v0.6.4

[0.6.4] - 2026-05-21

Changed

• Made py-roller install the sole supported audio runtime installation path and removed the legacy audio-core extra from package metadata.
• Switched SOCKS proxy support from httpx[socks]/socksio to requests[socks]/PySocks to match Hugging Face's download stack.
• Removed torchvision from the official install profiles because py-roller does not use vision workloads.

Fixed

• Fixed Hugging Face large-file download failure hints so interrupted partial streams recommend retrying, lowering --transcriber-hf-max-workers, increasing HF timeouts, and checking the configured proxy.
• Removed the unused g2p_en probing path so English ARPABET conversion consistently uses the declared pronouncing dependency and grapheme fallback.