Lite-Whisper Large v3 (Accuracy)

Lite-Whisper is a family of compressed Whisper checkpoints produced using LiteASR (Kamahori et al., arXiv:2502.20583), a post-training compression technique that replaces dense linear layers in the Whisper encoder with low-rank approximations derived from activation statistics.

• Encoder–decoder Transformer (inherits OpenAI Whisper Large v3 architecture)
• Low-rank-compressed encoder linears (principal-component / SVD-based)
• Uses the standard Whisper processor (128-mel input, 16 kHz)

• Drop-in replacement for Whisper Large v3 via transformers
• The -acc checkpoint targets the highest accuracy in the Lite-Whisper family, matching or outperforming Whisper Large v3 while being meaningfully smaller and faster
• No retraining required, compression is calibrated from a small activation dataset

High-quality multilingual ASR (99 languages) in latency- or memory-constrained deployments; a cheaper alternative to running vanilla Whisper Large v3.