Fasih-TTS-V1 — Arabic (MSA / Fusha) Professional Male TTS

▶ Try the live demo (ZeroGPU): https://huggingface.co/spaces/NightPrince/Fasih-TTS

Fasih (فَصِيح, "eloquent") is a single-speaker Modern Standard Arabic text-to-speech model with a professional male "news-anchor" voice, fine-tuned from Coqui XTTS v2. It voices the "Muslim" religious-Q&A assistant, and reaches human-level intelligibility with broadcast-grade consistency.

Samples

Greeting	Fiqh explanation
السَّلَامُ عَلَيْكُمْ... أَنَا مُسْلِم، مُسَاعِدُكَ الصَّوْتِيُّ	الْوُضُوءُ شَرْطٌ لِصِحَّةِ الصَّلَاةِ...

---

Capabilities

• Human-level intelligibility — 1.3% CER (ASR-measured), matching the 1.8% error floor of the original human recordings. The synthetic voice is as understandable as the actor.
• Broadcast consistency — the same sentence synthesized 4× gave identical CER, and across 24 stress generations there were zero autoregressive failures (no loops, skips, or early cut-offs) — the classic XTTS weakness, eliminated.
• True Fusha, correct iʿrāb — trained on fully-diacritized text and paired with a built-in CATT diacritizer, so even bare, undiacritized Arabic is pronounced correctly, case-endings and all.
• Production-ready text handling — auto-expands numbers to words, applies a sacred-term pronunciation lexicon, and chunks long passages — raw assistant text becomes speech-ready in a single call.
• Real-time and streaming — RTF ≈ 0.60, streaming first-audio ≈ 675 ms, clean 24 kHz output — suitable for live conversational agents.

---

Benchmarks

CER between the intended text and a Whisper-large-v3 transcription of the synthesized audio (both diacritics-stripped and orthography-normalized), judged against the ASR floor on the human originals.

Test set	Clips	Mean CER	Worst CER
Varied MSA sentences	8	1.3%	2.2%
Same sentence ×4 (variance)	4	2.0%	2.0%
Long text (auto-chunked)	2	0.8%	0.9%
Hard stress (numbers, lists, terms)	6	2.1%	8.2%
Human originals (ASR floor)	8	1.8%	4.8%

Efficiency (1× RTX 2080 Ti, FP32)	Value
Real-time factor	~0.60
Streaming time-to-first-audio	~675 ms
Output	24 kHz mono

---

SILMA open-source Arabic TTS benchmark

Evaluated on SILMA's Open-Source Arabic TTS Benchmark (MSA, 10 fixed sentences), scored by two independent ASR judges — Whisper-large-v3 and NVIDIA NeMo Arabic FastConformer — plus UTMOS naturalness. Two judges keep the ranking honest.

Model	WER · Whisper	WER · NeMo	UTMOS
Fasih-TTS-V1 (ours)	6.5	2.5	3.16
xtts (base)	10.3	2.5	2.99
chatterbox	12.8	5.4	3.20
silma_tts	11.1	5.8	3.15
omnivoice	15.3	7.3	3.62
habibi_specialized	21.9	23.3	2.33

Fasih is top-tier on intelligibility — the best-or-tied lowest WER across both ASR judges (tied with base XTTS at 2.5% on the stronger Arabic ASR, NeMo). On naturalness (UTMOS) it is mid-pack (#3) — the smoothest model (omnivoice) is also the least accurate, and Fasih is tuned toward pronunciation correctness, which is what matters most for a religious agent. WER measures intelligibility, not naturalness; SILMA's own benchmark is a human auditory comparison.

Full per-clip provenance and all clips live in the companion dataset NightPrince/Fasih-TTS-Benchmark. Reproduce: scripts/silma_compare.py (Whisper), scripts/nemo_compare.py (NeMo), scripts/utmos_compare.py (UTMOS).

---

Architecture

---

Quick start

from huggingface_hub import snapshot_download
from TTS.tts.configs.xtts_config import XttsConfig
from TTS.tts.models.xtts import Xtts

path = snapshot_download("NightPrince/Fasih-TTS-V1")
config = XttsConfig(); config.load_json(f"{path}/config.json")
model = Xtts.init_from_config(config)
model.load_checkpoint(config, checkpoint_path=f"{path}/model.pth",
                      vocab_path=f"{path}/vocab.json", use_deepspeed=False)
model.cuda().eval()

gpt_cond, spk = model.get_conditioning_latents(audio_path=["reference.wav"])
out = model.inference("السَّلَامُ عَلَيْكُمْ وَرَحْمَةُ اللَّهِ", "ar", gpt_cond, spk,
                      temperature=0.65, repetition_penalty=2.0)
# out["wav"] -> 24 kHz mono waveform

Feed diacritized Fusha for correct iʿrāb. The included front-end (diacritization, number expansion, sacred-term lexicon, ≤166-char chunking) turns raw text into speech-ready input automatically.

---

Intended use and ethics

In scope: voicing MSA / Fusha explanatory religious and educational content that has been authored or reviewed by a qualified human.

Out of scope / prohibited

• Qur'anic recitation — requires tajwīd and human reciters; route āyāt to real audio.
• Autonomous religious rulings — the model only voices text; it does not verify content.
• Impersonation / misinformation — do not synthesize false statements in this voice.

---

Training

Fine-tuned from coqui/XTTS-v2 on NightPrince/Arabic-professional-original-voice (1297 clips, ~2.4 h, one male speaker, fully diacritized — plain transcripts diacritized with CATT, verified ≈ human gold). FP32 (Turing has no bf16; XTTS's GPT is unstable under FP16 autocast), single RTX 2080 Ti, batch 1 × grad-accum 24, gradient checkpointing, LR 5e-6. Best validation loss: 2.622.

Limitations

• Feed diacritized text for correct iʿrāb (the front-end handles it).
• Number gender-agreement (خمسة vs خمس) is not always correct.
• Source audio is 128 kbps MP3 — a soft ceiling on fidelity.
• ~2.4 h single-speaker; auto-diacritization of 371 training clips is ~95%+ (not fully human-verified).

License

Fine-tuned from Coqui XTTS v2 under the Coqui Public Model License (CPML) — non-commercial, attribution required; derivatives inherit these terms. Diacritization: CATT (MIT).

Copyright

Copyright 2026 Yahya Elnawasany (NightPrince). The Fasih-TTS-V1 model, its voice and generated audio, and the "Fasih / فَصِيح" name and branding are copyright the author. The model is distributed under the Coqui Public Model License (non-commercial, attribution); the accompanying code is MIT. Do not use the model or its outputs to impersonate, misrepresent, or generate misleading religious content. Full terms: COPYRIGHT and THIRD_PARTY_NOTICES.md in the repository.

Source and citation

Full pipeline, evaluation and serving code: https://github.com/NightPrinceY/Fasih-TTS-V1
Author and portfolio: Yahya Elnawasany (NightPrince) — https://nightprincey.github.io/Portfolio-App/

@software{fasih_tts_v1_2026,
  author = {Yahya Elnawasany (NightPrince)},
  title  = {Fasih-TTS-V1: Arabic Fusha Professional-Male Text-to-Speech},
  year   = {2026},
  url    = {https://github.com/NightPrinceY/Fasih-TTS-V1},
  note   = {Fine-tuned from Coqui XTTS v2}
}