michael05josh-netizen/hausa-tts-coen543

Hausa text- to speech model # 🎙️ Hausa Text-to-Speech — Fine-Tuning SpeechT5 **Course:** COEN 543 — Natural Language Processing **Institution:** Ahmadu Bello University, Zaria, Nigeria **Group:** Group 1 Fine-tuning `microsoft/speecht5_tts` on the Mozilla Common Voice 25.0 Hausa dataset to produce intelligible Hausa speech from a model originally trained only on English. --- ## Demo > *Before vs. After comparison — run Phase 9 of the notebook to launch a live Gradio interface.* --- ## Project Structure ``` hausa-tts-coen543/ ├── nlp-group1-tts.ipynb # Full pipeline notebook (Phases 1–9) ├── results/ # Output charts saved during evaluation │ ├── spectrogram_comparison.png │ ├── mcd_comparison.png │ └── cer_comparison.png └── assets/ # Any supplementary images / diagrams ``` --- ## Notebook Phases | Phase | Content | |-------|---------| | 1 | Installation & imports | | 2 | Dataset loading & exploratory analysis | | 3 | Audio preprocessing pipeline | | 4 | Baseline inference (before fine-tuning) | | 5 | Dataset preparation & data collation | | 6 | Fine-tuning loop | | 7 | Evaluation: CER, MCD, before-vs-after spectrogram | | 8 | Standalone inference cell | | 9 | Gradio before-vs-after demo | --- ## Custom Contributions Four modifications over the vanilla SpeechT5 fine-tuning recipe: 1. **Amplitude normalisation** — Peak-normalises each waveform to `[-1, 1]` before feature extraction, removing speaker volume as a confounding factor. 2. **Reduction-factor label trimming** — Trims padded mel-spectrogram targets to match the model's reduction factor, preventing loss spikes from dummy frames. 3. **Combined TSV splits** — Merges `validated`, `other`, and `invalidated` TSVs (deduplicated on clip path) to maximise training data (~10 k+ samples instead of the validated-only ~8 k). 4. **Warm-up learning rate schedule** — 500-step linear warm-up before a linear decay, stabilising early training on the low-resource Hausa data. --- ## Dataset Mozilla Comm …