dannwaneri/naija-voice

# Nigerian-accented voice cloning pipeline Single-stage TTS pipeline that generates speech in a target voice while preserving Nigerian accent: ``` text → Qwen3-TTS (voice + accent + intelligibility) → WAV ``` Built for content creation where existing options either wash out the Nigerian accent (XTTS, F5-TTS), produce unreliable audio (YarnGPT), or drift mid-audio on long content (Afro-TTS). Historical two-stage pipelines using RVC on top of Afro-TTS or YarnGPT are kept as legacy fallbacks — see PLAYBOOK.md. ## Status Working. See PLAYBOOK.md for realistic quality expectations, tuned parameters, and input formatting rules. Known ceiling: ~95% intelligibility on simple content, ~85% on slang-heavy content, voice drift on multi-minute takes. Not a code problem — it's the training-data-per-billion-params floor of 2024-era open-source TTS. See "Upgrade paths" below. ## Layout - `speak_qwen.py` — **current pipeline (Qwen3-TTS single-stage). Use this.** - `speak_afro.py` — legacy pipeline (Afro-TTS + RVC). Kept as fallback. - `speak.py` — legacy pipeline (YarnGPT + RVC). Historical reference. - `convert.py` — standalone RVC inference on any input WAV. - `prep_training_audio.py` — combines dataset clips into a single training WAV for RVC. - `rvc_convert.py` — early prototype, superseded by `convert.py`. - `PLAYBOOK.md` — production usage and lessons learned. - `models/` (gitignored) — RVC model checkpoints. - `outputs/` (gitignored) — generated audio. - `Applio/` (gitignored) — Applio RVC fork, cloned locally. ## External dependencies (not in this repo) The pipeline shells out to sibling projects: - **Qwen3-TTS** at `C:\Users\DELL\qwen-tts\` (current default) — QwenLM/Qwen3-TTS - **Afro-TTS** at `C:\Users\DELL\afro-tts\` (legacy) — intronhealth/afro-tts - **YarnGPT** at `C:\Users\DELL\yarngpt\` (legacy) — saheedniyi/YarnGPT Each has its own venv because their dep trees conflict. ## Setup (short version) 1. Clone this repo 2. Install Applio into `./Applio/` 3. …