khalidhabbash/Saudi-tts-v4

Saudi TTS V4

Saudi TTS V4 is a 162.7M-parameter, reference-conditioned F5-TTS/DiT acoustic model adapted from SILMA TTS v1. It produces 24-kHz Arabic speech and supports Najdi-, Hijazi-, and Khaliji-oriented text. A fixed Vocos decoder converts the predicted mel spectrogram to waveform audio.

Source code, preprocessing, training configuration, and validation reports are available at KhalidAlHabbash/saudi-tts-finetune.

This repository contains the inference-only exponential-moving-average (EMA) weights from training update 104,276. Optimizer, scheduler, random-number, and resumable training state are intentionally excluded.

• File: model.safetensors
• Size: 650,704,952 bytes
• SHA-256: 0f89df26517617ae935f8267452f2f196467cb39b80df0af93d13be5ead3d431

Architecture

• Backbone: F5-TTS conditional flow matching with an 18-layer DiT
• Acoustic parameters: 162,666,852
• Hidden size: 768; attention heads: 12
• Mel channels: 100
• Output sample rate: 24 kHz
• Vocoder: charactr/vocos-mel-24khz, pinned to revision 0feb3fdd929bcd6649e0e7c5a688cf7dd012ef21
• Base checkpoint: silma-ai/silma-tts, pinned to revision ac81834c5ce305504fe4c7602187042fdd6913db

Training summary

The model was trained on 104,176 eligible Saudi-labelled clips per complete dataset pass, selected from SADA22 and limited to audio of at most eight seconds. Training used eight completed passes, an effective batch size of eight, AdamW, gradient clipping, activation checkpointing, and EMA tracking. The training profile was CUDA BF16 on one NVIDIA GPU. Validation and test manifests were held out from optimization.

The source is multi-speaker broadcast audio. Consequently, this is a reference-conditioned model rather than a single fixed voice. Output identity, prosody, and quality depend heavily on the reference clip and the accuracy of its transcript.

Installation

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Inference

Download this model repository, then provide your own reference recording and its exact transcript:

python inference.py \
  --reference-audio ./my_reference.wav \
  --reference-text "هلا والله، هذا تسجيل مرجعي واضح للصوت." \
  --text "أهلاً وسهلاً، هذه تجربة للنموذج السعودي." \
  --output ./output.wav

The reference audio is a runtime input and is not bundled with this release. For best results, use a clean mono recording of roughly 5–12 seconds with minimal background noise and provide a verbatim Arabic transcript. Use only a voice you own or have permission to synthesize.

Device selection is automatic in the order CUDA → Apple MPS → CPU. Override it with --device cuda, --device mps, or --device cpu.

Limitations

• Pronunciation may be unreliable for diacritics, uncommon names, code switching, long numbers, and text outside the Saudi/Arabic training domain.
• The training mirror did not provide stable speaker, actor, or program IDs, so strict speaker-disjoint evaluation could not be established.
• Reference conditioning can reproduce identity cues. Obtain consent and do not use the model for impersonation, fraud, or deceptive media.
• Generated audio should be disclosed as synthetic when appropriate.
• This release has not been evaluated for safety-critical use.

Licenses and attribution

The Stage 4 weights are released under CC BY-NC-SA 4.0. SILMA TTS v1 is an Apache-2.0 upstream model; F5-TTS is MIT-licensed; SADA22 is attributed to SDAIA and the Saudi Broadcasting Authority. See THIRD_PARTY_NOTICES.md for pinned revisions and complete provenance pointers.