--- title: Bayan Arabic TTS emoji: ๐ŸŽ™๏ธ colorFrom: indigo colorTo: red sdk: gradio sdk_version: 6.31.0 app_file: app.py python_version: "3.12" startup_duration_timeout: 1h short_description: Arabic TTS with the fine-tuned Bayan "Murad" voice pinned: false license: other license_name: nvidia-open-model-license license_link: https://huggingface.co/nvidia/magpie_tts_multilingual_357m models: - anwar-dev/Bayan-Arabic-TTS-v1.0 - nvidia/magpie_tts_multilingual_357m --- # ๐ŸŽ™๏ธ Bayan Arabic TTS (ุจูŠุงู†) Arabic text-to-speech with a single male voice โ€” **ยซู…ุฑุงุฏยป / Murad** โ€” built on [`nvidia/magpie_tts_multilingual_357m`](https://huggingface.co/nvidia/magpie_tts_multilingual_357m) with LoRA fine-tuning and a dedicated baked speaker row. - ๐Ÿค— **Model weights:** [anwar-dev/Bayan-Arabic-TTS-v1.0](https://huggingface.co/anwar-dev/Bayan-Arabic-TTS-v1.0) - ๐Ÿค— **Base model:** [nvidia/magpie_tts_multilingual_357m](https://huggingface.co/nvidia/magpie_tts_multilingual_357m) - ๐ŸŽ›๏ธ **Codec:** [NVIDIA NanoCodec 22.05 kHz](https://huggingface.co/nvidia/nemo-nano-codec-22khz-1.89kbps-21.5fps) Type Arabic text, press **ุชูˆู„ูŠุฏ ุงู„ุตูˆุช ยท Generate speech**, and get 22.05 kHz mono audio back. Use the **Voice** dropdown to compare the fine-tuned Murad voice (the default) against the five original MagpieTTS voices. ## Tips for good results - **Diacritics matter.** Training used fully diacritized (ู…ุดูƒู‘ู„ุฉ) text. Undiacritized text still works, but accuracy is much better with the diacritics โ€” especially on word endings. - Keep utterances under **~37 seconds**; split longer passages. - Text normalization is **off** (matching the authors' reference settings), so write numbers and symbols out as words. - Stuttering? Lower **Temperature** toward `0.5` or raise **CFG scale** toward `3.0` in the advanced settings. ## Implementation notes This Space loads the author's merged, self-contained `Bayan-Arabic-TTS-v1.0.nemo` directly through NeMo's `load_magpie_model` (no separate `load_lora_weights` step), with `speaker_index=5`, `language="ar-MSA"`, `cfg_scale=2.5`, `temperature=0.6`, `topk=80`. The example prompts are the author's own diacritized MSA sentences from `samples/msa/texts.txt` in the model repo. ## License Derived from `nvidia/magpie_tts_multilingual_357m` and governed by the [NVIDIA Open Model License](https://huggingface.co/nvidia/magpie_tts_multilingual_357m). See the model repo's `NOTICE.md` for commercial-use terms.