7 Najdi Small --- # Fasee7 Najdi Small — فصيح نجدي ## Model Summary **Fasee7 Najdi Small** is the first public release of the Fasee7 model family, developed by [Wittify.ai](https://wittify.ai). It is an open Arabic Text-to-Speech (TTS) model that generates natural-sounding **Najdi (Saudi) dialect** speech, trained on a high-quality in-house Najdi Arabic dataset. Unlike many Arabic TTS systems that focus exclusively on Modern Standard Arabic (MSA), Fasee7 Najdi Small is designed and optimized for real **conversational Najdi speech**. The model supports Arabic text with diacritics (تشكيل) to improve pronunciation accuracy and naturalness. --- ## Model Details | | | |---|---| | **Model Name** | Fasee7 Najdi Small | | **Developer** | [Wittify.ai](https://wittify.ai) | | **Task** | Text-to-Speech (TTS) | | **Language** | Arabic Dialects | | **Dialect** | Najdi Arabic (Saudi Arabia) | | **Architecture** | Based on the Chatterbox Multilingual TTS architecture, implemented via [VoxCPM2](https://huggingface.co/openbmb/VoxCPM2) + LoRA adapter | | **Base Model** | `openbmb/VoxCPM2` (2B parameters) | | **Fine-tuning** | LoRA | | **Training Data** | High-quality in-house Najdi Arabic dataset | --- ## Audio Samples ### Sample 1 ### Sample 2 --- ## Usage Try the model in Google Colab — no local setup required: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1nrXZ9bn4vhnrEZJEDfKkFZset8Yqdphp?usp=sharing) The notebook downloads the base model and LoRA adapter, runs Najdi voice-cloning inference, and lets you listen to and download the generated audio. --- ## Known Limitations ### Repetition In some cases the model may repeat words or phrases. To reduce this: - Increase `--inference_timesteps` (e.g. 20–30) for more stable output - Adjust `--cfg_value` (recommended `2.0`; lower values allow more variation, higher values follow conditioning more strictly) ### Dialect Coverage This release is fine-tuned for **Najdi Arabic only**. Other dialects or heavily MSA-style text may sound unnatural or inconsistent. ### Voice Cloning Quality Output quality depends on the reference audio. For best results: - Use a clean Najdi reference clip (3–10 seconds) - Provide an **exact transcript** of the reference audio - Avoid noisy, clipped, or heavily compressed files ### Mixed Text Arabic/English mixed text, numbers, abbreviations, and unusual spellings may produce inconsistent pronunciation. ### Catastrophic Forgetting The base VoxCPM2 model was pre-trained on multilingual data. This fine-tune was trained exclusively on Arabic dialect data, with no multilingual data included in the training mix. As a result, the model may suffer from **catastrophic forgetting** — its ability to synthesize speech in languages other than Arabic has likely degraded significantly compared to the base model. --- ## Intended Use **Suitable for:** - Najdi Arabic text-to-speech synthesis - Voice cloning with a Najdi reference speaker - Research and prototyping for Saudi Arabic voice applications **Not suitable for:** - Impersonation, fraud, or any use without voice-owner consent - Safety-critical applications without human review --- ## License Released under the [Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0) license, consistent with the base VoxCPM2 model. --- ## Acknowledgements Built on [VoxCPM2](https://huggingface.co/openbmb/VoxCPM2) by [OpenBMB](https://github.com/OpenBMB). ---