urdu-piper-tts-mobile-job

Urdu Piper TTS Mobile fine-tuning on Hugging Face Jobs

This repo contains a self-contained HF Jobs script to fine-tune a lightweight Piper/VITS voice for Urdu and upload mobile-friendly ONNX artifacts to the Hub.

Chosen model family: Piper (VITS exported to ONNX) because it is small, fast, and already designed for phones/edge devices. The job starts from Piper's general English Lessac generator checkpoint and adapts it to Urdu text/audio.

Default dataset: google/fleurs config ur_pk. You can switch to any Hub TTS dataset by setting DATASET_ID, DATASET_CONFIG, TEXT_COLUMN, and AUDIO_COLUMN.

Launch

hf jobs run \
  --flavor a10g-small \
  --timeout 8h \
  --secrets HF_TOKEN \
  -e MODEL_REPO_ID=abidlabs/urdu-piper-tts-mobile \
  -e DATASET_ID=google/fleurs \
  -e DATASET_CONFIG=ur_pk \
  -e DATASET_SPLIT=train+validation \
  -e MAX_EPOCHS=2000 \
  -e BATCH_SIZE=16 \
  -d \
  -- \
  pytorch/pytorch:2.2.2-cuda12.1-cudnn8-runtime \
  bash -lc 'git clone https://huggingface.co/abidlabs/urdu-piper-tts-mobile-job /workspace/job && python /workspace/job/train_hf_job.py'

The trained model is uploaded to MODEL_REPO_ID as .onnx plus Piper JSON config.

Faster smoke test

hf jobs run --flavor a10g-small --timeout 1h --secrets HF_TOKEN \
  -e MODEL_REPO_ID=abidlabs/urdu-piper-tts-mobile-smoke \
  -e MAX_EXAMPLES=64 -e MAX_EPOCHS=2 -e BATCH_SIZE=8 \
  -d -- pytorch/pytorch:2.2.2-cuda12.1-cudnn8-runtime \
  bash -lc 'git clone https://huggingface.co/abidlabs/urdu-piper-tts-mobile-job /workspace/job && python /workspace/job/train_hf_job.py'