Datasetpublic

Alqayed2024/EmiratiTTS-smoke-samples

EmiratiTTS — Stage 0.5 LoRA Smoke Samples These 10 audio clips are the stage 0.5 acceptance check for the EmiratiTTS project (Chatterbox Multilingual fine-tuned for Emirati Arabic). This is NOT a model release. It is a sanity check that the data + tokenizer reference-clip + ChatterboxMultilingualTTS pipeline is wired correctly before committing GPUs to the long full-FT run. Quality is irrelevant at this stage — the only pass criterion is "intelligible Arabic from both reference… See the full description on the dataset page: https://huggingface.co/datasets/Alqayed2024/EmiratiTTS-smoke-samples.

sourceHugging Facemitupdated 5mo agoView on Hugging Face

1likes1.2kdownloads

Dataset Card

EmiratiTTS — Stage 0.5 LoRA Smoke Samples

These 10 audio clips are the stage 0.5 acceptance check for the EmiratiTTS project (Chatterbox Multilingual fine-tuned for Emirati Arabic).

This is NOT a model release. It is a sanity check that the data + tokenizer

—reference-clip + ChatterboxMultilingualTTS pipeline is wired correctly before committing GPUs to the long full-FT run. Quality is irrelevant at this stage — the only pass criterion is "intelligible Arabic from both reference voices".

Training that produced these samples

—Base model: ResembleAI/chatterbox (multilingual, T3 + S3Gen)
—Adapter: LoRA injected on the T3 transformer (210 layers) via the Ahmed-Ezzat20 chatterbox-finetuning-multilingual fork's lora.py
—Data: 696 jamal_mulla clips, 60 minutes of audio (highest-SNR Emirati segments, 4–10 s, single channel — chosen for repeatability, not coverage)
—Schedule: 1 epoch, ~4 minutes wall-clock on a single A100-80GB
—Final loss: train 0.8967, val 0.8485

Reference clips

Two held-out reference voices from the EmiratiTTS reference library:

Reference	Channel	Role
`jamal_mulla_01`	jamal_mulla	seen training channel — sanity check
`uaevideos_01`	uaevideos	held-out channel — voice-variation test

If both voices sound like distinct speakers, the reference-clip → speaker conditioning path is working at inference time.

Sentences

Five Arabic sentences spanning emphatics (ص/ض/ط/ظ), pharyngeals (ع/ح), uvulars (ق/خ/غ), and one with a question mark for prosody:

1.السلام عليكم ورحمة الله وبركاته، كيف حالكم اليوم؟
2.أهلاً وسهلاً بكم في الإمارات العربية المتحدة.
3.الطقس اليوم حار جداً ودرجة الحرارة مرتفعة.
4.هذا اختبار لنموذج توليد الكلام باللهجة الإماراتية.
5.نشكركم على متابعتكم ونلقاكم في الحلقة القادمة.

Each sentence is rendered once per reference voice → 10 files total. Naming: <reference_name>__<sentence_idx>.wav

Notes

—All audio is 24 kHz mono PCM WAV.
—The Perth perceptual watermark has been physically removed from the Chatterbox source used for this project — these clips carry no traceable signature.
—These are smoke artifacts. The real EmiratiTTS v0.1 release will be a separate repo after the full T3 + S3Gen-tail fine-tune lands.