NileTTS: Egyptian Arabic TTS Dataset

• NileTTS is an Egyptian Arabic TTS resource offering a fully synthetic dataset and reproducible pipeline to address under-resourced dialect challenges.
• It employs a four-step pipeline—LLM text generation, audio synthesis, transcription with speaker diarization, and quality verification—to create a balanced 38.1-hour corpus.
• The fine-tuned XTTS v2 model shows significant performance gains with a 29.9% reduction in WER and improved speaker similarity compared to the baseline.

Searching arXiv for the NileTTS paper and closely related context. NileTTS is an Egyptian Arabic text-to-speech resource introduced in "LLM-to-Speech: A Synthetic Data Pipeline for Training Dialectal Text-to-Speech Models" (Khamis et al., 17 Feb 2026). It combines a fully synthetic corpus-construction workflow, a 38.1-hour transcribed speech dataset from two speakers, and an open-source fine-tuned XTTS v2 model for dialectal speech synthesis. The work is positioned against the concentration of Arabic TTS resources on Modern Spoken Arabic and Gulf dialects, and identifies Egyptian Arabic as severely under-resourced despite being described as the most widely understood Arabic dialect. Its central contribution is a four-step pipeline in which LLMs generate dialectal text, audio synthesis tools render it as speech, automatic transcription and speaker diarization convert the material into utterance-level supervision, and manual quality verification filters the result (Khamis et al., 17 Feb 2026).

1. Scope and contribution profile

NileTTS is presented as having three primary contributions: the first publicly available Egyptian Arabic TTS dataset, a reproducible synthetic data generation pipeline for dialectal TTS, and an open-source fine-tuned model (Khamis et al., 17 Feb 2026). The dataset contains speech from two balanced voices and spans medical, sales and customer service, and general conversation domains. The accompanying model is obtained by fine-tuning XTTS v2 against this corpus and evaluating it against the pretrained Arabic baseline.

Within the paper’s framing, NileTTS addresses a dialect-resource asymmetry rather than proposing a new end-to-end TTS architecture. The emphasis is therefore on corpus creation, supervision derivation, and adaptation of an existing multilingual TTS backbone. This suggests that the work belongs simultaneously to dialectal speech synthesis, synthetic data generation, and low-resource speech technology.

2. Synthetic data generation pipeline

NileTTS is built entirely via a four-step synthetic pipeline: content generation by LLMs, audio synthesis with a two-speaker TTS system, automatic transcription and speaker diarization, and manual quality verification (Khamis et al., 17 Feb 2026). The pipeline is summarized in the paper’s Fig. 1 as an end-to-end flow of LLM $\rightarrow$ TTS $\rightarrow$ Whisper $\rightarrow$ ECAPA $\rightarrow$ QC.

For content generation, the models used are Google Gemini and Anthropic Claude. The prompt template is reported as: “Write a 1–2 paragraph report entirely in Egyptian Arabic dialect on <DOMAIN>. Do not use Modern Standard Arabic. Make it sound like a natural conversation among Egyptians.” The domain variable ranges over Medical, Sales & Customer Service, and General Conversation. No post-processing other than light normalization of punctuation was applied, although manual spot-checks of dialect authenticity were conducted during quality control.

For audio synthesis, the TTS engine is NotebookLM’s built-in Egyptian-Arabic vocoder, described as a proprietary neural vocoder. The setup uses two “virtual hosts,” one male and one female, who converse naturally over each text report. The audio length per report is approximately 10–15 minutes. No additional data augmentation, including noise or speed perturbation, was applied. The details block notes a typical sampling rate of 24 kHz and mel-spectrogram parameters of 80 bins, hop $= 256$, and win $= 1024$, while also stating that these settings were not specified in the paper.

Automatic transcription and segmentation are performed with OpenAI Whisper Large (2022), described as Arabic-capable and providing word-level timestamps. Segmentation breaks audio at natural pauses via Whisper timestamps, discards segments shorter than 1 second or silence-only intervals, and yields one WAV file plus one text transcription per utterance.

Speaker diarization uses ECAPA-TDNN with 192-dimensional embeddings from SpeechBrain, followed by K-Means clustering with $k = 2$ over all embeddings to obtain two centroids. Each segment is then assigned to the closest centroid by cosine similarity. Manual quality control reviews the entire Sales & Customer Service split and 10% samples of the Medical and General splits. The review criteria are transcription accuracy, including dialectal vocabulary and idioms, speaker-label correctness, and audio integrity, including absence of dropouts and clipping. Faulty segments are corrected or removed. Inter-annotator statistics are not reported, but disagreements were adjudicated by a senior annotator.

3. Dataset composition and corpus characteristics

The released corpus has a total duration of 38.1 hours and contains 9,521 utterances (Khamis et al., 17 Feb 2026). The two speakers are balanced: SPEAKER_01 is male with 4,865 utterances, and SPEAKER_02 is female with 4,656 utterances. The average utterance length is 14.4 seconds, and the train/eval split is 90%/10%, with topic-level held-out material in evaluation.

Component	Utterances	Hours
Sales & Customer Service	4,975	21.0
General Conversations	2,979	11.2
Medical	1,567	5.9

The details block additionally reports text-diversity estimates not explicitly reported in the paper: vocabulary size of approximately 12K unique word forms and average sentence length of approximately 20 words. Because these are explicitly labeled as estimated rather than directly reported, they should be treated as approximate corpus descriptors rather than canonical benchmark statistics.

The corpus composition indicates that Sales & Customer Service is the largest domain by both utterance count and duration. A plausible implication is that the fine-tuned model may have stronger domain exposure for transactional and service-oriented discourse than for medical content, although the paper does not provide per-domain synthesis performance.

4. Model adaptation and optimization setup

The base model is XTTS v2, described in the details block as a “Casuistic multilingual, GPT-style autoregressive with discrete audio tokens + DVAE wave-decoder” system (Khamis et al., 17 Feb 2026). Its input is up to 400 text tokens, using byte-pair or phoneme embedding, and its output is a sequence of discrete audio tokens decoded by a frozen DVAE into waveform.

The fine-tuning protocol modifies only the GPT-decoder, while keeping the DVAE frozen. No architectural changes beyond integration of log-mel conditioning for dialectal prosody are reported. In other words, NileTTS adapts an existing multilingual model to Egyptian Arabic primarily through parameter fine-tuning and conditioning, not through redesign of the acoustic tokenization or waveform decoder.

The training configuration is reported as follows:

Hyperparameter	Value
Epochs	30
Batch Size	2
Gradient Accumulation Steps	8
Effective Batch Size	16
Learning Rate (constant)	$5 \times 10^{-6}$
Optimizer	AdamW
Weight Decay	$1 \times 10^{-2}$
Max Text Length	400 tokens

The loss is standard cross-entropy over the audio-token sequence:

$$L_{\text{TTS}} = - \sum_t \log p(y_t \mid y_{(<t)}, x)$$

where $\rightarrow$0 are discrete audio tokens and $\rightarrow$1 is text input.

This optimization setup places the work within the now-common regime of adapting autoregressive discrete-audio TTS systems using relatively small dialect-specific corpora. The paper’s contribution lies in demonstrating that such adaptation is viable even when the corpus itself is produced by a synthetic pipeline.

5. Evaluation methodology and reported results

The evaluation compares the fine-tuned NileTTS model against pretrained XTTS v2, characterized as a baseline Arabic model trained mostly on Modern Standard Arabic and Gulf dialects (Khamis et al., 17 Feb 2026). Three objective metrics are used. Word Error Rate is computed by synthesizing audio, transcribing it with Whisper, and comparing the result to reference text. Character Error Rate is computed analogously. Speaker Similarity is the cosine similarity between ECAPA-TDNN embeddings of reference and synthesized speech.

Model	WER	CER	SpkSim
Baseline XTTS v2	26.8%	8.1%	0.713
Finetuned NileTTS	18.8%	4.1%	0.755

The reported relative improvements are WER down 29.9%, corresponding to an absolute reduction of 8 percentage points, CER down 49.4%, and Speaker Similarity up 5.9%. The paper also references Fig. 2, which shows training curves for evaluation loss, WER, CER, and SpkSim over steps.

For subjective evaluation, 50 random test utterances were assessed, and expert listeners confirmed natural prosody, accurate Egyptian-Arabic pronunciation, and strong speaker identity preservation. No formal MOS or AB tests were conducted, and these are identified as planned future work. This is an important methodological boundary: the study reports expert qualitative confirmation, but not standardized listener-preference or mean-opinion-score measurements.

6. Interpretation, limitations, and release ecosystem

NileTTS is described as demonstrating a fully synthetic yet high-quality workflow for creating a 38-hour Egyptian Arabic TTS corpus, together with substantial gains in intelligibility and voice similarity after dialect-specific fine-tuning (Khamis et al., 17 Feb 2026). The resource package includes code and pipeline components on GitHub, the 38-hour WAV+CSV dataset on Hugging Face, and fine-tuned model weights on Hugging Face. Licensing is described as open-source, with the note “Apache 2.0 / CC-BY 4.0 typical; check each repo.” Replication instructions are provided through the repository README, and a prebuilt Docker container includes all dependencies, specifically Whisper v1.1, SpeechBrain 0.5, and Coqui XTTS.

Several limitations are explicit in the reported material. The audio synthesis stage depends on NotebookLM’s built-in Egyptian-Arabic vocoder, which is proprietary. Inter-annotator statistics for manual quality control are not reported. Text-diversity quantities such as vocabulary size and average sentence length are estimates rather than directly reported measurements. Human evaluation does not include formal MOS or AB tests. These points matter because they delimit reproducibility, annotation reliability analysis, and the strength of perceptual claims.

A common misconception would be to treat NileTTS as a purely naturally recorded Egyptian Arabic corpus. The reported workflow states the opposite: the dataset is built entirely through synthetic generation, automated processing, and manual verification. Another possible misconception would be to interpret the work as proving universal dialectal transfer across Arabic. The evidence provided is specific to Egyptian Arabic, two synthetic speakers, and the stated domains. What the paper does establish is that an LLM-to-speech data pipeline can be used to produce a dialect-specific corpus of sufficient quality to improve a multilingual TTS model on intelligibility, orthographic recoverability under ASR-based evaluation, and speaker similarity.