2.2 KiB
2.2 KiB
Q2 — Fine-tuning an Arabic TTS or STT
⏱ Budget: ~15 minutes · Weight: 25%
Choose ONE option
- Option A — TTS. Fine-tune an open Arabic (or multilingual) text-to-speech model — e.g. Coqui XTTS-v2, VITS, or StyleTTS2 — on a single-speaker Arabic dataset.
- Option B — STT. Fine-tune Whisper (e.g.
whisper-small,large-v3) on Arabic speech — e.g. Mozilla Common Voicear, MGB-2, or a custom corpus.
Produce a runnable guide covering
- Model & dataset choice — justify it; cite dataset sizes, license, and dialect (MSA vs. Egyptian vs. Gulf vs. Maghrebi).
- Data preprocessing — audio resampling/normalization, text normalization, and the Arabic-specific gotchas: diacritization (تشكيل), letter variants (أ/إ/آ, ة/ه, ى/ي), and (for STT) whether to strip diacritics before computing WER. Train/val split.
- Training setup — framework (e.g. Hugging Face Trainer, Coqui TTS, NeMo), LoRA vs. full fine-tune, key hyperparameters (batch size, LR, epochs, warmup), hardware, and a rough time/cost estimate.
- Evaluation — for STT: WER/CER (and how to compute them fairly for Arabic); for TTS: MOS, speaker similarity (SECS), and using an ASR to measure intelligibility. Include the actual metric definitions/commands.
- Deployment on Ghaymah — GPU sizing, container vs. function, serving stack, latency, and how the model is served behind an API.
- Risks & mitigations — dialect coverage, hallucination, diacritic fidelity, MSA vs. dialect mismatch, data licensing.
Include real code snippets (Python / Hugging Face / CLI). Cite sources inline.
Deliverable
Fill in the template below. State clearly which option you chose.
ANSWER — Arabic TTS/STT Fine-tune
Option chosen: A / B (delete one)
1. Model & dataset choice
write here
2. Data preprocessing
write here + code snippets
3. Training setup
write here + code/config snippets
4. Evaluation
write here + metric definitions/commands
5. Deployment on Ghaymah
write here
6. Risks & mitigations
| Risk | Mitigation |
|---|---|