Johniblazee/naija-speech

# naija-speech Fine-tuning and benchmarking open **Speech-to-Text (STT)** and **Text-to-Speech (TTS)** models for **Nigerian-accented English** — the implementation artifact for the MIT thesis *"Speech-to-Text & Text-to-Speech for Nigerian English Accents"* (John Ikpe Michael, MIVA Open University). This repo follows the thesis's **Design Science Research** plan: build a curated corpus, fine-tune models with parameter-efficient methods (LoRA), and evaluate them on three axes — **accuracy**, **fairness** (per macro-accent), and **production efficiency**. > **Status:** corpus-first vertical slice. The first milestone is one STT model > (Whisper) fine-tuned with LoRA on the Nigerian subset of AfriSpeech-200, evaluated against > its zero-shot baseline. TTS comes after the STT slice works end-to-end. ## TTS roster (locked 2026-07-11; code lands at TTS kickoff) - **Fine-tuned (3):** StyleTTS 2 (primary, style-diffusion, MIT) → Orpheus-3B (codec-LM, Apache-2.0, Unsloth 4-bit QLoRA — same toolchain as the STT slice) → F5-TTS (flow-matching DiT, stretch; MIT code / CC-BY-NC weights). - **Zero-shot baselines (3):** YarnGPT (Nigerian domain baseline), Qwen3-TTS, XTTS-v2 (demoted from fine-tune slot: CPML non-commercial). Compute tiers: free Colab T4 for dev + 4-bit QLoRA runs → RunPod Community A5000 24GB (~$0.27/hr) / A40 48GB (~$0.44/hr) for most fine-tunes → A100-80 (~$1.40–1.65/hr RunPod, or HF Jobs per-minute) for final citable runs only. --- ## Why "vertical slice" first Instead of building all 3 STT + 3 TTS models at once, we prove the **entire pipeline** on one model first: `corpus → zero-shot baseline → LoRA fine-tune → accent-stratified evaluation`. Once that works, adding more architectures is repetition, not risk. This is the de-risking strategy described in the thesis methodology. ## Pipeline (run in order) | Step | Script | What it does | |------|--------|--------------| | 0 | `scripts/00_eda.py` | EDA on the Nigerian subset (metadata only — no audio): …