kihahu/kikuyu-tts

# kikuyu-tts English-to-Kikuyu translation and speech synthesis pipeline with an MMS-first TTS workflow. ## Current TTS Paths - Preferred: MMS full-checkpoint continuation on Waxal via `configs/train_mms_tts_kik_waxal.yaml` and `scripts/bootstrap_mms_kikuyu_tts_finetune.py` - Baseline: MMS base-model selection and export via `configs/finetune_mms_tts_kik.yaml` and `scripts/prepare_mms_tts_finetune.py` - Secondary: scratch Coqui VITS training via `scripts/colab_train_vits_scratch.py` ## Current ASR Path - Preferred: MMS ASR fine-tuning via `configs/train_mms_asr_kik.yaml` and `scripts/train_mms_asr_kik.py` Run it with: ```bash python scripts/train_mms_asr_kik.py \ --config configs/train_mms_asr_kik.yaml ``` This trains `facebook/mms-1b-all` with the Kikuyu MMS head (`kik`) on the paired `audio` + `text` data from `google/WaxalNLP`, config `kik_tts`. Metrics and hyperparameters are logged to **MLflow** by default (`report_to: mlflow`); see `docs/mms_asr_finetune_kikuyu.md`. ## Recommended Workflow ### English To Kikuyu Audio The first end-to-end app surface is a script. It translates English text to Kikuyu with NLLB, then synthesizes Kikuyu audio with the current best Waxal MMS/VITS checkpoint, `G_77100`. Smoke test the Waxal audio path with known Kikuyu text: ```bash python scripts/english_to_kikuyu_audio.py \ --translation-backend identity \ --text "Ni wega gukwona umuthi." \ --output-wav artifacts/english_to_kikuyu_audio/smoke_identity.wav ``` Run English text through translation and TTS: ```bash python scripts/english_to_kikuyu_audio.py \ --input path/to/shakespeare_chapter.txt \ --output-wav artifacts/english_to_kikuyu_audio/shakespeare_chapter.wav \ --translated-output artifacts/english_to_kikuyu_audio/shakespeare_chapter.kik.txt \ --manifest-json artifacts/english_to_kikuyu_audio/shakespeare_chapter.manifest.json \ --translation-device cpu \ --tts-device cpu ``` The compatibility wrapper `scripts/pipeline.py` calls the same implementation: ```b …