Najdi-TTS

Najdi Arabic text-to-speech in one consistent Najdi voice. Najdi-TTS is the Saudi model of Habibi-TTS fine-tuned on 28.4 hours of Najdi speech in a single voice (the Dr. Mohammed corpus). Give it everyday Najdi text, such as «وش أخبارك؟ والله إني مشتاق لك من زمان», and a short clip of that voice, and it reads the text in that voice.

Research model from the Emotional Najdi TTS project (Imam Mohammad Ibn Saud Islamic University). Non-commercial use only (CC BY-NC-SA 4.0).

Before and after

The same new sentences, never seen in training, read with the same reference clip of the Dr. Mohammed voice.

Reference clip (from the Dr. Mohammed dataset, drm_000001):

«وش أخبارك؟ والله إني مشتاق لك من زمان»

Before: original Habibi-TTS Saudi	After: Najdi-TTS

«يا شيخ خلنا نطلع البر نهاية الأسبوع ونشب النار ونسوي قهوة»

Before: original Habibi-TTS Saudi	After: Najdi-TTS

On 50 held-out sentences of the training voice, speaker similarity to the original recordings rose from 0.605 (original model) to 0.813 (Najdi-TTS), with a character error rate of 0.014.

Getting the best results

Option 1: speak in the Dr. Mohammed voice (recommended)

Najdi-TTS was trained on one voice and sounds best in it. Use a reference clip from the Dr. Mohammed dataset together with its exact transcript:

from importlib.resources import files

from f5_tts.infer.utils_infer import load_model, load_vocoder, preprocess_ref_audio_text
from f5_tts.model import DiT
from habibi_tts.infer.utils_infer import infer_process
from huggingface_hub import hf_hub_download
from omegaconf import OmegaConf

ckpt = hf_hub_download("nsa01n/najdi-tts", "model_40290_ema.safetensors")
vocab = hf_hub_download("nsa01n/najdi-tts", "vocab.txt")
arch = OmegaConf.to_container(OmegaConf.load(str(files("f5_tts").joinpath("configs/F5TTS_v1_Base.yaml"))).model.arch)
model = load_model(DiT, arch, ckpt, mel_spec_type="vocos", vocab_file=vocab)
vocoder = load_vocoder(vocoder_name="vocos")

# reference: a Dr. Mohammed clip and its exact transcript
ref_wav = hf_hub_download("nsa01n/najdi-synthetic-dr-mohammed", "data/train/000/drm_000001.wav", repo_type="dataset")
ref_audio, ref_text = preprocess_ref_audio_text(ref_wav, "أبوي يصحى قبل الفجر ويشغل الدلة قبل لا يصحى أحد")

wav, sr, _ = infer_process(ref_audio, ref_text, "وش أخبارك؟ والله إني مشتاق لك من زمان", model, vocoder,
                           mel_spec_type="vocos", nfe_step=32)

Any other clip from the dataset works the same way: pass its audio and the text field from metadata.jsonl.

Tips:

• The transcript must match the reference audio exactly, or the output drifts.
• Write numbers as words (خمسة وتسعين, not ٩٥).
• One-word phrases are too short for the model's length estimate; give them at least about one second (fix_duration = reference length + 1.0).

Option 2: speak in other voices (the mixing formula)

Because the training data had a single speaker, Najdi-TTS pulls other reference voices toward the Dr. Mohammed voice. To copy a different voice, mix Najdi-TTS with the original Habibi-TTS Saudi model:

θ_mixed = α · θ_Najdi-TTS + (1 − α) · θ_original

Every weight is a weighted average of the two models. α = 1 is Najdi-TTS, α = 0 is the original model, and values in between move gradually from one to the other. We measured the trade-off on 50 natural Najdi test items, each with a different real speaker as the reference:

α	Similarity to the reference speaker ↑	Similarity to the training voice	Dialect score	CER ↓
0 (original)	0.698	0.108	0.757	0.060
0.25	0.673	0.152	0.754	0.062
0.5	0.587	0.219	0.763	0.063
0.75	0.404	0.365	0.749	0.063
1 (Najdi-TTS)	0.288	0.460	0.761	0.072

α = 0.25 keeps almost all of the voice cloning. Build the mixed checkpoint with the included script and load it in place of model_40290_ema.safetensors:

python interpolate.py --alpha 0.25 --out najdi-tts-a0.25.safetensors

Example with a female voice the model was not trained on, sentence «يا شيخ خلنا نطلع البر…»:

Reference voice	α = 0	α = 0.25	α = 1
similarity to this voice	0.77	0.78	0.27

Evaluation

173 natural Najdi clips from 13 SADA television shows, reviewed by a native speaker; each item read in the voice of its own real speaker (automatic speech length, 32 steps, seed 0). Mean (95% CI).

Model	CER ↓	Dialect score P(dialect) ↑	Speaker similarity ↑	UTMOS
Natural recordings	0.105	0.819	–	1.84
Original Habibi-TTS Saudi	0.071 (0.054–0.089)	0.745 (0.716–0.773)	0.717 (0.700–0.733)	1.73
Najdi-TTS	0.080 (0.063–0.097)	0.763 (0.737–0.786)	0.315 (0.297–0.334)	2.91

CER: Cohere Transcribe Arabic. P(dialect): a dialect-versus-MSA classifier trained on SADA; it does not separate Najdi from other Saudi dialects. Speaker similarity: ECAPA-TDNN, output against the reference clip. UTMOS is English-trained and secondary; it favours clean studio audio.

Intelligibility and dialect score stay at the level of the original model, which already speaks Saudi Arabic. The change is the voice: Najdi-TTS speaks in the Dr. Mohammed voice, which is why Option 1 gives the best results and Option 2 exists for other voices.

Files

File	What
model_40290_ema.safetensors	Najdi-TTS weights (EMA, update 40,290 of 40,290, 102 epochs)
vocab.txt	Habibi-TTS Saudi character vocabulary (unchanged)
interpolate.py	builds a mixed checkpoint for Option 2
base.yaml, data_summary.json	training settings and data split
samples/	the audio on this page

Training

Habibi-TTS Saudi (Specialized/SAU/model_200000, an F5-TTS v1 model, 337 M parameters) fine-tuned on 20,285 clips (28.43 h) of synthetic Najdi speech in one cloned voice, 200 clips held out. Learning rate 1e-5 with 1,000 warmup updates and linear decay, batches of 25,600 mel frames, bf16, 40,290 updates on one H100 (about 4.9 h).

Limitations

• One training voice: other voices are pulled toward it (use Option 2).
• The text has no diacritics, so ambiguous words such as ميت (mayyit) follow the model's guess.
• No listening test with native speakers yet; the dialect score is automatic.

License

CC BY-NC-SA 4.0, inherited from Habibi-TTS. Research and non-commercial use only; derivatives must keep the same license.