AI4Bharat IndicF5 setup for local Hindi text-to-speech

This tutorial sets up AI4Bharat IndicF5 for the text-to-speech phase of a local AI dubbing pipeline. Earlier phases handle transcription and translation; this session installs the speech environment and runs a Hindi sample. Converting translated JSON segments into full audio tracks is left for the next session.

The demonstrated setup uses Windows with WSL2 Ubuntu 24.04, a Conda environment with Python 3.10, and PyTorch packages with CUDA 12.8 support. The presenter checks NVIDIA access with nvidia-smi and a Python GPU test. The displayed GPU has 16 GB of VRAM, but the tutorial does not establish that as a minimum requirement. It also covers FFmpeg, cache paths, pinned dependencies such as Transformers 4.49.0, cloning the IndicF5 repository, and an editable installation. Hugging Face access includes accepting the repository terms and authenticating with a token.

The Hindi test supplies reference audio and a matching transcript, with output set to 24,000 Hz. The presenter says the reference clip should be no longer than 15 seconds. An initial run downloads a model shown as 1.4 GB, then fails because the reference file is in the wrong folder. Moving it into the expected Hindi directory lets the script generate speech and save it to the TTS output folder. The presenter describes the result as matching the sample voice; that is a demonstration claim, rather than a general quality guarantee.