IN BUILDML · Malayalam TTS2026
Astroid Voice brings zero-shot voice cloning to Malayalam, a language most TTS stacks ignore. Built on IndicF5 and served from a home RTX 3090, it synthesizes natural speech in real time from a few seconds of reference audio, with a live waveform/spectrum visualization in the demo.
IndicF5PyTorchTTSRTX 3090FastAPI
Self-host the GPU inference
Running IndicF5 on a local 3090 removes per-request API cost and keeps reference audio private — important for voice data.
Stream synthesis, don't batch
Audio is generated and streamed in chunks so the first sound plays well before the full clip finishes rendering.
// OUTCOME
Real-time Malayalam synthesis from a home rig; demo renders a live waveform as audio streams.